Keyu He, Tejas Srinivasan, Brihi Joshi, Xiang Ren, Jesse Thomason, Swabha Swayamdipta. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

As the adoption of language models advances, so does the need to better represent individual users to the model.Are there aspects of an individual's belief system that a language model can utilize for improved alignment?Following prior research, we investigate this question in the domain of opinion prediction by developing PRIMEX, a dataset of public opinion survey data from 858 US residents with two additional sources of belief information: written explanations from the respondents for why they hold specific opinions, and the Primal World Belief survey for assessing respondent worldview.We provide an extensive initial analysis of our data and show the value of belief explanations and worldview for personalizing language models.Our results demonstrate how the additional belief information in PRIMEX can benefit both the NLP and psychological research communities, opening up avenues for further study.Public Opinion Surveys on multiple topics 858 Participants of various age, gender, education, race, & US region 1 2 3 Free-response explanations of 3 questions per survey Good vs. Bad Safe vs.

Today, large language models are widely used as judges to evaluate responses from other language models.Hence, it is imperative to benchmark and improve these LLM-judges on realworld language model usage: a typical humanassistant conversation is lengthy, and shows significant diversity in topics, intents, and requirements across turns, e.g.social interactions, task requests, feedback.We present AMULET, a framework that leverages pertinent linguistic concepts of dialog-acts and maxims to improve the accuracy of LLM-judges on preference data with complex, multi-turn conversational context.AMULET presents valuable insights about (a) the communicative structures and intents present in the conversation (dialog acts), and (b) the satisfaction of conversational principles (maxims) by the preference pair responses, and uses them to make judgments.On 4 challenging datasets, AMULET shows that (a) humans frequently (60-70% of the time) change their intents from one turn of the conversation to the next, and (b) in ∼75% of instances, the preference pair responses can be differentiated via dialog acts and/or maxims, reiterating the latter's significance in judging such data.AMULET can be used either as a judge by applying the framework to a single LLM, or integrated into a jury with different LLM judges; our judges and juries show strong improvements on relevant baselines for all 4 datasets.(code, data).

Jaspreet Ranjit, Brihi Joshi, Rebecca Dorn, Laura Petry, Olga Koumoundouros, Jayne Bottarini, Peichen Liu, Eric Rice, Swabha Swayamdipta. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Large language models (LMs) are capable of generating free-text rationales to aid question answering. However, prior work 1) suggests that useful self-rationalization is emergent only at significant scales (e.g., 175B parameter GPT-3); and 2) focuses largely on downstream performance, ignoring the semantics of the rationales themselves, e.g., are they faithful, true, and helpful for humans? In this work, we enable small-scale LMs (∼200x smaller than GPT-3) to generate rationales that not only improve downstream task performance, but are also more plausible, consistent, and diverse, assessed both by automatic and human evaluation. Our method, MaRio (Multi-rewArd RatIOnalization), is a multi-reward conditioned self-rationalization algorithm that optimizes multiple distinct properties like plausibility, diversity and consistency. Results on three difficult question-answering datasets StrategyQA, QuaRel and OpenBookQA show that not only does MaRio improve task accuracy, but it also improves the self-rationalization quality of small LMs across the aforementioned axes better than a supervised fine-tuning (SFT) baseline. Extensive human evaluations confirm that MaRio rationales are preferred vs. SFT rationales, as well as qualitative improvements in plausibility and consistency.

Brihi Joshi, Ziyi Liu, Sahana Ramnath, Aaron Chan, Zhewei Tong, Shaoliang Nie, Qifan Wang, Yejin Choi, Xiang Ren. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Contextual embeddings derived from transformer-based neural language models have shown state-of-the-art performance for various tasks such as question answering, sentiment analysis, and textual similarity in recent years.Extensive work shows how accurately such models can represent abstract, semantic information present in text.In this expository work, we explore a tangent direction and analyze such models' performance on tasks that require a more granular level of representation.We focus on the problem of textual similarity from two perspectives: matching documents on a granular level (requiring embeddings to capture fine-grained attributes in the text), and an abstract level (requiring embeddings to capture overall textual semantics).We empirically demonstrate, across two datasets from different domains, that despite high performance in abstract document matching as expected, contextual embeddings are consistently (and at times, vastly) outperformed by simple baselines like TF-IDF for more granular tasks.We then propose a simple but effective method to incorporate TF-IDF into models that use contextual embeddings, achieving relative improvements of up to 36% on granular tasks.

With the rise in popularity of social media platforms like Twitter, having higher influence on these platforms has a greater value attached to it, since it has the power to influence many decisions in the form of brand promotions and shaping opinions. However, blackmarket services that allow users to inorganically gain influence are a threat to the credibility of these social networking platforms. Twitter users can gain inorganic appraisals in the form of likes, retweets, and follows through these blackmarket services either by paying for them or by joining syndicates wherein they gain such appraisals by providing similar appraisals to other users. These customers tend to exhibit a mix of organic and inorganic retweeting behavior, making it tougher to detect them. In this article, we investigate these blackmarket customers engaged in collusive retweeting activities. We collect and annotate a novel dataset containing various types of information about blackmarket customers and use these sources of information to construct multiple user representations. We adopt Weighted Generalized Canonical Correlation Analysis (WGCCA) to combine these individual representations to derive user embeddings that allow us to effectively classify users as: genuine users, bots, promotional customers, and normal customers. Our method significantly outperforms state-of-the-art approaches (32.95% better macro F1-score than the best baseline).

Twitter's popularity has fostered the emergence of various illegal user activities - one such activity is to artificially bolster visibility of tweets by gaining large number of retweets within a short time span. The natural way to gain visibility is time-consuming. Therefore, users who want their tweets to get quick visibility try to explore shortcuts - one such shortcut is to approach the blackmarket services, and gain retweets for their own tweets by retweeting other customers' tweets. Thus the users intrinsically become a part of a collusive ecosystem controlled by these services. In this paper, we propose CoReRank, an unsupervised framework to detect collusive users (who are involved in producing artificial retweets), and suspicious tweets (which are submitted to the blackmarket services) simultaneously. CoReRank leverages the retweeting (or quoting) patterns of users, and measures two scores - the 'credibility' of a user and the 'merit' of a tweet. We propose a set of axioms to derive the interdependency between these two scores, and update them in a recursive manner. The formulation is further extended to handle the cold start problem. CoReRank is guaranteed to converge in a finite number of iterations and has linear time complexity. We also propose a semi-supervised version of CoReRank (called CoReRank+) which leverages a partial ground-truth labeling of users and tweets. Extensive experiments are conducted to show the superiority of CoReRank compared to six baselines on a novel dataset we collected and annotated. CoReRank beats the best unsupervised baseline method by 269% (20%) (relative) average precision and 300% (22.22%) (relative) average recall in detecting collusive (genuine) users. CoReRank+ beats the best supervised baseline method by 33.18% AUC. CoReRank also detects suspicious tweets with 0.85 (0.60) average precision (recall). To our knowledge, CoReRank is the first unsupervised method to detect collusive users and suspicious tweets simultaneously with theoretical guarantees.

Twitter has increasingly become a popular platform to share news and user opinion. A tweet is considered to be important if it receives high number of affirmative reactions from other Twitter users via Retweets. Retweet count is thus considered as a surrogate measure for positive crowd-sourced reactions - high number of retweets of a tweet not only help the tweet being broadcasted, but also aid in making its topic trending. This in turn bolsters the social reputation of the author of the tweet. Since social reputation/impact of users/t weets influences many decisions (such as promoting brands, advertisement etc.), several blackmarket syndicates have actively been engaged in producing fake retweets in a collusive manner. Users who want to boost the impact of their tweets approach the blackmarket services, and gain retweets for their own tweets by retweeting other customers' tweets. Thus they become customers of blackmarket syndicates and engage in fake activities. Interestingly, these customers are neither bots, nor even fake users - they are usually normal human beings; they express a mix of organic and inorganic retweeting activities, and there is no synchronicity across their behaviors. In this paper, we make a first attempt to investigate such blackmarket customers engaged in producing fake retweets. We collected and annotated a novel dataset comprising of customers of many blackmarket services and characterize them using a set of 64 novel features. We show how their social behavior differs from genuine users. We then use state-of-the-art supervised models to detect three types of customers (bots, promotional, normal) and genuine users. We achieve a Macro Fl-score of 0.87 with SVM, outperforming four other baselines significantly. We further design a browser extension, SCoRe which, given the link of a tweet, spots its fake retweeters in real-time. We also collected users' feedback on the performance of SCoRe and obtained 85% accuracy.