Identifying harmful memes is challenging due to their implicit meanings, which are not always evident from texts and images alone. Existing solutions often lack clear explanations to justify their decisions. To address this gap, we propose an explainable approach,ExplainHM++, which detects harmful memes by reasoning over competing rationales from both harmful and harmless perspectives. First, inspired by the capabilities of Large Multimodal Models (LMMs) in text generation and multimodal reasoning, we developExplainHM, a one-stage multimodal debate in which LMMs generate explanations through contradictory arguments. Second, we fine-tune a small language model to serve as a judge in the debate, improving the integration of harmfulness rationales with the multimodal content of memes. However, we observe that a naive multimodal debate remains vulnerable, as it heavily depends on the inherent reasoning ability of LMMs to understand the memes. Given the evolving and noisy nature of memes, we further introduce a meme sample retrieval mechanism and a retrieval-augmented debate paradigm to strengthen and refine LMM-generated explanations. Extensive experiments on three public meme datasets demonstrate thatExplainHM++not only outperforms state-of-the-art methods but also provides superior, interpretable explanations for harmful meme detection.

The proliferation of misinformation, such as rumors on social media, has drawn significant attention, prompting various expressions of stance among users. Although rumor detection and stance detection are distinct tasks, they can complement each other. Rumors can be identified by cross-referencing stances in related posts, and stances are influenced by the nature of the rumor. However, existing stance detection methods often require post-level stance annotations, which are costly to obtain. We propose a novel LLM-enhanced MIL approach to jointly predict post stance and claim class labels, supervised solely by claim labels, using an undirected microblog propagation model. Our weakly supervised approach relies only on bag-level labels of claim veracity, aligning with multi-instance learning (MIL) principles. To achieve this, we transform the multi-class problem into multiple MIL-based binary classification problems. We then employ a discriminative attention layer to aggregate the outputs from these classifiers into finer-grained classes. Experiments conducted on three rumor datasets and two stance datasets demonstrate the effectiveness of our approach, highlighting strong connections between rumor veracity and expressed stances in responding posts. Our method shows promising performance in joint rumor and stance detection compared to the state-of-the-art methods.

The age of social media is flooded with Internet memes, necessitating a clear grasp and effective identification of harmful ones. This task presents a significant challenge due to the implicit meaning embedded in memes, which is not explicitly conveyed through the surface text and image. However, existing harmful meme detection methods do not present readable explanations that unveil such implicit meaning to support their detection decisions. In this paper, we propose an explainable approach to detect harmful memes, achieved through reasoning over conflicting rationales from both harmless and harmful positions. Specifically, inspired by the powerful capacity of Large Language Models (LLMs) on text generation and reasoning, we first elicit multimodal debate between LLMs to generate the explanations derived from the contradictory arguments. Then we propose to fine-tune a small language model as the debate judge for harmfulness inference, to facilitate multimodal fusion between the harmfulness rationales and the intrinsic multimodal information within memes. In this way, our model is empowered to perform dialectical reasoning over intricate and implicit harm-indicative patterns, utilizing multimodal explanations originating from both harmless and harmful arguments. Extensive experiments on three public meme datasets demonstrate that our harmful meme detection approach achieves much better performance than state-of-the-art methods and exhibits a superior capacity for explaining the meme harmfulness of the model predictions.

Conversational Recommender Systems (CRS) has become an emerging research topic seeking to perform recommendations through interactive conversations, which generally consist of generation and recommendation modules. Prior work on CRS tends to incorporate more external and domain-specific knowledge like item reviews to enhance performance. Despite the fact that the collection and annotation of theexternal domain-specificinformation needs much human effort and degenerates the generalizability, too much extra knowledge introduces more difficulty to balance among them. Therefore, we propose to fully discover and extract theinternalknowledge from the context. We capture both entity-level and contextual-level representations to jointly model user preferences for the recommendation, where a time-aware attention is designed to emphasize the recently appeared items in entity-level representations. We further use the pre-trained BART to initialize the generation module to alleviate the data scarcity and enhance the context modeling. In addition to conducting experiments on a popular dataset (ReDial), we also include a multi-domain dataset (OpenDialKG) to show the effectiveness of our model. Experiments on both datasets show that our model achieves better performance on most evaluation metrics with less external knowledge and generalizes well to other domains. Additional analyses on the recommendation and generation tasks demonstrate the effectiveness of our model in different scenarios.

Rumors can cause devastating consequences to individuals and our society. Analysis shows that the widespread of rumors typically results from deliberate promotion of information with unknown veracity aiming to shape the collective public opinions on the concerned news event. In this paper, we attempt to combat such chaotic phenomenon with a countermeasure by mirroring against how such chaos is created in order to make automatic rumor detection more robust and effective. Our idea is inspired by adversarial learning method originated from Generative Adversarial Networks (GAN). We propose a GAN-style approach, where a generator is designed to produce uncertain or conflicting voices, further polarizing the original conversation threads with the intention of pressurizing the discriminator to learn stronger rumor indicative features from the augmented, more challenging examples. We reveal that feature learning effectiveness is highly relevant to the quality of generated parody, viz., how hard it is to get distinguished from real posts. Given the strong natural language generation performance of transformer, we propose a transformer-based method to improve the generated posts, so that they appear to be closely responsive to the source post and retain the authentic propagation structure and context of information. Different from traditional data-driven approach to rumor detection, our method can capture low-frequency but more salient non-trivial discriminant patterns via adversarial training. Extensive experiments on THREE benchmark datasets demonstrate that our rumor detection methods and the transformer-based model achieve much better results than state-of-the-art methods.

The diffusion of rumors on social media generally follows a propagation tree structure, which provides valuable clues on how an original message is transmitted and responded by users over time. Recent studies reveal that rumor verification and stance detection are two relevant tasks that can jointly enhance each other despite their differences. For example, rumors can be debunked by cross-checking the stances conveyed by their relevant posts, and stances are also conditioned on the nature of the rumor. However, stance detection typically requires a large training set of labeled stances at post level, which are rare and costly to annotate. Enlightened by Multiple Instance Learning (MIL) scheme, we propose a novel weakly supervised joint learning framework for rumor verification and stance detection which only requires bag-level class labels concerning the rumor's veracity. Specifically, based on the propagation trees of source posts, we convert the two multi-class problems into multiple MIL-based binary classification problems where each binary model is focused on differentiating a target class (of rumor or stance) from the remaining classes. Then, we propose a hierarchical attention mechanism to aggregate the binary predictions, including (1) a bottom-up/top-down tree attention layer to aggregate binary stances into binary veracity; and (2) a discriminative attention layer to aggregate the binary class into finer-grained classes. Extensive experiments conducted on three Twitter-based datasets demonstrate promising performance of our model on both claim-level rumor detection and post-level stance classification compared with state-of-the-art methods.

We study the task of span-level emotion cause analysis (SECA), which is focused on identifying the specific emotion cause span(s) triggering a certain emotion in the text. Compared to the popular clause-level emotion cause analysis (CECA), it is a finer-grained emotion cause analysis (ECA) task. In this paper, we design a BERT-based graph attention network for emotion cause span(s) identification. The proposed model takes advantage of the structure of BERT to capture the relationship information between emotion and text, and utilizes graph attention network to model the structure information of the text. Our SECA method can be easily used for extracting clause-level emotion causes for CECA as well. Experimental results show that the proposed method consistently outperforms the state-of-the-art ECA methods on benchmark emotion cause dataset.

This paper addresses the task of span-level emotion cause analysis (SECA). It is a finer-grained emotion cause analysis (ECA) task, which aims to identify the specific emotion cause span(s) behind certain emotions in text. In this paper, we formalize SECA as a sequence tagging task for which several variants of neural network-based sequence tagging models to extract specific emotion cause span(s) in the given context. These models combine different types of encoding and decoding approaches. Furthermore, to make our models more "emotionally sensitive'', we utilize the multi-head attention mechanism to enhance the representation of context. Experimental evaluations conducted on two benchmark datasets demonstrate the effectiveness of the proposed models.

In open-domain dialogue systems, knowledge information such as unstructured persona profiles, text descriptions and structured knowledge graph can help incorporate abundant background facts for delivering more engaging and informative responses. Existing studies attempted to model a general posterior distribution over candidate knowledge by considering the entire response utterance as a whole at the beginning of decoding process for knowledge selection. However, a single smooth distribution could fail to model the variability of knowledge selection patterns over different decoding steps, and make the knowledge expression less consistent. To remedy this issue, we propose an adaptive posterior knowledge selection framework, which sequentially introduces a series of discriminative distributions to dynamically control when and what knowledge should be used in specific decoding steps. The adaptive distributions can also capture knowledge-relevant semantic dependencies between adjacent words to refine response generation. In particular, for knowledge graph-grounded dialogue generation, we further incorporate the adaptive distributions into generative word distributions to help express the knowledge entity words. The experimental results show that our developed methods outperform strong baseline systems by large margins.

In open-domain dialogue systems, dialogue cues such as emotion, persona, and emoji can be incorporated into conversation models for strengthening the semantic relevance of generated responses. Existing neural response generation models either incorporate dialogue cue into decoder’s initial state or embed the cue indiscriminately into the state of every generated word, which may cause the gradients of the embedded cue to vanish or disturb the semantic relevance of generated words during back propagation. In this paper, we propose a Cue Adaptive Decoder (CueAD) that aims to dynamically determine the involvement of a cue at each generation step in the decoding. For this purpose, we extend the Gated Recurrent Unit (GRU) network with an adaptive cue representation for facilitating cue incorporation, in which an adaptive gating unit is utilized to decide when to incorporate cue information so that the cue can provide useful clues for enhancing the semantic relevance of the generated words. Experimental results show that CueAD outperforms state-of-the-art baselines with large margins.

Rumor spread in social media severely jeopardizes the credibility of online content. Thus, automatic debunking of rumors is of great importance to keep social media a healthy environment. While facing a dubious claim, people often dispute its truthfulness sporadically in their posts containing various cues, which can form useful evidence with long-distance dependencies. In this work, we propose to learn discriminative features from microblog posts by following their non-sequential propagation structure and generate more powerful representations for identifying rumors. For modeling non-sequential structure, we first represent the diffusion of microblog posts with propagation trees, which provide valuable clues on how a claim in the original post is transmitted and developed over time. We then present a bottom-up and a top-down tree-structured models based on Recursive Neural Networks (RvNN) for rumor representation learning and classification, which naturally conform to the message propagation process in microblogs. To enhance the rumor representation learning, we reveal that effective rumor detection is highly related to finding evidential posts, e.g., the posts expressing specific attitude towards the veracity of a claim, as an extension of the previous RvNN-based detection models that treat every post equally. For this reason, we design discriminative attention mechanisms for the RvNN-based models to selectively attend on the subset of evidential posts during the bottom-up/top-down recursive composition. Experimental results on four datasets collected from real-world microblog platforms confirm that (1) our RvNN-based models achieve much better rumor detection and classification performance than state-of-the-art approaches; (2) the attention mechanisms for focusing on evidential posts can further improve the performance of our RvNN-based method; and (3) our approach possesses superior capacity on detecting rumors at a very early stage.

Understanding the spread of false information in social networks has gained a lot of recent attention. In this paper, we explore the role community structures play in determining how people get exposed to fake news. Inspired by approaches in epidemiology, we propose a novel Community Health Assessment model, whose goal is to understand the vulnerability of communities to fake news spread. We define the concepts of neighbor, boundary and core nodes of a community and propose appropriate metrics to quantify the vulnerability of nodes (individual-level) and communities (group-level) to spreading fake news. We evaluate our model on communities identified using three popular community detection algorithms for twelve real-world news spreading networks collected from Twitter. Experimental results show that the proposed metrics perform significantly better on the fake news spreading networks than on the true news, indicating that our community health assessment model is effective.

Rumors can cause devastating consequences to individual and/or society. Analysis shows that widespread of rumors typically results from deliberately promoted information campaigns which aim to shape collective opinions on the concerned news events. In this paper, we attempt to fight such chaos with itself to make automatic rumor detection more robust and effective. Our idea is inspired by adversarial learning method originated from Generative Adversarial Networks (GAN). We propose a GAN-style approach, where a generator is designed to produce uncertain or conflicting voices, complicating the original conversational threads in order to pressurize the discriminator to learn stronger rumor indicative representations from the augmented, more challenging examples. Different from traditional data-driven approach to rumor detection, our method can capture low-frequency but stronger non-trivial patterns via such adversarial training. Extensive experiments on two Twitter benchmark datasets demonstrate that our rumor detection method achieves much better results than state-of-the-art methods.

Ubiquitous use of social media such as microblogging platforms brings about ample opportunities for the false information to diffuse online. It is very important not just to determine the veracity of information but also the authenticity of the users who spread the information, especially in time-critical situations like real-world emergencies, where urgent measures have to be taken for stopping the spread of fake information. In this work, we propose a novel machine learning based approach for automatic identification of the users spreading rumorous information by leveraging the concept of believability, i.e., the extent to which the propagated information is likely to be perceived as truthful, based on the trust measures of users in Twitter's retweet network. We hypothesize that the believability between two users is proportional to the trustingness of the retweeter and the trustworthiness of the tweeter, which are two complementary measures of user trust and can be inferred from retweeting behaviors using a variant of HITS algorithm. With the retweet network edge-weighted by believability scores, we use network representation learning to generate user embeddings, which are then leveraged to classify users into as rumor spreaders or not. Based on experiments on a very large real-world rumor dataset collected from Twitter, we demonstrate that our method can effectively identify rumor spreaders and outperform four strong baselines with large margin.

In recent years there has been a growing interest in text quantification, a supervised learning task where the goal is to accurately estimate, in an unlabelled set of items, the prevalence (or "relative frequency") of each class c in a predefined set C. Text quantification has several applications, and is a dominant concern in fields such as market research, the social sciences, political science, and epidemiology. In this paper we tackle, for the first time, the problem of ordinal text quantification, defined as the task of performing text quantification when a total order is defined on the set of classes; estimating the prevalence of "five stars" reviews in a set of reviews of a given product, and monitoring this prevalence across time, is an example application. We present OQT, a novel tree-based OQ algorithm, and discuss experimental results obtained on a dataset of tweets classified according to sentiment strength.

In the research of building emotion lexicons, we witness the exploitation of crowd-sourced affective annotation given by readers of online news articles. Such approach ignores the relationship between topics and emotion expressions which are often closely correlated. We build an emotion lexicon by developing a novel joint non-negative matrix factorization model which not only incorporates crowd-annotated emotion labels of articles but also generates the lexicon using the topic-specific matrices obtained from the factorization process. We evaluate our lexicon via emotion classification on both benchmark and built-in-house datasets. Results demonstrate the high-quality of our lexicon.

Sentiment classification has become a ubiquitous enabling technology in the Twittersphere, since classifying tweets according to the sentiment they convey towards a given entity (be it a product, a person, a political party, or a policy) has many applications in political science, social science, market research, and many others. In this paper we contend that most previous studies dealing with tweet sentiment classification (TSC) use a suboptimal approach. The reason is that the final goal of most such studies is not estimating the class label (e.g., Positive, Negative, or Neutral) of individual tweets, but estimating the relative frequency (a.k.a. "prevalence") of the different classes in the dataset. The latter task is called quantification, and recent research has convincingly shown that it should be tackled as a task of its own, using learning algorithms and evaluation measures different from those used for classification. In this paper we show, on a multiplicity of TSC datasets, that using a quantification-specific algorithm produces substantially better class frequency estimates than a state-of-the-art classification-oriented algorithm routinely used in TSC. We thus argue that researchers interested in tweet sentiment prevalence should switch to quantification-specific (instead of classification-specific) learning algorithms and evaluation measures.

Automatically identifying rumors from online social media especially microblogging websites is an important research issue. Most of existing work for rumor detection focuses on modeling features related to microblog contents, users and propagation patterns, but ignore the importance of the variation of these social context features during the message propagation over time. In this study, we propose a novel approach to capture the temporal characteristics of these features based on the time series of rumor's lifecycle, for which time series modeling technique is applied to incorporate various social context information. Our experiments using the events in two microblog datasets confirm that the method outperforms state-of-the-art rumor detection approaches by large margins. Moreover, our model demonstrates strong performance on detecting rumors at early stage after their initial broadcast.

Single-document summarization is a challenging task. In this paper, we explore effective ways using the tweets linking to news for generating extractive summary of each document. We reveal the very basic value of tweets that can be utilized by regarding every tweet as a vote for candidate sentences. Base on such finding, we resort to unsupervised summarization models by leveraging the linking tweets to master the ranking of candidate extracts via random walk on a heterogeneous graph. The advantage is that we can use the linking tweets to opportunistically "supervise" the summarization with no need of reference summaries. Furthermore, we analyze the influence of the volume and latency of tweets on the quality of output summaries since tweets come after news release. Compared to truly supervised summarizer unaware of tweets, our method achieves significantly better results with reasonably small tradeoff on latency; compared to the same using tweets as auxiliary features, our method is comparable while needing less tweets and much shorter time to achieve significant outperformance.

Web search ranking models are learned from features originated from different views or perspectives of document relevancy, such as query dependent or independent features. This seems intuitively conformant to the principle of multi-view approach that leverages distinct complementary views to improve model learning. In this paper, we aim to obtain optimal separation of ranking features into non-overlapping subsets (i.e., views), and use such different views for rank learning and adaptation. We present a novel semi-supervised multi-view ranking model, which is then extended into an adaptive ranker for search domains where no training data exists. The core idea is to proactively strengthen view consistency (i.e., the consistency between different rankings each predicted by a distinct view-based ranker) especially when training and test data follow divergent distributions. For this purpose, we propose a unified framework based on listwise ranking scheme to mutually reinforce the view consistency of target queries and the appropriate weighting of source queries that act as prior knowledge. Based on LETOR and Yahoo Learning to Rank datasets, our method significantly outperforms some strong baselines including single-view ranking models commonly used and multi-view ranking models that do not impose view consistency on target data.

Social media data are produced continuously by a large and uncontrolled number of users. The dynamic nature of such data requires the sentiment and topic analysis model to be also dynamically updated, capturing the most recent language use of sentiments and topics in text. We propose a dynamic Joint Sentiment-Topic model (dJST) which allows the detection and tracking of views of current and recurrent interests and shifts in topic and sentiment. Both topic and sentiment dynamics are captured by assuming that the current sentiment-topic-specific word distributions are generated according to the word distributions at previous epochs. We study three different ways of accounting for such dependency information: (1) sliding window where the current sentiment-topic word distributions are dependent on the previous sentiment-topic-specific word distributions in the last S epochs; (2) skip model where history sentiment topic word distributions are considered by skipping some epochs in between; and (3) multiscale model where previous long- and short- timescale distributions are taken into consideration. We derive efficient online inference procedures to sequentially update the model with newly arrived data and show the effectiveness of our proposed model on the Mozilla add-on reviews crawled between 2007 and 2011.

Social media streams such as Twitter are regarded as faster first-hand sources of information generated by massive users. The content diffused through this channel, although noisy, provides important complement and sometimes even a substitute to the traditional news media reporting. In this paper, we propose a novel unsupervised approach based on topic modeling to summarize trending subjects by jointly discovering the representative and complementary information from news and tweets. Our method captures the content that enriches the subject matter by reinforcing the identification of complementary sentence-tweet pairs. To valuate the complementarity of a pair, we leverage topic modeling formalism by combining a two-dimensional topic-aspect model and a cross-collection approach in the multi-document summarization literature. The final summaries are generated by co-ranking the news sentences and tweets in both sides simultaneously. Experiments give promising results as compared to state-of-the-art baselines.

Learning to adapt in a new setting is a common challenge to our knowledge and capability. New life would be easier if we actively pursued supervision from the right mentor chosen with our relevant but limited prior knowledge. This variant principle of active learning seems intuitively useful to many domain adaptation problems. In this paper, we substantiate its power for advancing automatic ranking adaptation, which is important in web search since it's prohibitive to gather enough labeled data for every search domain for fully training domain-specific rankers. For the cost-effectiveness, it is expected that only those most informative instances in target domain are collected to annotate while we can still utilize the abundant ranking knowledge in source domain. We propose a unified ranking framework to mutually reinforce the active selection of informative target-domain queries and the appropriate weighting of source training data as related prior knowledge. We select to annotate those target queries whose documents' order most disagrees among the members of a committee built on the mixture of source training data and the already selected target data. Then the replenished labeled set is used to adjust the importance of source queries for enhancing their rank transfer. This procedure iterates until labeling budget exhausts. Based on LETOR3.0 and Yahoo! Learning to Rank Challenge data sets, our approach significantly outperforms the random query annotation commonly used in ranking adaptation and the active rank learner on target-domain data only.

Like traditional supervised and semi-supervised algorithms, learning to rank for information retrieval requires document annotations provided by domain experts. It is costly to annotate training data for different search domains and tasks. We propose to exploit training data annotated for a related domain to learn to rank retrieved documents in the target domain, in which no labeled data is available. We present a simple yet effective approach based on instance-weighting scheme. Our method first estimates the importance of each related-domain document relative to the target domain. Then heuristics are studied to transform the importance of individual documents to the pairwise weights of document pairs, which can be directly incorporated into the popular ranking algorithms. Due to importance weighting, ranking model trained on related domain is highly adaptable to the data of target domain. Ranking adaptation experiments on LETOR3.0 dataset [27] demonstrate that with a fair amount of related-domain training data, our method significantly outperforms the baseline without weighting, and most of time is not significantly worse than an "ideal" model directly trained on target domain.

Query suggestion aims to suggest relevant queries for a given query, which helps users better specify their information needs. Previous work on query suggestion has been limited to the same language. In this article, we extend it to cross-lingual query suggestion (CLQS): for a query in one language, we suggest similar or relevant queries in other languages. This is very important to the scenarios of cross-language information retrieval (CLIR) and other related cross-lingual applications. Instead of relying on existing query translation technologies for CLQS, we present an effective means to map the input query of one language to queries of the other language in the query log. Important monolingual and cross-lingual information such as word translation relations and word co-occurrence statistics, and so on, are used to estimate the cross-lingual query similarity with a discriminative model. Benchmarks show that the resulting CLQS system significantly outperforms a baseline system that uses dictionary-based query translation. Besides, we evaluate CLQS with French-English and Chinese-English CLIR tasks on TREC-6 and NTCIR-4 collections, respectively. The CLIR experiments using typical retrieval models demonstrate that the CLQS-based approach has significantly higher effectiveness than several traditional query translation methods. We find that when combined with pseudo-relevance feedback, the effectiveness of CLIR using CLQS is enhanced for different pairs of languages.

Query suggestion aims to suggest relevant queries for a given query, which help users better specify their information needs. Previously, the suggested terms are mostly in the same language of the input query. In this paper, we extend it to cross-lingual query suggestion (CLQS): for a query in one language, we suggest similar or relevant queries in other languages. This is very important to scenarios of cross-language information retrieval (CLIR) and cross-lingual keyword bidding for search engine advertisement. Instead of relying on existing query translation technologies for CLQS, we present an effective means to map the input query of one language to queries of the other language in the query log. Important monolingual and cross-lingual information such as word translation relations and word co-occurrence statistics, etc. are used to estimate the cross-lingual query similarity with a discriminative model. Benchmarks show that the resulting CLQS system significantly out performs a baseline system based on dictionary-based query translation. Besides, the resulting CLQS is tested with French to English CLIR tasks on TREC collections. The results demonstrate higher effectiveness than the traditional query translation methods.