Weiqi Wang, Limeng Cui, Xin Liu, Sreyashi Nag, Wenju Xu, Chen Luo, Sheikh Muhammad Sarwar, Yang Li, Hansu Gu, Hui Liu, Changlong Yu, Jiaxin Bai, Yifan Gao, Haiyang Zhang, Qi He, Shuiwang Ji, Yangqiu Song. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Online shopping platforms, such as Amazon, offer services to billions of people worldwide. Unlike web search or other search engines, product search engines have their unique characteristics, primarily featuring short queries which are mostly a combination of product attributes and structured product search space. The uniqueness of product search underscores the crucial importance of the query understanding component. However, there are limited studies focusing on exploring this impact within real-world product search engines. In this work, we aim to bridge this gap by conducting a comprehensive study and sharing our year-long journey investigating how the query understanding service impacts Amazon Product Search. Firstly, we explore how query understanding-based ranking features influence the ranking process. Next, we delve into how the query understanding system contributes to understanding the performance of a ranking model. Building on the insights gained from our study on the evaluation of the query understanding-based ranking model, we propose a query understanding-based multi-task learning framework for ranking. We present our studies and investigations on Amazon Search.

Aiming at a better understanding of the search goals in the user search sessions, recent query recommender systems explicitly model the reformulations of queries, which hopes to estimate the intents behind these reformulations and thus benefit the next-query recommendation. However, in real-world e-commercial search scenarios, user intents are much more complicated and may evolve dynamically. Existing methods merely consider trivial reformulation intents from semantic aspects and fail to model dynamic reformulation intent flows in search sessions, leading to sub-optimal capacities to recommend desired queries. To deal with these limitations, we first explicitly define six types of query reformulation intents according to the desired products of two consecutive queries. We then apply two self-attentive encoders on top of two pre-trained large language models to learn the transition dynamics from semantic query and intent reformulation sequences, respectively. We develop an intent-aware query decoder to utilize the predicted intents for suggesting the next queries. We instantiate such a framework with an Intent-aware Variational AutoEncoder (IVAE) under deployment at Amazon. We conduct comprehensive experiments on two real-world e-commercial datasets from Amazon and one public dataset from BestBuy. Specifically, IVAE improves the Recall@15 by 25.44% and 60.47% on two Amazon datasets and 13.91% on BestBuy, respectively.

Graph Neural Networks (GNNs) have achieved great success in modeling graph-structured data. However, recent works show that GNNs are vulnerable to adversarial attacks which can fool the GNN model to make desired predictions of the attacker. In addition, training data of GNNs can be leaked under membership inference attacks. This largely hinders the adoption of GNNs in high-stake domains such as e-commerce, finance and bioinformatics. Though investigations have been made in conducting robust predictions and protecting membership privacy, they generally fail to simultaneously consider the robustness and membership privacy. Therefore, in this work, we study a novel problem of developing robust and membership privacy-preserving GNNs. Our analysis shows that Information Bottleneck (IB) can help filter out noisy information and regularize the predictions on labeled samples, which can benefit robustness and membership privacy. However, structural noises and lack of labels in node classification challenge the deployment of IB on graph-structured data. To mitigate these issues, we propose a novel graph information bottleneck framework that can alleviate structural noises with neighbor bottleneck. Pseudo labels are also incorporated in the optimization to minimize the gap between the predictions on the labeled set and unlabeled set for membership privacy. Extensive experiments on real-world datasets demonstrate that our method can give robust predictions and simultaneously preserve membership privacy.

Health thread recommendation methods aim to suggest the most relevant existing threads for a user. Most of the existing methods tend to rely on modeling the post contents to retrieve relevant answers. However, some posts written by users with different clinical conditions can be lexically similar, as unrelated diseases (e.g., Angina and Osteoporosis) may have the same symptoms (e.g., back pain), yet irrelevant threads to a user. Therefore, it is critical to not only consider the connections between users and threads, but also the descriptions of users' symptoms and clinical conditions. In this paper, towards this problem of thread recommendation in online healthcare forums, we propose a knowledge graph enhanced Threads Recommendation (KETCH) model, which leverages graph neural networks to model the interactions among users and threads, and learn their representations. In our model, the users, threads and posts are three types of nodes in a graph, linked through their associations. KETCH uses the message passing strategy by aggregating information along with the network. In addition, we introduce a knowledge-enhanced attention mechanism to capture the latent conditions and symptoms. We also apply the method to the task of predicting the side effects of drugs, to show that KETCH has the potential to complement the medical knowledge graph. Comparing with the best results of seven competing methods, in terms of MRR, KETCH outperforms all methods by at least 0.125 on the MedHelp dataset, 0.048 on the Patient dataset and 0.092 on HealthBoards dataset, respectively. We release the source code of KETCH at: https://github.com/cuilimeng/KETCH.

Human labeling is time-consuming and costly. This problem is further exacerbated in extremely imbalanced class label scenarios, such as detecting fraudsters in online websites. Active learning selects the most relevant example for human labelers to improve the model performance at a lower cost. However, existing methods for active learning for graph data often assumes that both data and label distributions are balanced. These assumptions fail in extreme rare-class classification scenarios, such as classifying abusive reviews in an e-commerce website.

Rare diseases affect hundreds of millions of people worldwide but are hard to detect since they have extremely low prevalence rates (varying from 1/1,000 to 1/200,000 patients) and are massively underdiagnosed. How do we reliably detect rare diseases with such low prevalence rates? How to further leverage patients with possibly uncertain diagnosis to improve detection? In this paper, we propose a Complementary pattern Augmentation (CONAN) framework for rare disease detection. CONAN combines ideas from both adversarial training and max-margin classification. It first learns self-attentive and hierarchical embedding for patient pattern characterization. Then, we develop a complementary generative adversarial networks (GAN) model to generate candidate positive and negative samples from the uncertain patients by encouraging a max-margin between classes. In addition, CONAN has a disease detector that serves as the discriminator during the adversarial training for identifying rare diseases. We evaluated CONAN on two disease detection tasks. For low prevalence inflammatory bowel disease (IBD) detection, CONAN achieved .96 precision recall area under the curve (PR-AUC) and 50.1% relative improvement over the best baseline. For rare disease idiopathic pulmonary fibrosis (IPF) detection, CONAN achieves .22 PR-AUC with 41.3% relative improvement over the best baseline.

To provide accurate and explainable misinformation detection, it is often useful to take an auxiliary source (e.g., social context and knowledge base) into consideration. Existing methods use social contexts such as users' engagements as complementary information to improve detection performance and derive explanations. However, due to the lack of sufficient professional knowledge, users seldom respond to healthcare information, which makes these methods less applicable. In this work, to address these shortcomings, we propose a novel knowledge guided graph attention network for detecting health misinformation better. Our proposal, named as DETERRENT, leverages on the additional information from medical knowledge graph by propagating information along with the network, incorporates a Medical Knowledge Graph and an Article-Entity Bipartite Graph, and propagates the node embeddings through Knowledge Paths. In addition, an attention mechanism is applied to calculate the importance of entities to each article, and the knowledge guided article embeddings are used for misinformation detection. DETERRENT addresses the limitation on social contexts in the healthcare domain and is capable of providing useful explanations for the results of detection. Empirical validation using two real-world datasets demonstrated the effectiveness of DETERRENT. Comparing with the best results of eight competing methods, in terms of F1 Score, DETERRENT outperforms all methods by at least 4.78% on the diabetes dataset and 12.79% on cancer dataset. We release the source code of DETERRENT at: https://github.com/cuilimeng/DETERRENT.

How to effectively detect fake news and prevent its diffusion on social media has gained much attention in recent years. However, relatively little focus has been given on exploiting user comments left for posts and latent sentiments therein in detecting fake news. Inspired by the rich information available in user comments on social media, therefore, we investigate whether the latent sentiments hidden in user comments can potentially help distinguish fake news from reliable content. We incorporate users' latent sentiments into an end-to-end deep embedding framework for detecting fake news, named as SAME. First, we use multi-modal networks to deal with heterogeneous data modalities. Second, to learn semantically meaningful spaces per data source, we adopt an adversarial mechanism. Third, we define a novel regularization loss to bring embeddings of relevant pairs closer. Our comprehensive validation using two real-world datasets, PolitiFact and GossipCop, demonstrates the effectiveness of SAME in detecting fake news, significantly outperforming state-of-the-art methods.

Despite recent advancements in computationally detecting fake news, we argue that a critical missing piece be the explainability of such detection--i.e., why a particular piece of news is detected as fake--and propose to exploit rich information in users' comments on social media to infer the authenticity of news. In this demo paper, we present our system for an explainable fake news detection called dEFEND, which can detect the authenticity of a piece of news while identifying user comments that can explain why the news is fake or real. Our solution develops a sentence-comment co-attention sub-network to exploit both news contents and user comments to jointly capture explainable top-k check-worthy sentences and user comments for fake news detection. The system is publicly accessible.

In recent years, to mitigate the problem of fake news, computational detection of fake news has been studied, producing some promising early results. While important, however, we argue that a critical missing piece of the study be the explainability of such detection, i.e., why a particular piece of news is detected as fake. In this paper, therefore, we study the explainable detection of fake news. We develop a sentence-comment co-attention sub-network to exploit both news contents and user comments to jointly capture explainable top-k check-worthy sentences and user comments for fake news detection. We conduct extensive experiments on real-world datasets and demonstrate that the proposed method not only significantly outperforms 7 state-of-the-art fake news detection methods by at least 5.33% in F1-score, but also (concurrently) identifies top-k user comments that explain why a news piece is fake, better than baselines by 28.2% in NDCG and 30.7% in Precision.

Cross-modal retrieval, which takes text queries to retrieve relevant images or vice versa, has drawn much attention in recent years. This topic exhibits dual-heterogeneity: heterogeneity of different modalities and heterogeneous features obtained from multiple views. To address this issue, we propose an effective multi-view fusion method for cross-modal retrieval based on tensor modeling (CMTM) for cross-modal retrieval from the full-order feature interactions within the multimodal data. In order to facilitate integration of heterogeneous features from multiple views, we adopt the tensor structure to model the full-order interactions among the multi-view features effectively. Besides, a tensor factorization is applied to derive model parameters. Extensive experiments demonstrate the effectiveness of CMTM on cross-modal retrieval.

Multimedia data available in various disciplines are usually heterogeneous, containing representations in multi-views, where the cross-modal search techniques become necessary and useful. It is a challenging problem due to the heterogeneity of data with multiple modalities, multi-views in each modality and the diverse data categories. In this paper, we propose a novel multi-view cross-modal hashing method named Multi-view Collective Tensor Decomposition (MCTD) to fuse these data effectively, which can exploit the complementary feature extracted from multi-modality multi-view while simultaneously discovering multiple separated subspaces by leveraging the data categories as supervision information. Our contributions are summarized as follows: 1) we exploit tensor modeling to get better representation of the complementary features and redefine a latent representation space; 2) a block-diagonal loss is proposed to explicitly pursue a more discriminative latent tensor space by exploring supervision information; 3) we propose a new feature projection method to characterize the data and to generate the latent representation for incoming new queries. An optimization algorithm is proposed to solve the objective function designed for MCTD, which works under an iterative updating procedure. Experimental results prove the state-of-the-art precision of MCTD compared with competing methods.

In large-scale learning problem, the scalability of learning algorithms is usually the key factor affecting the algorithm practical performance, which is determined by both the time complexity of the learning algorithms and the amount of supervision information (i.e., labeled data). Learning with label proportions (LLP) is a new kind of machine learning problem which has drawn much attention in recent years. Different from the well-known supervised learning, LLP can estimate a classifier from groups of weakly labeled data, where only the positive/negative class proportions of each group are known. Due to its weak requirements for the input data, LLP presents a variety of real-world applications in almost all the fields involving anonymous data, like computer vision, fraud detection and spam filtering. However, even through the required labeled data is of a very small amount, LLP still suffers from the long execution time a lot due to the high time complexity of the learning algorithm itself. In this paper, we propose a very fast learning method based on inversing output scaling process and extreme learning machine, namely Inverse Extreme Learning Machine (IELM), to address the above issues. IELM can speed up the training process by order of magnitudes for large datasets, while achieving highly competitive classification accuracy with the existing methods at the same time. Extensive experiments demonstrate the significant speedup of the proposed method. We also demonstrate the feasibility of IELM with a case study in real-world setting: modeling image attributes based on ImageNet Object Attributes dataset.

Employees in companies can be divided into different social communities, and those who frequently socialize with each other will be treated as close friends and are grouped in the same community. In the enterprise context, a large amount of information about the employees is available in both (1) offline company internal sources and (2) online enterprise social networks (ESNs). Each of the information sources also contain multiple categories of employees' socialization activities at the same time. In this paper, we propose to detect the social communities of the employees in companies based on the broad learning setting with both these online and offline information sources simultaneously, and the problem is formally called the "Broad Learning based Enterprise Community Detection" (BL-ECD) problem. To address the problem, a novel broad learning based community detection framework named "HeterogeneoUs Multi-sOurce ClusteRing" (HUMOR) is introduced in this paper. Based on the various enterprise social intimacy measures introduced in this paper, HUMOR detects a set of micro community structures of the employees based on each of the socialization activities respectively. To obtain the (globally) consistent community structure of employees in the company, HUMOR further fuses these micro community structures via two broad learning phases: (1) intra-fusion of micro community structures to obtain the online and offline (locally) consistent communities respectively, and (2) inter-fusion of the online and offline communities to achieve the (globally) consistent community structure of employees. Extensive experiments conducted on real-world enterprise datasets demonstrate our method can perform very well in addressing the BL-ECD problem.

Network embedding aims at projecting the network data into a low-dimensional feature space, where the nodes are represented as a unique feature vector and network structure can be effectively preserved. In recent years, more and more online application service sites can be represented as massive and complex networks, which are extremely challenging for traditional machine learning algorithms to deal with. Effective embedding of the complex network data into low-dimension feature representation can both save data storage space and enable traditional machine learning algorithms applicable to handle the network data. Network embedding performance will degrade greatly if the networks are of a sparse structure, like the emerging networks with few connections. In this paper, we propose to learn the embedding representation for a target emerging network based on the broad learning setting, where the emerging network is aligned with other external mature networks at the same time. To solve the problem, a new embedding framework, namely "Deep alIgned autoencoder based eMbEdding" (DIME), is introduced in this paper. DIME handles the diverse link and attribute in a unified analytic based on broad learning, and introduces the multiple aligned attributed heterogeneous social network concept to model the network structure. A set of meta paths are introduced in the paper, which define various kinds of connections among users via the heterogeneous link and attribute information. The closeness among users in the networks are defined as the meta proximity scores, which will be fed into DIME to learn the embedding vectors of users in the emerging network. Extensive experiments have been done on real-world aligned social networks, which have demonstrated the effectiveness of DIME in learning the emerging network embedding vectors.

Cracks are a growing threat to road conditions and have drawn much attention to the construction of intelligent transportation systems. However, as the key part of an intelligent transportation system, automatic road crack detection has been challenged because of the intense inhomogeneity along the cracks, the topology complexity of cracks, the inference of noises with similar texture to the cracks, and so on. In this paper, we propose CrackForest, a novel road crack detection framework based on random structured forests, to address these issues. Our contributions are shown as follows: 1) apply the integral channel features to redefine the tokens that constitute a crack and get better representation of the cracks with intensity inhomogeneity; 2) introduce random structured forests to generate a high-performance crack detector, which can identify arbitrarily complex cracks; and 3) propose a new crack descriptor to characterize cracks and discern them from noises effectively. In addition, our method is faster and easier to parallel. Experimental results prove the state-of-the-art detection precision of CrackForest compared with competing methods.