Zhaoheng Huang, Dacheng Wen, Yutao Zhu, Xiaoying Lian, Yushi Liang, Kai Hao, Nan Li, Liangjie Zhang, Qi Zhang, Ji-Rong Wen, Zhicheng Dou, Fangzhao Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Open-weight large language models (LLMs) enable extensive customization but remain susceptible to post-release misuse via malicious fine-tuning.While existing defenses attempt to constrain parameter-space dynamics or mitigate harmful internal representations, malicious fine-tuning continues to erode these safeguards leaving the development of fundamental, persistent defenses for open-weight models an unresolved challenge.In this paper, we characterize a safety region for open-weight LLMs and propose Safety Guidance Trigger (SGT), a framework that preserves alignment by guiding finetuning toward the safety manifold.It has two stages: (1) optimizing a safety trigger to steer the base model outputs toward safe responses and ( 2) training the open-weight model to align its internal features with trigger-induced safety representations.We demonstrate that SGT substantially improves robustness against malicious fine-tuning, forcing adversaries to significantly increase data budgets to bypass safeguards.Our analysis further confirms that this approach anchors model representations within a safety region that remains resilient under adversarial attacks.

Yueqi Xie, Tao Qi, Jingwei Yi, Xiyuan Yang, Ryan Whalen, Junming Huang, Qian Ding, Yu Xie, Xing Xie, Fangzhao Wu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The integration of large language models (LLMs) with external content has enabled applications such as Microsoft Copilot but also introduced vulnerabilities to indirect prompt injection attacks. In these attacks, malicious instructions embedded within external content can manipulate LLM outputs, causing deviations from user expectations. To address this critical yet under-explored issue, we introduce the first benchmark for bindirect prompt injection attacks, named BIPIA, to assess the risk of such vulnerabilities. Using BIPIA, we evaluate existing LLMs and find them universally vulnerable. Our analysis identifies two key factors contributing to their success: LLMs' inability to distinguish between informational context and actionable instructions, and their lack of awareness in avoiding the execution of instructions within external content. Based on these findings, we propose two novel defense mechanisms -- boundary awareness and explicit reminder -- to address these vulnerabilities in both black-box and white-box settings. Extensive experiments demonstrate that our black-box defense provides substantial mitigation, while our white-box defense reduces the attack success rate to near-zero levels, all while preserving the output quality of LLMs. We hope this work inspires further research into securing LLM applications and fostering their safe and reliable use. Our code is available at https://github.com/microsoft/BIPIA.

Federated learning has recently been applied to recommendation systems to protect user privacy. In federated learning settings, recommendation systems can train recommendation models by collecting the intermediate parameters instead of the real user data, which greatly enhances user privacy. In addition, federated recommendation systems (FedRSs) can cooperate with other data platforms to improve recommendation performance while meeting the regulation and privacy constraints. However, FedRSs face many new challenges such as privacy, security, heterogeneity, and communication costs. While significant research has been conducted in these areas, gaps in the surveying literature still exist. In this article, we: 1) summarize some common privacy mechanisms used in FedRSs and discuss the advantages and limitations of each mechanism; 2) review several novel attacks and defenses against security; 3) summarize some approaches to address heterogeneity and communication costs problems; 4) introduce some realistic applications and public benchmark datasets for FedRSs; and 5) present some prospective research directions in the future. This article can guide researchers and practitioners understand the research progress in these areas.

Wenjun Peng, Jingwei Yi, Fangzhao Wu, Shangxi Wu, Bin Bin Zhu, Lingjuan Lyu, Binxing Jiao, Tong Xu, Guangzhong Sun, Xing Xie. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Federated learning (FL) is important for privacy-preserving services by training models without collecting raw user data. Most FL algorithms assume all data is annotated, which is impractical due to the high cost of labeling data in real applications. To alleviate the reliance on labeled data, semi-supervised federated learning (SSFL) has been proposed to utilize unlabeled data on clients to improve model performance. However, most existing methods either have privacy issues which share models trained on other clients, or generate pseudo-labels for unlabeled local datasets with the global model, which is usually biased towards the global data distribution. The latter may lead to sub-optimal accuracy of pseudo-labels, due to the gap between the local data distribution and the global model, especially in non-IID settings. In this paper, we propose a semi-supervised heterogeneous federated learning method with local knowledge enhancement, called FedLoKe, which aims to train an accurate global model from both labeled and unlabeled local data with non-IID distributions. Specifically, in FedLoKe, the server maintains a global model to capture global data distribution, and each client learns a local model to capture local data distribution. Since the distribution captured by the local model is aligned with the local data distribution, we utilize it to generate high-accuracy pseudo-labels of the unlabeled dataset for global model training. To prevent the local model from severely overfitting the small number of local labeled data, we further use the exponential moving average and apply the global model to generate pseudo-labels for local modeling training. Experiments on four datasets show the effectiveness of FedLoKe. Our code is available at: https://github.com/zcfinal/FedLoKe.

Federated learning is used to train a shared model in a decentralized way without clients sharing private data with each other. Federated learning systems are susceptible to poisoning attacks when malicious clients send false updates to the central server. Existing defense strategies are ineffective under non-IID data settings. This paper proposes a new defense strategy, FedCPA (Federated learning with Critical Parameter Analysis). Our attack-tolerant aggregation method is based on the observation that benign local models have similar sets of top-k and bottom-k critical parameters, whereas poisoned local models do not. Experiments with different attack scenarios on multiple datasets demonstrate that our model outperforms existing defense strategies in defending against poisoning attacks.

Federated learning has exhibited vulnerabilities to Byzantine attacks, where the Byzantine attackers can send arbitrary gradients to a central server to destroy the convergence and performance of the global model. A wealth of robust AGgregation Rules (AGRs) have been proposed to defend against Byzantine attacks. However, Byzantine clients can still circumvent robust AGRs when data is non-Identically and Independently Distributed (non-IID). In this paper, we first reveal the root causes of performance degradation of current robust AGRs in non-IID settings: the curse of dimensionality and gradient heterogeneity. In order to address this issue, we propose GAS, a GrAdient Splitting approach that can successfully adapt existing robust AGRs to non-IID settings. We also provide a detailed convergence analysis when the existing robust AGRs are combined with GAS. Experiments on various real-world datasets verify the efficacy of our proposed GAS. The implementation code is provided in https://github.com/YuchenLiu-a/byzantine-gas.

Personalized federated learning (FL) aims to collaboratively train a personalized model for each client. Previous methods do not adaptively determine who to collaborate at a fine-grained level, making them difficult to handle diverse data heterogeneity levels and those cases where malicious clients exist. To address this issue, our core idea is to learn a collaboration graph, which models the benefits from each pairwise collaboration and allocates appropriate collaboration strengths. Based on this, we propose a novel personalized FL algorithm, pFedGraph, which consists of two key modules: (1) inferring the collaboration graph based on pairwise model similarity and dataset size at server to promote fine-grained collaboration and (2) optimizing local model with the assistance of aggregated model at client to promote personalization. The advantage of pFedGraph is flexibly adaptive to diverse data heterogeneity levels and model poisoning attacks, as the proposed collaboration graph always pushes each client to collaborate more with similar and beneficial clients. Extensive experiments show that pFedGraph consistently outperforms the other $14$ baseline methods across various heterogeneity levels and multiple cases where malicious clients exist. Code will be available at https://github.com/MediaBrain-SJTU/pFedGraph.

Federated learning (FL) is an important technique for learning models from decentralized data in a privacy-preserving way. Existing FL methods usually uniformly sample clients for local model learning in each round. However, different clients may have significantly different data sizes, and the clients with more data cannot have more opportunities to contribute to model training, which may lead to inferior performance. In this paper, instead of client uniform sampling, we propose a novel data uniform sampling strategy for federated learning (FedSampling), which can effectively improve the performance of federated learning especially when client data size distribution is highly imbalanced across clients. In each federated learning round, local data on each client is randomly sampled for local model learning according to a probability based on the server desired sample size and the total sample size on all available clients. Since the data size on each client is privacy-sensitive, we propose a privacy-preserving way to estimate the total sample size with a differential privacy guarantee. Experiments on four benchmark datasets show that FedSampling can effectively improve the performance of federated learning.

Federated learning enables learning from decentralized data sources without compromising privacy, which makes it a crucial technique. However, it is vulnerable to model poisoning attacks, where malicious clients interfere with the training process. Previous defense mechanisms have focused on the server-side by using careful model aggregation, but this may not be effective when the data is not identically distributed or when attackers can access the information of benign clients. In this paper, we propose a new defense mechanism that focuses on the client-side, called FedDefender, to help benign clients train robust local models and avoid the adverse impact of malicious model updates from attackers, even when a server-side defense cannot identify or remove adversaries. Our method consists of two main components: (1) attack-tolerant local meta update and (2) attack-tolerant global knowledge distillation. These components are used to find noise-resilient model parameters while accurately extracting knowledge from a potentially corrupted global model. Our client-side defense strategy has a flexible structure and can work in conjunction with any existing server-side strategies. Evaluations of real-world scenarios across multiple datasets show that the proposed method enhances the robustness of federated learning against model poisoning attacks.

News recommendation is essential for personalized news distribution. Federated news recommendation, which enables collaborative model learning from multiple clients without sharing their raw data, is a promising approach for preserving users' privacy. However, the security of federated news recommendation is still unclear. In this paper, we study this problem by proposing an untargeted attack on federated news recommendation called UA-FedRec. By exploiting the prior knowledge of news recommendation and federated learning, UA-FedRec can effectively degrade the model performance with a small percentage of malicious clients. First, the effectiveness of news recommendation highly depends on user modeling and news modeling. We design a news similarity perturbation method to make representations of similar news farther and those of dissimilar news closer to interrupt news modeling, and propose a user model perturbation method to make malicious user updates in opposite directions of benign updates to interrupt user modeling. Second, updates from different clients are typically aggregated with a weighted average based on their sample sizes. We propose a quantity perturbation method to enlarge sample sizes of malicious clients in a reasonable range to amplify the impact of malicious updates. Extensive experiments on two real-world datasets show that UA-FedRec can effectively degrade the accuracy of existing federated news recommendation methods, even when defense is applied. Our study reveals a critical security issue in existing federated news recommendation systems and calls for research efforts to address the issue. Our code is available at https://github.com/yjw1029/UA-FedRec.

Existing research efforts for multi-interest candidate matching in recommender systems mainly focus on improving model architecture or incorporating additional information, neglecting the importance of training schemes. This work revisits the training framework and uncovers two major problems hindering the expressiveness of learned multi-interest representations. First, the current training objective (i.e., uniformly sampled softmax) fails to effectively train discriminative representations in a multi-interest learning scenario due to the severe increase in easy negative samples. Second, a routing collapse problem is observed where each learned interest may collapse to express information only from a single item, resulting in information loss. To address these issues, we propose the REMI framework, consisting of an Interest-aware Hard Negative mining strategy (IHN) and a Routing Regularization (RR) method. IHN emphasizes interest-aware hard negatives by proposing an ideal sampling distribution and developing a Monte-Carlo strategy for efficient approximation. RR prevents routing collapse by introducing a novel regularization term on the item-to-interest routing matrices. These two components enhance the learned multi-interest representations from both the optimization objective and the composition information. REMI is a general framework that can be readily applied to various existing multi-interest candidate matching methods. Experiments on three real-world datasets show our method can significantly improve state-of-the-art methods with easy implementation and negligible computational overhead. The source code is available at https://github.com/Tokkiu/REMI.

The increasing data privacy concerns in recommendation systems have made federated recommendations attract more and more attention. Existing federated recommendation systems mainly focus on how to effectively and securely learn personal interests and preferences from their on-device interaction data. Still, none of them considers how to efficiently erase a user's contribution to the federated training process. We argue that such a dual setting is necessary. First, from the privacy protection perspective, "the right to be forgotten (RTBF)" requires that users have the right to withdraw their data contributions. Without the reversible ability, federated recommendation systems risk breaking data protection regulations. On the other hand, enabling a federated recommender to forget specific users can improve its robustness and resistance to malicious clients' attacks.

Algorithmic fairness has become an important machine learning problem, especially for mission-critical Web applications. This work presents a self-supervised model, called DualFair, that can debias sensitive attributes like gender and race from learned representations. Unlike existing models that target a single type of fairness, our model jointly optimizes for two fairness criteria—group fairness and counterfactual fairness—and hence makes fairer predictions at both the group and individual levels. Our model uses contrastive loss to generate embeddings that are indistinguishable for each protected group, while forcing the embeddings of counterfactual pairs to be similar. It then uses a self-knowledge distillation method to maintain the quality of representation for the downstream tasks. Extensive analysis over multiple datasets confirms the model’s validity and further shows the synergy of jointly addressing two fairness criteria, suggesting the model’s potential value in fair intelligent Web applications.

Personalized news recommendation is important for users to find interesting news information and alleviate information overload. Although it has been extensively studied over decades and has achieved notable success in improving user experience, there are still many problems and challenges that need to be further studied. To help researchers master the advances in personalized news recommendation, in this article, we present a comprehensive overview of personalized news recommendation. Instead of following the conventional taxonomy of news recommendation methods, in this article, we propose a novel perspective to understand personalized news recommendation based on its core problems and the associated techniques and challenges. We first review the techniques for tackling each core problem in a personalized news recommender system and the challenges they face. Next, we introduce the public datasets and evaluation methods for personalized news recommendation. We then discuss the key points on improving the responsibility of personalized news recommender systems. Finally, we raise several research directions that are worth investigating in the future. This article can provide up-to-date and comprehensive views on personalized news recommendation. We hope this article can facilitate research on personalized news recommendation as well as related fields in natural language processing and data mining.

Nowadays, due to the breakthrough in natural language generation (NLG), including machine translation, document summarization, image captioning, etc NLG models have been encapsulated in cloud APIs to serve over half a billion people worldwide and process over one hundred billion word generations per day. Thus, NLG APIs have already become essential profitable services in many commercial companies. Due to the substantial financial and intellectual investments, service providers adopt a pay-as-you-use policy to promote sustainable market growth. However, recent works have shown that cloud platforms suffer from financial losses imposed by model extraction attacks, which aim to imitate the functionality and utility of the victim services, thus violating the intellectual property (IP) of cloud APIs. This work targets at protecting IP of NLG APIs by identifying the attackers who have utilized watermarked responses from the victim NLG APIs. However, most existing watermarking techniques are not directly amenable for IP protection of NLG APIs. To bridge this gap, we first present a novel watermarking method for text generation APIs by conducting lexical modification to the original outputs. Compared with the competitive baselines, our watermark approach achieves better identifiable performance in terms of p-value, with fewer semantic losses. In addition, our watermarks are more understandable and intuitive to humans than the baselines. Finally, the empirical studies show our approach is also applicable to queries from different domains, and is effective on the attacker trained on a mixture of the corpus which includes less than 10% watermarked samples.

Query-aware webpage snippet extraction is widely used in search engines to help users better understand the content of the returned webpages before clicking.Although important, it is very rarely studied.In this paper, we propose an effective query-aware webpage snippet extraction method named DeepQSE, aiming to select a few sentences which can best summarize the webpage content in the context of input query.DeepQSE first learns query-aware sentence representations for each sentence to capture the fine-grained relevance between query and sentence, and then learns document-aware query-sentence relevance representations for snippet extraction.Since the query and each sentence are jointly modeled in DeepQSE, its online inference may be slow.Thus, we further propose an efficient version of DeepQSE, named Efficient-DeepQSE, which can significantly improve the inference speed of Deep-QSE without affecting its performance.The core idea of Efficient-DeepQSE is to decompose the query-aware snippet extraction task into two stages, i.e., a coarse-grained candidate sentence selection stage where sentence representations can be cached, and a fine-grained relevance modeling stage.Experiments on two real-world datasets validate the effectiveness and efficiency of our methods.

News recommendation is a widely adopted technique to provide personalized news feeds for the user.Recently, pre-trained language models (PLMs) have demonstrated the great capability of natural language understanding and benefited news recommendation via improving news modeling.However, most existing works simply finetune the PLM with the news recommendation task, which may suffer from the known domain shift problem between the pre-training corpus and downstream news texts.Moreover, PLMs usually contain a large volume of parameters and have high computational overhead, which imposes a great burden on low-latency online services.In this paper, we propose Tiny-NewsRec, which can improve both the effectiveness and the efficiency of PLM-based news recommendation.We first design a self-supervised domain-specific posttraining method to better adapt the general PLM to the news domain with a contrastive matching task between news titles and news bodies.We further propose a two-stage knowledge distillation method to improve the efficiency of the large PLM-based news recommendation model while maintaining its performance.Multiple teacher models originated from different time steps of our post-training procedure are used to transfer comprehensive knowledge to the student model in both its posttraining stage and finetuning stage.Extensive experiments on two real-world datasets validate the effectiveness and efficiency of our method.

InfoNCE is a widely used contrastive training loss. It aims to estimate the mutual information between a pair of variables by discriminating between each positive pair and its associated K negative pairs. It is proved that when the sample labels are clean, the lower bound of mutual information estimation is tighter when more negative samples are incorporated, which usually yields better model performance. However, in practice the labels often contain noise, and incorporating too many noisy negative samples into model training may be suboptimal. In this paper, we study how many negative samples are optimal for InfoNCE in different scenarios via a semi-quantitative theoretical framework. More specifically, we first propose a probabilistic model to analyze the influence of the negative sampling ratio K on training sample informativeness. Then, we design a training effectiveness function to measure the overall influence of training samples based on their informativeness. We estimate the optimal negative sampling ratio using the K value that maximizes the training effectiveness function. Based on our framework, we further propose an adaptive negative sampling method that can dynamically adjust the negative sampling ratio to improve InfoNCE-based model training. Extensive experiments in three different tasks show our framework can accurately predict the optimal negative sampling ratio, and various models can benefit from our adaptive negative sampling method.

Chit-chat has been shown effective in engaging users in human-computer interaction. We find with a user study that generating appropriate chit-chat for news articles can help expand user interest and increase the probability that a user reads a recommended news article. Based on this observation, we propose a method to generate personalized chit-chat for news recommendation. Different from existing methods for personalized text generation, our method only requires an external chat corpus obtained from an online forum, which can be disconnected from the recommendation dataset from both the user and item (news) perspectives. This is achieved by designing a weak supervision method for estimating users' personalized interest in a chit-chat post by transferring knowledge learned by a news recommendation model. Based on the method for estimating user interest, a reinforcement learning framework is proposed to generate personalized chit-chat. Extensive experiments, including the automatic offline evaluation and user studies, demonstrate the effectiveness of our method.

Federated learning (FL) is an important paradigm for training global models from decentralized data in a privacy-preserving way. Existing FL methods usually assume the global model can be trained on any participating client. However, in real applications, the devices of clients are usually heterogeneous, and have different computing power. Although big models like BERT have achieved huge success in AI, it is difficult to apply them to heterogeneous FL with weak clients. The straightforward solutions like removing the weak clients or using a small model to fit all clients will lead to some problems, such as under-representation of dropped clients and inferior accuracy due to data loss or limited model representation ability. In this work, we propose InclusiveFL, a client-inclusive federated learning method to handle this problem. The core idea of InclusiveFL is to assign models of different sizes to clients with different computing capabilities, bigger models for powerful clients and smaller ones for weak clients. We also propose an effective method to share the knowledge among local models with different sizes. In this way, all the clients can participate in FL training, and the final model can be big and powerful enough. Besides, we propose a momentum knowledge distillation method to better transfer knowledge in big models on powerful clients to the small models on weak clients. Extensive experiments on many real-world benchmark datasets demonstrate the effectiveness of InclusiveFL in learning accurate models from clients with heterogeneous devices under the FL framework.

Federated learning (FL) is a feasible technique to learn personalized recommendation models from decentralized user data. Unfortunately, federated recommender systems are vulnerable to poisoning attacks by malicious clients. Existing recommender system poisoning methods mainly focus on promoting the recommendation chances of target items due to financial incentives. In fact, in real-world scenarios, the attacker may also attempt to degrade the overall performance of recommender systems. However, existing general FL poisoning methods for degrading model performance are either ineffective or not concealed in poisoning federated recommender systems. In this paper, we propose a simple yet effective and covert poisoning attack method on federated recommendation, named FedAttack. Its core idea is using globally hardest samples to subvert model training. More specifically, the malicious clients first infer user embeddings based on local user profiles. Next, they choose the candidate items that are most relevant to the user embeddings as hardest negative samples, and find the candidates farthest from the user embeddings as hardest positive samples. The model gradients inferred from these poisoned samples are then uploaded for aggregation. Extensive experiments on two benchmark datasets show that FedAttack can effectively degrade the performance of various federated recommender systems, meanwhile cannot be effectively detected nor defended by many existing methods.

News recommendation calls for deep insights of news articles' underlying semantics. Therefore, pretrained language models (PLMs), like BERT and RoBERTa, may substantially contribute to the recommendation quality. However, it's extremely challenging to have news recommenders trained together with such big models: the learning of news recommenders requires intensive news encoding operations, whose cost is prohibitive if PLMs are used as the news encoder. In this paper, we propose a novel framework, SpeedyFeed, which efficiently trains PLMs-based news recommenders of superior quality. SpeedyFeed is highlighted for its light-weight encoding pipeline, which gives rise to three major advantages. Firstly, it makes the intermediate results fully reusable for the training workflow, which removes most of the repetitive but redundant encoding operations. Secondly, it improves the data efficiency of the training workflow, where non-informative data can be eliminated from encoding. Thirdly, it further saves the cost by leveraging simplified news encoding and compact news representation.

Previous works have validated that text generation APIs can be stolen through imitation attacks, causing IP violations. In order to protect the IP of text generation APIs, recent work has introduced a watermarking algorithm and utilized the null-hypothesis test as a post-hoc ownership verification on the imitation models. However, we find that it is possible to detect those watermarks via sufficient statistics of the frequencies of candidate watermarking words. To address this drawback, in this paper, we propose a novel Conditional wATERmarking framework (CATER) for protecting the IP of text generation APIs. An optimization method is proposed to decide the watermarking rules that can minimize the distortion of overall word distributions while maximizing the change of conditional word selections. Theoretically, we prove that it is infeasible for even the savviest attacker (they know how CATER works) to reveal the used watermarks from a large pool of potential word pairs based on statistical inspection. Empirically, we observe that high-order conditions lead to an exponential growth of suspicious (unused) watermarks, making our crafted watermarks more stealthy. In addition, CATER can effectively identify IP infringement under architectural mismatch and cross-domain imitation attacks, with negligible impairments on the generation quality of victim APIs. We envision our work as a milestone for stealthily protecting the IP of text generation APIs.

Vertical federated learning (VFL) is a privacy-preserving machine learning paradigm that can learn models from features distributed on different platforms in a privacy-preserving way. Since in real-world applications the data may contain bias on fairness-sensitive features (e.g., gender), VFL models may inherit bias from training data and become unfair for some user groups. However, existing fair machine learning methods usually rely on the centralized storage of fairness-sensitive features to achieve model fairness, which are usually inapplicable in federated scenarios. In this paper, we propose a fair vertical federated learning framework (FairVFL), which can improve the fairness of VFL models. The core idea of FairVFL is to learn unified and fair representations of samples based on the decentralized feature fields in a privacy-preserving way. Specifically, each platform with fairness-insensitive features first learns local data representations from local features. Then, these local representations are uploaded to a server and aggregated into a unified representation for the target task. In order to learn a fair unified representation, we send it to each platform storing fairness-sensitive features and apply adversarial learning to remove bias from the unified representation inherited from the biased data. Moreover, for protecting user privacy, we further propose a contrastive adversarial learning method to remove private information from the unified representation in server before sending it to the platforms keeping fairness-sensitive features. Experiments on three real-world datasets validate that our method can effectively improve model fairness with user privacy well-protected.

News recommendation aims to match news with personalized user interest. Existing methods for news recommendation usually model user interest from historical clicked news without the consideration of candidate news. However, each user usually has multiple interests, and it is difficult for these methods to accurately match a candidate news with a specific user interest. In this paper, we present a candidate-aware user modeling method for personalized news recommendation, which can incorporate candidate news into user modeling for better matching between candidate news and user interest. We propose a candidate-aware self-attention network that uses candidate news as clue to model candidate-aware global user interest. In addition, we propose a candidate-aware CNN network to incorporate candidate news into local behavior context modeling and learn candidate-aware short-term user interest. Besides, we use a candidate-aware attention network to aggregate previously clicked news weighted by their relevance with candidate news to build candidate-aware user representation. Experiments on real-world datasets show the effectiveness of our method in improving news recommendation performance.

User modeling is important for news recommendation. Existing methods usually first encode user's clicked news into news embeddings independently and then aggregate them into user embedding. However, the word-level interactions across different clicked news from the same user, which contain rich detailed clues to infer user interest, are ignored by these methods. In this paper, we propose a fine-grained and fast user modeling framework (FUM) to model user interest from fine-grained behavior interactions for news recommendation. The core idea of FUM is to concatenate the clicked news into a long document and transform user modeling into a document modeling task with both intra-news and inter-news word-level interactions. Since vanilla transformer cannot efficiently handle long document, we apply an efficient transformer named Fastformer to model fine-grained behavior interactions. Extensive experiments on two real-world datasets verify that FUM can effectively and efficiently model user interest for news recommendation.

News recommendation aims to help online news platform users find their preferred news articles. Existing news recommendation methods usually learn models from historical user behaviors on news. However, these behaviors are usually biased on news providers. Models trained on biased user data may capture and even amplify the biases on news providers, and are unfair for some minority news providers. In this paper, we propose a provider fairness-aware news recommendation framework (named ProFairRec), which can learn news recommendation models fair for different news providers from biased user data. The core idea of ProFairRec is to learn provider-fair news representations and provider-fair user representations to achieve provider fairness. To learn provider-fair representations from biased data, we employ provider-biased representations to inherit provider bias from data. Provider-fair and -biased news representations are learned from news content and provider IDs respectively, which are further aggregated to build fair and biased user representations based on user click history. All of these representations are used in model training while only fair representations are used for user-news matching to achieve fair news recommendation. Besides, we propose an adversarial learning task on news provider discrimination to prevent provider-fair news representation from encoding provider bias. We also propose an orthogonal regularization on provider-fair and -biased representations to better reduce provider bias in provider-fair representations. Moreover, ProFairRec is a general framework and can be applied to different news recommendation methods. Extensive experiments on a public dataset verify that our ProFairRec approach can effectively improve the provider fairness of many existing methods and meanwhile maintain their recommendation accuracy.

User modeling is critical for personalization. Existing methods usually train user models from task-specific labeled data, which may be insufficient. In fact, there are usually abundant unlabeled user behavior data that encode rich universal user information, and pre-training user models on them can empower user modeling in many downstream tasks. In this paper, we propose a user model pre-training method named UserBERT to learn universal user models on unlabeled user behavior data with two contrastive self-supervision tasks. The first one is masked behavior prediction and discrimination, aiming to model the contexts of user behaviors. The second one is behavior sequence matching, aiming to capture user interest stable in different periods. Besides, we propose a medium-hard negative sampling framework to select informative negative samples for better contrastive pre-training. Extensive experiments validate the effectiveness of UserBERT in user model pre-training.

News recommendation is often modeled as a sequential recommendation task, assuming there are rich short-term dependencies over historical clicked news. However, users usually have strong preferences on the temporal diversity of news information and may not tend to click similar news successively, which is very different from many sequential recommendation scenarios such as e-commerce recommendation. In this paper, we study whether news recommendation can be regarded as a standard sequential recommendation problem. Through extensive experiments on two real-world datasets, we find it suboptimal to model news recommendation as a conventional sequential recommendation problem. To handle this issue, we further propose a temporal diversity-aware sequential news recommendation method that can promote candidate news that are diverse from recently clicked news to help predict future clicks more accurately. Experiments show that our method can empower various news recommendation methods.

News representation is critical for news recommendation. Most existing methods learn news representations only from news texts while ignoring the visual information of news. In fact, users may click news not only due to the interest in news titles but also the attraction of news images. Thus, images are useful for representing news and predicting news clicks. Pretrained visiolinguistic models are powerful in multi-modal understanding, which can represent news from both textual and visual contents. In this paper, we propose a multimodal news recommendation method that can incorporate both textual and visual information of news to learn multimodal news representations. We first extract region-of-interests (ROIs) from news images via object detection. We then use a pre-trained visiolinguistic model to encode both news texts and image ROIs and model their inherent relatedness using co-attentional Transformers. In addition, we propose a crossmodal candidate-aware attention network to select relevant historical clicked news for the accurate modeling of user interest in candidate news. Experiments validate that incorporating multimodal news information can effectively improve the performance of news recommendation.

Accurate user interest modeling is important for news recommendation. Most existing methods for news recommendation rely on implicit feedbacks like click for inferring user interests and model training. However, click behaviors usually contain heavy noise, and cannot help infer complicated user interest such as dislike. Besides, the feed recommendation models trained solely on click behaviors cannot optimize other objectives such as user engagement. In this paper, we present a news feed recommendation method that can exploit various kinds of user feedbacks to enhance both user interest modeling and model training. We propose a unified user modeling framework to incorporate various explicit and implicit user feedbacks to infer both positive and negative user interests. In addition, we propose a strong-to-weak attention network that uses the representations of stronger feedbacks to distill positive and negative user interests from implicit weak feedbacks for accurate user interest modeling. Besides, we propose a multi-feedback model training framework to learn an engagement-aware feed recommendation model. Extensive experiments on a real-world dataset show that our approach can effectively improve the model performance in terms of both news clicks and user engagement.

Native ad is a popular type of online advertisement that has similar forms with the native content displayed on websites. Native ad click-through rate (CTR) prediction is useful for improving user experience and platform revenue. However, it is challenging due to the lack of explicit user intent, and user behaviors on the platform with native ads may be insufficient to infer users’ interest in ads. Fortunately, user behaviors exist on many online platforms that can provide complementary information for user-interest mining. Thus, leveraging multi-platform user behaviors is useful for native ad CTR prediction. However, user behaviors are highly privacy-sensitive, and the behavior data on different platforms cannot be directly aggregated due to user privacy concerns and data protection regulations. Existing CTR prediction methods usually require centralized storage of user behavior data for user modeling, which cannot be directly applied to the CTR prediction task with multi-platform user behaviors. In this article, we propose a federated native ad CTR prediction method named FedCTR, which can learn user-interest representations from cross-platform user behaviors in a privacy-preserving way. On each platform a local user model learns user embeddings from the local user behaviors on that platform. The local user embeddings from different platforms are uploaded to a server for aggregation, and the aggregated ones are sent to the ad platform for CTR prediction. Besides, we apply local differential privacy and differential privacy to the local and aggregated user embeddings, respectively, for better privacy protection. Moreover, we propose a federated framework for collaborative model training with distributed models and user behaviors. Extensive experiments on real-world dataset show that FedCTR can effectively leverage multi-platform user behaviors for native ad CTR prediction in a privacy-preserving manner.

News recommendation is important for online news services. Existing news recommendation models are usually learned from users' news click behaviors. Usually the behaviors of users with the same sensitive attributes (e.g., genders) have similar patterns and news recommendation models can easily capture these patterns. It may lead to some biases related to sensitive user attributes in the recommendation results, e.g., always recommending sports news to male users, which is unfair since users may not receive diverse news information. In this paper, we propose a fairness-aware news recommendation approach with decomposed adversarial learning and orthogonality regularization, which can alleviate unfairness in news recommendation brought by the biases of sensitive user attributes. In our approach, we propose to decompose the user interest model into two components. One component aims to learn a bias-aware user embedding that captures the bias information on sensitive user attributes, and the other aims to learn a bias-free user embedding that only encodes attribute-independent user interest information for fairness-aware news recommendation. In addition, we propose to apply an attribute prediction task to the bias-aware user embedding to enhance its ability on bias modeling, and we apply adversarial learning to the bias-free user embedding to remove the bias information from it. Moreover, we propose an orthogonality regularization method to encourage the bias-free user embeddings to be orthogonal to the bias-aware one to better distinguish the bias-free user embedding from the bias-aware one. For fairness-aware news ranking, we only use the bias-free user embedding. Extensive experiments on benchmark dataset show that our approach can effectively improve fairness in news recommendation with minor performance loss.

Tao Qi, Fangzhao Wu, Chuhan Wu, Yongfeng Huang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Tao Qi, Fangzhao Wu, Chuhan Wu, Peiru Yang, Yang Yu, Xing Xie, Yongfeng Huang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

News recommendation is critical for personalized news access.Most existing news recommendation methods rely on centralized storage of users' historical news click behavior data, which may lead to privacy concerns and hazards.Federated Learning is a privacy-preserving framework for multiple clients to collaboratively train models without sharing their private data.However, the computation and communication cost of directly learning many existing news recommendation models in a federated way are unacceptable for user clients.In this paper, we propose an efficient federated learning framework for privacy-preserving news recommendation.Instead of training and communicating the whole model, we decompose the news recommendation model into a large news model maintained in the server and a light-weight user model shared on both server and clients, where news representations and user model are communicated between server and clients.More specifically, the clients request the user model and news representations from the server, and send their locally computed gradients to the server for aggregation.The server updates its global user model with the aggregated gradients, and further updates its news model to infer updated news representations.Since the local gradients may contain private information, we propose a secure aggregation method to aggregate gradients in a privacy-preserving way.Experiments on two real-world datasets show that our method can reduce the computation and communication cost on clients while keep promising model performance.

Accurate user modeling is critical for news recommendation. Existing news recommendation methods usually model users' interest from their behaviors via sequential or attentive models. However, they cannot model the rich relatedness between user behaviors, which can provide useful contexts of these behaviors for user interest modeling. In this paper, we propose a novel user modeling approach for news recommendation, which models each user as a personalized heterogeneous graph built from user behaviors to better capture the fine-grained behavior relatedness. In addition, in order to learn user interest embedding from the personalized heterogeneous graph, we propose a novel heterogeneous graph pooling method, which can summarize both node features and graph topology, and be aware of the varied characteristics of different types of nodes. Experiments on large-scale benchmark dataset show the proposed methods can effectively improve the performance of user modeling for news recommendation.

Transformer has achieved great success in the NLP field by composing various advanced models like BERT and GPT.However, Transformer and its existing variants may not be optimal in capturing token distances because the position or distance embeddings used by these methods usually cannot keep the precise information of real distances, which may not be beneficial for modeling the orders and relations of contexts.In this paper, we propose DA-Transformer, which is a distanceaware Transformer that can exploit the real distance.We propose to incorporate the real distances between tokens to re-scale the raw self-attention weights, which are computed by the relevance between attention query and key.Concretely, in different self-attention heads the relative distance between each pair of tokens is weighted by different learnable parameters, which control the different preferences on long-or short-term information of these heads.Since the raw weighted real distances may not be optimal for adjusting selfattention weights, we propose a learnable sigmoid function to map them into re-scaled coefficients that have proper ranges.We first clip the raw self-attention weights via the ReLU function to keep non-negativity and introduce sparsity, and then multiply them with the rescaled coefficients to encode real distance information into self-attention.Extensive experiments on five benchmark datasets show that DA-Transformer can effectively improve the performance of many tasks and outperform the vanilla Transformer and its several variants.

The most important task in personalized news recommendation is accurate matching between candidate news and user interest. Most of existing news recommendation methods model candidate news from its textual content and user interest from their clicked news in an independent way. However, a news article may cover multiple aspects and entities, and a user usually has different kinds of interest. Independent modeling of candidate news and user interest may lead to inferior matching between news and users. In this paper, we propose a knowledge-aware interactive matching method for news recommendation. Our method interactively models candidate news and user interest to facilitate their accurate matching. We design a knowledge-aware news co-encoder to interactively learn representations for both clicked news and candidate news by capturing their relatedness in both semantic and entities with the help of knowledge graphs. We also design a user-news co-encoder to learn candidate news-aware user interest representation and user-aware candidate news representation for better interest matching. Experiments on two real-world datasets validate that our method can effectively improve the performance of news recommendation.

Recently, exploiting a knowledge graph (KG) to enrich the semantic representation of a news article have been proven to be effective for news recommendation. These solutions focus on the representation learning for news articles with additional information in the knowledge graph, where the user representations are mainly derived based on these news representations later. However, different users would hold different interests on the same news article. In other words, directly identifying the entities relevant to the user's interest and deriving the resultant user representation could enable a better news recommendation and explanation.

Personalized news recommendation is an essential technique for online news services. News articles usually contain rich textual content, and accurate news modeling is important for personalized news recommendation.

Product reviews can provide rich information about the opinions users have of products. However, it is nontrivial to effectively infer user preference and item characteristics from reviews due to the complicated semantic understanding. Existing methods usually learn features for users and items from reviews in single static fashions and cannot fully capture user preference and item features. In this article, we propose a neural review-based recommendation approach that aims to learn comprehensive representations of users/items under a three-tier attention framework. We design a review encoder to learn review features from words via a word-level attention, an aspect encoder to learn aspect features via a review-level attention, and a user/item encoder to learn the final representations of users/items via an aspect-level attention. In word- and review-level attentions, we adopt the context-aware mechanism to indicate importance of words and reviews dynamically instead of static attention weights. In addition, the attentions in the word and review levels are of multiple paradigms to learn multiple features effectively, which could indicate the diversity of user/item features. Furthermore, we propose a personalized aspect-level attention module in user/item encoder to learn the final comprehensive features. Extensive experiments are conducted and the results in rating prediction validate the effectiveness of our method.

Natural language inference (NLI) task requires an agent to determine the semantic relation between a premise sentence ( p) and a hypothesis sentence ( h), which demands sufficient understanding about sentences semantic. Due to the issues, such as polysemy, ambiguity, as well as fuzziness of sentences, intense sentence understanding is very challenging. To this end, in this article, we introduce the corresponding image of sentences as reference information for enhancing sentence semantic understanding and representing. Specifically, we first propose an image-enhanced multilevel sentence representation net (IEMLRN), that utilizes the image features from pretrained models for enhancing the sentence semantic understanding at different scales, i.e., lexical-level, phrase-level, and sentence-level. The proposed model advances the performance on NLI tasks by leveraging the pretrained global features of images. However, as these pretrained image features are optimized on specific image classification datasets, they may not show the best performance on NLI tasks. Therefore, we further propose to design an adaptive image feature generator that extracts fine-grained image labels from the corresponding sentences. After that, we extend the IEMLRN to multilevel image-enhanced sentence representation net (MIESR) by utilizing not only the coarse-grained pretrained features of images, but also the fine-grained adaptive features of images. Therefore, sentence semantic can be evaluated and enhanced more comprehensively and precisely. Extensive experiments on two benchmark datasets (SNLI, SICK) clearly show our proposed IEMLRN significantly outperform the state-of-the-art baselines, and our proposed MIESR model achieves the best performance by considering not only the text but also images in an adaptive multigranularities way.

Personalized news recommendation is a critical technology to improve users' online news reading experience.The core of news recommendation is accurate matching between user's interests and candidate news.The same user usually has diverse interests that are reflected in different news she has browsed.Meanwhile, important semantic features of news are implied in text segments of different granularities.Existing studies generally represent each user as a single vector and then match the candidate news vector, which may lose fine-grained information for recommendation.In this paper, we propose FIM, a Finegrained Interest Matching method for neural news recommendation.Instead of aggregating user's all historical browsed news into a unified vector, we hierarchically construct multilevel representations for each news via stacked dilated convolutions.Then we perform finegrained matching between segment pairs of each browsed news and the candidate news at each semantic level.High-order salient signals are then identified by resembling the hierarchy of image recognition for final click prediction.Extensive experiments on a real-world dataset from MSN news validate the effectiveness of our model on news recommendation.

Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, Ming Zhou. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. 2020.

Pooling is an important technique for learning text representations in many neural NLP models. In conventional pooling methods such as average, max and attentive pooling, text representations are weighted summations of the L1 or L∞ norm of input features. However, their pooling norms are always fixed and may not be optimal for learning accurate text representations in different tasks. In addition, in many popular pooling methods such as max and attentive pooling some features may be over-emphasized, while other useful ones are not fully exploited. In this paper, we propose an Attentive Pooling with Learnable Norms (APLN) approach for text representation. Different from existing pooling methods that use a fixed pooling norm, we propose to learn the norm in an end-to-end manner to automatically find the optimal ones for text representation in different tasks. In addition, we propose two methods to ensure the numerical stability of the model training. The first one is scale limiting, which re-scales the input to ensure non-negativity and alleviate the risk of exponential explosion. The second one is re-formulation, which decomposes the exponent operation to avoid computing the real-valued powers of the input and further accelerate the pooling operation. Experimental results on four benchmark datasets show that our approach can effectively improve the performance of attentive pooling.

Emotion-cause pair extraction, which aims at extracting both the emotion and its corresponding cause in text, is a significant and challenging task in emotion analysis. Previous work formulated the task in a two-step framework, i.e., emotion and cause extraction, and emotion-cause relation classification. However, different tasks may correlate with each other and the two-step framework does not fully exploit the interactions between tasks. In this paper, we propose a multi-task neural network to perform emotion-cause pair extraction in a unified model. The task of relation classification is learned together with emotion and cause extraction. To this end, we develop a method to obtain training samples for relation classification without the dependence on the result of emotion and cause extraction. To fully exploit the interactions between different tasks, our model shares useful features across tasks. Moreover, we propose a method to incorporate position-aware emotion information in cause extraction to further improve the performance. Experimental results show that our model outperforms the state-of-the-art model on emotion-cause pair extraction.

Modeling user interest is critical for accurate news recommendation. Existing news recommendation methods usually infer user interest from click behaviors on news. However, users may click a news article because attracted by its title shown on the news website homepage, but may not be satisfied with its content after reading. In many cases users close the news page quickly after click. In this paper we propose to model user interest from both click behaviors on news titles and reading behaviors on news content for news recommendation. More specifically, we propose a personalized reading speed metric to measure users’ satisfaction with news content. We learn embeddings of users from the news content they have read and their satisfaction with these news to model their interest in news content. In addition, we also learn another user embedding from the news titles they have clicked to model their preference in news titles. We combine both kinds of user embeddings into a unified user representation for news recommendation. We train the user representation model using two supervised learning tasks built from user behaviors, i.e., news title based click prediction and news content based satisfaction prediction, to encourage our model to recommend the news articles which not only are likely to be clicked but also have the content satisfied by the user. Experiments on real-world dataset show our method can effectively boost the performance of user modeling for news recommendation.

With the explosion of online news, personalized news recommendation becomes increasingly important for online news platforms to help their users find interesting information. Existing news recommendation methods achieve personalization by building accurate news representations from news content and user representations from their direct interactions with news (e.g., click), while ignoring the high-order relatedness between users and news. Here we propose a news recommendation method which can enhance the representation learning of users and news by modeling their relatedness in a graph setting. In our method, users and news are both viewed as nodes in a bipartite graph constructed from historical user click behaviors. For news representations, a transformer architecture is first exploited to build news semantic representations. Then we combine it with the information from neighbor news in the graph via a graph attention network. For user representations, we not only represent users from their historically clicked news, but also attentively incorporate the representations of their neighbor users in the graph. Improved performances on a large-scale real-world dataset validate the effectiveness of our proposed method.

Chatbots such as Xiaoice have gained huge popularity in recent years. Users frequently mention their favorite works such as songs and movies in conversations with chatbots. Detecting these entities can help design better chat strategies and improve user experience. Existing named entity recognition methods are mainly designed for formal texts, and their performance on the informal chatbot conversation texts may not be optimal. In addition, these methods rely on massive manually annotated data for model training. In this article, we propose a neural approach to detect entities of works for Chinese chatbot. Our approach is based on a language model (LM) long-short term memory (LSTM) convolutional neural network (CNN) conditional random value (CRF), or LM-LSTM-CNN-CRF, framework, which contains a language model to generate context-aware character embeddings, a Bi-LSTM network to learn contextual character representations from global contexts, a CNN to learn character representations from local contexts, and a CRF layer to jointly decode the character label sequence. In addition, we propose an automatic text annotation method via quote marks to reduce the effort of manual annotation. Besides, we propose an iterative data purification method to improve the quality of the automatically constructed labeled data. Massive experiments on a real-world dataset validate that our approach can achieve good performance on entity detection for Chinese chatbots.

Sentence semantic matching requires an agent to determine the semantic relation between two sentences, which is widely used in various natural language tasks such as Natural Language Inference (NLI) and Paraphrase Identification (PI). Among all matching methods, attention mechanism plays an important role in capturing the semantic relations and properly aligning the elements of two sentences. Previous methods utilized attention mechanism to select important parts of sentences at one time. However, the important parts of the sentence during semantic matching are dynamically changing with the degree of sentence understanding. Selecting the important parts at one time may be insufficient for semantic understanding. To this end, we propose a Dynamic Re-read Network (DRr-Net) approach for sentence semantic matching, which is able to pay close attention to a small region of sentences at each step and re-read the important words for better sentence semantic understanding. To be specific, we first employ Attention Stack-GRU (ASG) unit to model the original sentence repeatedly and preserve all the information from bottom-most word embedding input to up-most recurrent output. Second, we utilize Dynamic Re-read (DRr) unit to pay close attention to one important word at one time with the consideration of learned information and re-read the important words for better sentence semantic understanding. Extensive experiments on three sentence matching benchmark datasets demonstrate that DRr-Net has the ability to model sentence semantic more precisely and significantly improve the performance of sentence semantic matching. In addition, it is very interesting that some of finding in our experiments are consistent with the findings of psychological research.

Personalized news recommendation is important to help users find their interested news and improve reading experience.A key problem in news recommendation is learning accurate user representations to capture their interests.Users usually have both long-term preferences and short-term interests.However, existing news recommendation methods usually learn single representations of users, which may be insufficient.In this paper, we propose a neural news recommendation approach which can learn both long-and short-term user representations.The core of our approach is a news encoder and a user encoder.In the news encoder, we learn representations of news from their titles and topic categories, and use attention network to select important words.In the user encoder, we propose to learn long-term user representations from the embeddings of their IDs.In addition, we propose to learn short-term user representations from their recently browsed news via GRU network.Besides, we propose two methods to combine long-term and short-term user representations.The first one is using the long-term user representation to initialize the hidden state of the GRU network in short-term user representation.The second one is concatenating both long-and short-term user representations as a unified user vector.Extensive experiments on a real-world dataset show our approach can effectively improve the performance of neural news recommendation.

Aspect term extraction (ATE) aims at identifying all aspect terms in a sentence and is usually modeled as a sequence labeling problem.However, sequence labeling based methods cannot make full use of the overall meaning of the whole sentence and have the limitation in processing dependencies between labels.To tackle these problems, we first explore to formalize ATE as a sequence-tosequence (Seq2Seq) learning task where the source sequence and target sequence are composed of words and labels respectively.At the same time, to make Seq2Seq learning suit to ATE where labels correspond to words one by one, we design the gated unit networks to incorporate corresponding word representation into the decoder, and position-aware attention to pay more attention to the adjacent words of a target word.The experimental results on two datasets show that Seq2Seq learning is effective in ATE accompanied with our proposed gated unit networks and position-aware attention mechanism.

News recommendation can help users find interested news and alleviate information overload.The topic information of news is critical for learning accurate news and user representations for news recommendation.However, it is not considered in many existing news recommendation methods.In this paper, we propose a neural news recommendation approach with topic-aware news representations.The core of our approach is a topic-aware news encoder and a user encoder.In the news encoder we learn representations of news from their titles via CNN networks and apply attention networks to select important words.In addition, we propose to learn topic-aware news representations by jointly training the news encoder with an auxiliary topic classification task.In the user encoder we learn the representations of users from their browsed news and use attention networks to select informative news for user representation learning.Extensive experiments on a real-world dataset validate the effectiveness of our approach.

Estimating in-hospital costs from medical records is an important task with many applications such as accountable care. Existing methods for this task usually rely on manual feature engineering which needs massive domain knowledge, and do not exploit the textual information in medical records, e.g., diagnosis and operation texts. In this paper, we propose a neural in-hospital cost estimation (NICE) approach to estimate the in-hospital costs of patients from their admission records. Our approach can exploit the heterogeneous information in records, such as patient features, diagnosis/operation texts, and the diagnosis/operation IDs, via a multi-view learning framework. In addition, since different words, diagnoses and operations have different importance for cost estimation, we propose a hierarchical attention network to select important words, diagnoses and operations for learning informative record representations. Extensive experiments on a real-world medical dataset validate the effectiveness of our approach.

Sentiment classification is an important task in the sentiment analysis field. Many deep learning based sentiment classification methods have been proposed in recent years. However, these methods usually rely on massive labeled texts to train sentiment classifiers, which are expensive and time-consuming to annotate. Luckily, many high-quality sentiment lexicons have been constructed and can cover a large number of sentiment words. Since sentiment words are the basic units to convey sentiments in texts, these sentiment lexicons have the potential to improve the performance of neural sentiment classification. In this paper, we propose two approaches to exploit sentiment lexicons to enhance neural sentiment classification. In our first approach we use sentiment lexicons to learn sentiment-aware attentions. We propose a word sentiment classification task to classify the sentiments of words in a sentence based on their hidden representations in the attention network of neural sentiment classification models. We jointly train this task with neural sentiment classifier to facilitate the attention network to recognize and highlight sentiment-bearing words. In our second approach we use sentiment lexicons to learn sentiment-aware word embeddings. We design an auxiliary task to classify the sentiments of words in sentiment lexicons based on their word embeddings, and jointly train this task with neural sentiment classifier to encode sentiment information in sentiment lexicons to word embeddings. Extensive experiments on three benchmark datasets validate the effectiveness of our approach.

Review rating prediction is an important task in data mining and natural language processing fields, and has wide applications. Users usually express opinions towards many aspects in their reviews, and the overall review rating is a synthesis of these opinions. However, most existing review rating prediction methods ignore users' opinions on aspects, which is insufficient. In this paper, we propose a neural aspect-aware rating prediction approach for Chinese reviews. In our approach we propose a collaborative learning framework to jointly train review-level rating predictor and multiple aspect-level rating predictors. In our framework different rating predictors share the same review encoder model to exploit the inherent relatedness between them, but have different attention networks to focus on different informative texts for each task. The final review representation for rating prediction is a concatenation of the review representations from all predictors. Since word segmentation of Chinese reviews is usually inaccurate, we propose a multi-view learning model to learn review representations from both words and characters. Extensive experiments on real-world dataset validate the effectiveness of our approach.

Demographics of social media users such as gender are very important for personalized online services. However, the gender information of many users is usually not available. Luckily, the messages posted by social media users can provide rich clues for inferring their genders, since male and female users usually have differences in their message content. In addition, users with different genders often have different patterns in expressing emotions. In this paper, we propose a neural approach for gender prediction in social media based on both content and emotion of messages posted by users. The core of our approach is an emotion-aware hierarchical user representation model. Our model first learns message representations from words using message encoder and then learns user representations from messages using user encoder with hierarchical attention networks selecting important words and messages to learn informative user representations. In addition, we propose two methods to incorporate emotion information in messages into user representation learning. The first one is to incorporate emotion-aware message representations generated by a pre-trained emotion classifier into message representations. The second one is to train emotion-aware message encoders via jointly training our model with an auxiliary emotion classification task. Extensive experiments on two real-world datasets validate the effectiveness of our approach.

Neural network methods have achieved great success in sentiment classification. Recent studies have found that incorporating user and product information can effectively improve the performance of review sentiment classification. However, most of these studies only concentrate on the influence of users and products, ignoring the inherent correlation between users or products. This information is important for users or products since they can obtain more information from similar users or products. In this paper, we propose a novel framework for review rating prediction with user and product memory. First, besides the original user or product representations, we construct inferred representations from representative users or products which are stored in memory slots. These memory units can be viewed as refined knowledge representations of users or products learned from the data. Then, we employ two hierarchical networks with user attention and product attention using both the original and inferred representations. Experiments on benchmark datasets show that our method can achieve state-of-the-art performance. Besides, our approach performs much more better in cold-start scenarios where the training data is scarce.

Chuhan Wu, Fangzhao Wu, Mingxiao An, Tao Qi, Jianqiang Huang, Yongfeng Huang, Xing Xie. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Chuhan Wu, Fangzhao Wu, Suyu Ge, Tao Qi, Yongfeng Huang, Xing Xie. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Chuhan Wu, Fangzhao Wu, Tao Qi, Suyu Ge, Yongfeng Huang, Xing Xie. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Deep learning techniques have been widely applied to modern recommendation systems, bringing in flexible and effective ways of user representation. Conventionally, user representations are generated purely in the offline stage. Without referencing to the specific candidate item for recommendation, it is difficult to fully capture user preference from the perspective of interest. More recent algorithms tend to generate user representation at runtime, where user's historical behaviors are attentively summarized w.r.t. the presented candidate item. In spite of the improved efficacy, it is too expensive for many real-world scenarios because of the repetitive access to user's entire history. In this work, a novel user representation framework, Hi-Fi Ark, is proposed. With Hi-Fi Ark, user history is summarized into highly compact and complementary vectors in the offline stage, known as archives. Meanwhile, user preference towards a specific candidate item can be precisely captured via the attentive aggregation of such archives. As a result, both deployment feasibility and superior recommendation efficacy are achieved by Hi-Fi Ark. The effectiveness of Hi-Fi Ark is empirically validated on three real-world datasets, where remarkable and consistent improvements are made over a variety of well-recognized baseline methods.

Personalized news recommendation is very important for online news platforms to help users find interested news and improve user experience. News and user representation learning is critical for news recommendation. Existing news recommendation methods usually learn these representations based on single news information, e.g., title, which may be insufficient. In this paper we propose a neural news recommendation approach which can learn informative representations of users and news by exploiting different kinds of news information. The core of our approach is a news encoder and a user encoder. In the news encoder we propose an attentive multi-view learning model to learn unified news representations from titles, bodies and topic categories by regarding them as different views of news. In addition, we apply both word-level and view-level attention mechanism to news encoder to select important words and views for learning informative news representations. In the user encoder we learn the representations of users based on their browsed news and apply attention mechanism to select informative news for user representation learning. Extensive experiments on a real-world dataset show our approach can effectively improve the performance of news recommendation.

News recommendation is very important to help users find interested news and alleviate information overload. Different users usually have different interests and the same user may have various interests. Thus, different users may click the same news article with attention on different aspects. In this paper, we propose a neural news recommendation model with personalized attention (NPA). The core of our approach is a news representation model and a user representation model. In the news representation model we use a CNN network to learn hidden representations of news articles based on their titles. In the user representation model we learn the representations of users based on the representations of their clicked news articles. Since different words and different news articles may have different informativeness for representing news and users, we propose to apply both word- and news-level attention mechanism to help our model attend to important words and news articles. In addition, the same news article and the same word may have different informativeness for different users. Thus, we propose a personalized attention network which exploits the embedding of user ID to generate the query vector for the word- and news-level attentions. Extensive experiments are conducted on a real-world news recommendation dataset collected from MSN news, and the results validate the effectiveness of our approach on news recommendation.

Existing review-based recommendation methods usually use the same model to learn the representations of all users/items from reviews posted by users towards items. However, different users have different preference and different items have different characteristics. Thus, the same word or the similar reviews may have different informativeness for different users and items. In this paper we propose a neural recommendation approach with personalized attention to learn personalized representations of users and items from reviews. We use a review encoder to learn representations of reviews from words, and a user/item encoder to learn representations of users or items from reviews. We propose a personalized attention model, and apply it to both review and user/item encoders to select different important words and reviews for different users/items. Experiments on five datasets validate our approach can effectively improve the performance of neural recommendation.

Demographics of online users such as age and gender play an important role in personalized web applications. However, it is difficult to directly obtain the demographic information of online users. Luckily, search queries can cover many online users and the search queries from users with different demographics usually have some difference in contents and writing styles. Thus, search queries can provide useful clues for demographic prediction. In this paper, we study predicting users' demographics based on their search queries, and propose a neural approach for this task. Since search queries can be very noisy and many of them are not useful, instead of combining all queries together for user representation, in our approach we propose a hierarchical user representation with attention (HURA) model to learn informative user representations from their search queries. Our HURA model first learns representations for search queries from words using a word encoder, which consists of a CNN network and a word-level attention network to select important words. Then we learn representations of users based on the representations of their search queries using a query encoder, which contains a CNN network to capture the local contexts of search queries and a query-level attention network to select informative search queries for demographic prediction. Experiments on two real-world datasets validate that our approach can effectively improve the performance of search query based age and gender prediction and consistently outperform many baseline methods.

Twitter is a popular social media platform for information sharing and dissemination. Many Twitter users post tweets to share their experiences about drugs and adverse drug reactions. Automatic detection of tweets mentioning drug names and adverse drug reactions at a large scale has important applications such as pharmacovigilance. However, detecting drug name and adverse drug reaction mentioning tweets is very challenging, because tweets are usually very noisy and informal, and there are massive misspellings and user-created abbreviations for these mentions. In addition, these mentions are usually context dependent. In this paper, we propose a neural approach with hierarchical tweet representation and multi-head self-attention mechanism to jointly detect tweets mentioning drug names and adverse drug reactions. In order to alleviate the influence of massive misspellings and user-created abbreviations in tweets, we propose to use a hierarchical tweet representation model to first learn word representations from characters and then learn tweet representations from words. In addition, we propose to use multi-head self-attention mechanism to capture the interactions between words to fully model the contexts of tweets. Besides, we use additive attention mechanism to select the informative words to learn more informative tweet representations. Experimental results validate the effectiveness of our approach.

Chinese word segmentation (CWS) is very important for Chinese text processing. Existing methods for CWS usually rely on a large number of labeled sentences to train word segmentation models, which are expensive and time-consuming to annotate. Luckily, the unlabeled data is usually easy to collect and many high-quality Chinese lexicons are off-the-shelf, both of which can provide useful information for CWS. In this paper, we propose a neural approach for Chinese word segmentation which can exploit both lexicon and unlabeled data. Our approach is based on a variant of posterior regularization algorithm, and the unlabeled data and lexicon are incorporated into model training as indirect supervision by regularizing the prediction space of CWS models. Extensive experiments on multiple benchmark datasets in both in-domain and cross-domain scenarios validate the effectiveness of our approach.

Chinese named entity recognition (CNER) is an important task in Chinese natural language processing field. However, CNER is very challenging since Chinese entity names are highly context-dependent. In addition, Chinese texts lack delimiters to separate words, making it difficult to identify the boundary of entities. Besides, the training data for CNER in many domains is usually insufficient, and annotating enough training data for CNER is very expensive and time-consuming. In this paper, we propose a neural approach for CNER. First, we introduce a CNN-LSTM-CRF neural architecture to capture both local and long-distance contexts for CNER. Second, we propose a unified framework to jointly train CNER and word segmentation models in order to enhance the ability of CNER model in identifying entity boundaries. Third, we introduce an automatic method to generate pseudo labeled samples from existing labeled data which can enrich the training data. Experiments on two benchmark datasets show that our approach can effectively improve the performance of Chinese named entity recognition, especially when training data is insufficient.

Supervised learning methods are widely used in sentiment classification. However, when sentiment distribution is imbalanced, the performance of these methods declines. In this paper, we propose an effective approach for imbalanced sentiment classification. In our approach, multiple balanced subsets are sampled from the imbalanced training data and a multi-task learning based framework is proposed to learn robust sentiment classifier from these subsets collaboratively. In addition, we incorporate prior knowledge of sentiment expressions extracted from both existing sentiment lexicons and massive unlabeled data into our approach to enhance the learning of sentiment classifier in imbalanced scenario. Experimental results on benchmark datasets validate the effectiveness of our approach in improving imbalanced sentiment classification.

It is important to detect social spammers and spam messages in microblogging platforms. Existing methods usually handle the detection of social spammers and spam messages as two separate tasks using supervised learning techniques. However, labeled samples are usually scarce and manual annotation is expensive. In this paper, we propose a semi-supervised collaborative learning approach to jointly detect social spammers and spam messages in microblogging platforms. In our approach, the social spammer classifier and spam message classifier are collaboratively trained by exploiting the inherent relatedness between these tasks. In addition, unlabeled samples are incorporated into model training with the help of social contexts of users and messages. Experiments on real-world dataset show our approach can effectively improve the performance of both social spammer detection and spam message detection.

Sentence-level sentiment classification aims to mine fine-grained sentiment information from texts. Existing methods for this task are usually based on supervised learning and rely on massive labeled sentences for model training. However, annotating sufficient sentences is expensive and time-consuming. In this paper, we propose a neural sentence-level sentiment classification approach which can exploit heterogeneous sentiment supervision and reduce the dependence on labeled sentences. Besides the sentence-level supervision from labeled sentences, our approach can also incorporate the word-level supervision extracted from sentiment lexicons, document-level supervision extracted from labeled documents and sentiment relations between sentences extracted from unlabeled documents. A unified neural framework is proposed to fuse heterogeneous sentiment supervision to train sentence-level sentiment classification model. Experiments on benchmark datasets validate the effectiveness of our approach.

Natural Language Inference (NLI) task requires an agent to determine the semantic relation between a premise sentence (p) and a hypothesis sentence (h), which demands sufficient understanding about sentences from lexical knowledge to global semantic. Due to the issues such as polysemy, ambiguity, as well as fuzziness of sentences, fully understanding sentences is still challenging. To this end, we propose an Image-Enhanced Multi-Level Sentence Representation Net (IEMLRN), a novel architecture that is able to utilize the image to enhance the sentence semantic understanding at different scales. To be specific, we introduce the corresponding image of sentences as reference information, which can be helpful for sentence semantic understanding and inference relation evaluation. Since image information might be related to the sentence semantics at different scales, we design a multi-level architecture to understand sentences from different granularity and generate the sentence representation more precisely. Experimental results on the large-scale NLI corpus and real-world NLI alike corpus demonstrate that IEMLRN can simultaneously improve the performance. It is noteworthy that IEMLRN significantly outperforms the state-of-the-art sentence-encoding based models on the challenging hard subset and challenging lexical subset of SNLI corpus.

Domain adaptation is an important technology to handle domain dependence problem in sentiment analysis field.Existing methods usually rely on sentiment classifiers trained in source domains.However, their performance may heavily decline if the distributions of sentiment features in source and target domains have significant difference.In this paper, we propose an active sentiment domain adaptation approach to handle this problem.Instead of the source domain sentiment classifiers, our approach adapts the general-purpose sentiment lexicons to target domain with the help of a small number of labeled samples which are selected and annotated in an active learning mode, as well as the domain-specific sentiment similarities among words mined from unlabeled samples of target domain.A unified model is proposed to fuse different types of sentiment information and train sentiment classifier for target domain.Extensive experiments on benchmark datasets show that our approach can train accurate sentiment classifier with less labeled samples.

Sentence-level sentiment classification is important to understand users' fine-grained opinions. Existing methods for sentence-level sentiment classification are mainly based on supervised learning. However, it is difficult to obtain sentiment labels of sentences since manual annotation is expensive and time-consuming. In this paper, we propose an approach for sentence-level sentiment classification without the need of sentence labels. More specifically, we propose a unified framework to incorporate two types of weak supervision, i.e., document-level and word-level sentiment labels, to learn the sentence-level sentiment classifier. In addition, the contextual information of sentences and words extracted from unlabeled sentences is incorporated into our approach to enhance the learning of sentiment classifier. Experiments on benchmark datasets show that our approach can effectively improve the performance of sentence-level sentiment classification.

We propose a collaborative multi-domain sentiment classification approach to train sentiment classifiers for multiple domains simultaneously. In our approach, the sentiment information in different domains is shared to train more accurate and robust sentiment classifiers for each domain when labeled data is scarce. Specifically, we decompose the sentiment classifier of each domain into two components, a global one and a domain-specific one. The global model can capture the general sentiment knowledge and is shared by various domains. The domain-specific model can capture the specific sentiment expressions in each domain. In addition, we extract domain-specific sentiment knowledge from both labeled and unlabeled samples in each domain and use it to enhance the learning of domain-specific sentiment classifiers. Besides, we incorporate the similarities between domains into our approach as regularization over the domain-specific sentiment classifiers to encourage the sharing of sentiment information between similar domains. Two kinds of domain similarity measures are explored, one based on textual content and the other one based on sentiment expressions. Moreover, we introduce two efficient algorithms to solve the model of our approach. Experimental results on benchmark datasets show that our approach can effectively improve the performance of multi-domain sentiment classification and significantly outperform baseline methods.

Microblog sentiment classification is an interesting and important research topic with wide applications. Traditional microblog sentiment classification methods usually use a single model to classify the messages from different users and omit individuality. However, microblogging users frequently embed their personal character, opinion bias and language habits into their messages, and the same word may convey different sentiments in messages posted by different users. In this paper, we propose a personalized approach for microblog sentiment classification. In our approach, each user has a personalized sentiment classifier, which is decomposed into two components, a global one and a user-specific one. Our approach can capture the individual personality and at the same time leverage the common sentiment knowledge shared by all users. The personalized sentiment classifiers of massive users are trained in a collaborative way based on multi-task learning to handle the data sparseness problem. In addition, we incorporate users' social relations into our model to strengthen the learning of the personalized models. Moreover, we propose a distributed optimization algorithm to solve our model in parallel. Experiments on two real-world microblog sentiment datasets validate that our approach can improve microblog sentiment classification accuracy effectively and efficiently.

Domain adaptation is an important research topic in sentiment analysis area.Existing domain adaptation methods usually transfer sentiment knowledge from only one source domain to target domain.In this paper, we propose a new domain adaptation approach which can exploit sentiment knowledge from multiple source domains.We first extract both global and domain-specific sentiment knowledge from the data of multiple source domains using multi-task learning.Then we transfer them to target domain with the help of words' sentiment polarity relations extracted from the unlabeled target domain data.The similarities between target domain and different source domains are also incorporated into the adaptation process.Experimental results on benchmark dataset show the effectiveness of our approach in improving cross-domain sentiment classification performance.

Sentiment domain adaptation is widely studied to tackle the domain-dependence problem in sentiment analysis field. Existing domain adaptation methods usually train a sentiment classifier in a source domain and adapt it to the target domain using transfer learning techniques. However, when the sentiment feature distributions of the source and target domains are significantly different, the adaptation performance will heavily decline. In this paper, we propose a new sentiment domain adaptation approach by adapting the sentiment knowledge in general-purpose sentiment lexicons to a specific domain. Since the general sentiment words of general-purpose sentiment lexicons usually convey consistent sentiments in different domains, they have better generalization performance than the sentiment classifier trained in a source domain. In addition, we propose to extract various kinds of contextual sentiment knowledge from massive unlabeled samples in target domain and formulate them as sentiment relations among sentiment expressions. It can propagate the sentiment information in general sentiment words to massive domain-specific sentiment expressions. Besides, we propose a unified framework to incorporate these different kinds of sentiment knowledge and learn an accurate domain-specific sentiment classifier for target domain. Moreover, we propose an efficient optimization algorithm to solve the model of our approach. Extensive experiments on benchmark datasets validate the effectiveness and efficiency of our approach.

Microblog sentiment classification is an important research topic which has wide applications in both academia and industry. Because microblog messages are short, noisy and contain masses of acronyms and informal words, microblog sentiment classification is a very challenging task. Fortunately, collectively the contextual information about these idiosyncratic words provide knowledge about their sentiment orientations. In this paper, we propose to use the microblogs' contextual knowledge mined from a large amount of unlabeled data to help improve microblog sentiment classification. We define two kinds of contextual knowledge: word-word association and word-sentiment association. The contextual knowledge is formulated as regularization terms in supervised learning algorithms. An efficient optimization procedure is proposed to learn the model. Experimental results on benchmark datasets show that our method can consistently and significantly outperform the state-of-the-art methods.

The popularity of microblogging platforms, such as Twitter, makes them important for information dissemination and sharing. However, they are also recognized as ideal places by spammers to conduct social spamming. Massive social spammers and spam messages heavily hurt the user experience and hinder the healthy development of microblogging systems. Thus, effectively detecting the social spammers and spam messages in microblogging is of great value. Existing studies mainly regard social spammer detection and spam message detection as two separate tasks. However, social spammers and spam messages have strong connections, since social spammers tend to post more spam messages and spam messages have high probabilities to be posted by social spammers. Combining social spammer detection with spam message detection has the potential to boost the performance of each task. In this paper, we propose a unified framework for social spammer and spam message co-detection in microblogging. Our framework utilizes the posting relations between users and messages to combine social spammer detection with spam message detection. In addition, we extract the social relations between users as well as the connections between messages, and incorporate them into our framework as regularization terms over the prediction results. Besides, we introduce an efficient optimization method to solve our framework. Extensive experiments on a real-world microblog dataset demonstrate that our framework can significantly and consistently improve the performance of both social spammer detection and spam message detection.

Sentiment classification is a hot research topic in both industrial and academic fields. The mainstream sentiment classification methods are based on machine learning and treat sentiment classification as a text classification problem. However, sentiment classification is widely recognized as a highly domain-dependent task. The sentiment classifier trained in one domain may not perform well in another domain. A simple solution to this problem is training a domain-specific sentiment classifier for each domain. However, it is difficult to label enough data for every domain since they are in a large quantity. In addition, this method omits the sentiment information in other domains. In this paper, we propose to train sentiment classifiers for multiple domains in a collaborative way based on multi-task learning. Specifically, we decompose the sentiment classifier in each domain into two components, a general one and a domain-specific one. The general sentiment classifier can capture the global sentiment information and is trained across various domains to obtain better generalization ability. The domain-specific sentiment classifier is trained using the labeled data in one domain to capture the domain-specific sentiment information. In addition, we explore two kinds of relations between domains, one based on textual content and the other one based on sentiment word distribution. We build a domain similarity graph using domain relations and encode it into our approach as regularization over the domain-specific sentiment classifiers. Besides, we incorporate the sentiment knowledge extracted from sentiment lexicons to help train the general sentiment classifier more accurately. Moreover, we introduce an accelerated optimization algorithm to train the sentiment classifiers efficiently. Experimental results on two benchmark sentiment datasets show that our method can outperform baseline methods significantly and consistently.

This paper addresses the problem of post-processing of ranking in search, referred to as post ranking. Although important, no research seems to have been conducted on the problem, particularly with a principled approach, and in practice ad-hoc ways of performing the task are being adopted. This paper formalizes the problem as constrained optimization in which the constraints represent the post-processing rules and the objective function represents the trade-off between adherence to the original ranking and satisfaction of the rules. The optimization amounts to refining the original ranking result based on the rules. We further propose a specific probabilistic implementation of the general formalization on the basis of the Bradley-Terry model, which is theoretically sound, effective, and efficient. Our experimental results, using benchmark datasets and enterprise search dataset, show that the proposed method works much better than several baseline methods of utilizing rules.

It is important for many different applications such as government and business intelligence to analyze and explore the diffusion of public opinions on social media. However, the rapid propagation and great diversity of public opinions on social media pose great challenges to effective analysis of opinion diffusion. In this paper, we introduce a visual analysis system called OpinionFlow to empower analysts to detect opinion propagation patterns and glean insights. Inspired by the information diffusion model and the theory of selective exposure, we develop an opinion diffusion model to approximate opinion propagation among Twitter users. Accordingly, we design an opinion flow visualization that combines a Sankey graph with a tailored density map in one view to visually convey diffusion of opinions among many users. A stacked tree is used to allow analysts to select topics of interest at different levels. The stacked tree is synchronized with the opinion flow visualization to help users examine and compare diffusion patterns across topics. Experiments and case studies on Twitter data demonstrate the effectiveness and usability of OpinionFlow.

Correlated topical trend detection is very useful in analyzing public and social media influence. In this paper, we propose an algorithm that can both detect the correlation and discover the corresponding keywords that trigger the correlation. To detect the correlation, we use a projection vector to project two text streams onto the same space, and then use a least square cost function to regress one text stream over the other with different time lags. To extract the corresponding keywords, we impose the non-negative sparsity constraints over the projection parameters. In addition, we present an accelerated algorithm based on Nesterov's method to efficiently solve the optimization problem. In our experiments, we use both syntehtic and real data sets to demonstrate the advantages and capabilities of the proposed algorithm over CCA on the follower link prediction problem.