Rebecca Hwa

Rr. Nefriana, Muheng Yan, Rebecca Hwa et al.

Extremist communities increasingly rely on social media to sustain and amplify divisive discourse. However, the relationship between their internal participation structures, audience engagement, and narrative expression remains underexplored. This study analyzes ten years of Facebook activity by hate groups related to the Israel-Palestine conflict, focusing on anti-Semitic and Islamophobic ideologies. Consistent with prior work, we find that higher participation centralization in online hate groups is associated with greater user engagement across hate ideologies, suggesting the role of key actors in sustaining group activity over time. Meanwhile, our narrative frame detection models--based on an eight-frame extremist taxonomy (e.g., dehumanization, violence justification)--reveal a clear contrast across hate ideologies: centralized Islamophobic groups employ more uniform messaging, while centralized anti-Semitic groups demonstrate greater framing diversity and topical breadth, potentially reflecting distinct historical trajectories and leader coordination patterns. Analysis of the inter-group network indicates that, although centralization and homophily are not clearly linked, ideological distinctions emerge: Islamophobic groups cluster tightly, whereas anti-Semitic groups remain more evenly connected. Overall, these findings clarify how participation structure may shape the dissemination pattern and resonance of extremist narratives online and provide a foundation for tailored strategies to disrupt or mitigate such discourse.

Omid Kashefi, Tazin Afrin, Meghan Dale et al.

Analyzing how humans revise their writings is an interesting research question, not only from an educational perspective but also in terms of artificial intelligence. Better understanding of this process could facilitate many NLP applications, from intelligent tutoring systems to supportive and collaborative writing environments. Developing these applications, however, requires revision corpora, which are not widely available. In this work, we present ArgRewrite V.2, a corpus of annotated argumentative revisions, collected from two cycles of revisions to argumentative essays about self-driving cars. Annotations are provided at different levels of purpose granularity (coarse and fine) and scope (sentential and subsentential). In addition, the corpus includes the revision goal given to each writer, essay scores, annotation verification, pre- and post-study surveys collected from participants as meta-data. The variety of revision unit scope and purpose granularity levels in ArgRewrite, along with the inclusion of new types of meta-data, can make it a useful resource for research and applications that involve revision analysis. We demonstrate some potential applications of ArgRewrite V.2 in the development of automatic revision purpose predictors, as a training source and benchmark.

Yongsu Ahn, Muheng Yan, Yu-Ru Lin et al.

With the rise of AI and data mining techniques, group profiling and group-level analysis have been increasingly used in many domains including policy making and direct marketing. In some cases, the statistics extracted from data may provide insights to a group's shared characteristics; in others, the group-level analysis can lead to problems including stereotyping and systematic oppression. How can analytic tools facilitate a more conscientious process in group analysis? In this work, we identify a set of accountable group analytics design guidelines to explicate the needs for group differentiation and preventing overgeneralization of a group. Following the design guidelines, we develop TribalGram, a visual analytic suite that leverages interpretable machine learning algorithms and visualization to offer inference assessment, model explanation, data corroboration, and sense-making. Through the interviews with domain experts, we showcase how our design and tools can bring a richer understanding of "groups" mined from the data.

Dae-young Kim, Rebecca Hwa, Muhammad Mahbubur Rahman

This paper introduces mhGPT, a lightweight generative pre-trained transformer trained on mental health-related social media and PubMed articles. Fine-tuned for specific mental health tasks, mhGPT was evaluated under limited hardware constraints and compared with state-of-the-art models like MentaLLaMA and Gemma. Despite having only 1.98 billion parameters and using just 5% of the dataset, mhGPT outperformed larger models and matched the performance of models trained on significantly more data. The key contributions include integrating diverse mental health data, creating a custom tokenizer, and optimizing a smaller architecture for low-resource settings. This research could advance AI-driven mental health care, especially in areas with limited computing power.

Iyad Ait Hou, Shrenik Borad, Harsh Sharma et al.

Machine unlearning is increasingly important for clinical language models, where privacy regulations and institutional policies may require removing sensitive information from deployed systems without retraining from scratch. In practice, deletion requests must balance effective forgetting of targeted information with preservation of model utility and minimal parameter modification. We introduce Sparse Token Embedding Unlearning (STEU), a parameter-efficient method for behavioral class-level unlearning that updates only PMI-selected token embeddings together with a small classifier head while keeping all encoder layers frozen. Across experiments on MIMIC-IV, MIMIC-III, and eICU using BioClinicalBERT, BERT-base, and DistilBERT, STEU consistently suppresses the target class while largely preserving retained task performance. In the primary MIMIC-IV setting, STEU achieves near-complete forgetting (forget F1 = 0.0004) while maintaining competitive retained utility (retain avg F1 = 0.4766) after modifying only 0.19\% of model parameters. These results suggest that targeted behavioral unlearning can be achieved through sparse embedding edits without modifying deeper encoder representations.

Rr. Nefriana, Muheng Yan, Ahmad Diab et al.

This short paper explores trends in extremist Facebook data from July 2023 to June 2024. We examined engagement, sentiment, and topics within Facebook groups categorized as anti-Israel/Semitic, anti-Palestine/Muslim, and anti-both, mapping these trends against five major events related to the recent Israel-Hamas conflict. Our findings support the hypothesis that shifts in trends correspond with these key events, showing varying patterns across different group categories. We observed decreased activity proportion in anti-both groups and increased activity proportion in the two one-sided hate groups at the conflict's onset. This pattern reversed after the Israeli troop withdrawal from Khan Yunis, Gaza. During the conflict, negative content proportion surged, and neutral content proportion fell in all the three group categories. Anti-Palestine/Muslim groups' discourses shifted from religious to social media activism and political/protest around the time the war began, while anti-Israel/Semitic groups moved from political/protest to religious topics a couple of weeks before the war.

Iyad Ait Hou, Rebecca Hwa

If the same neuron activates for both "lender" and "riverside," standard metrics attribute the overlap to superposition--the neuron must be compressing two unrelated concepts. This work explores how much of the overlap is due a lexical confound: neurons fire for a shared word form (such as "bank") rather than for two compressed concepts. A 2x2 factorial decomposition reveals that the lexical-only condition (same word, different meaning) consistently exceeds the semantic-only condition (different word, same meaning) across models spanning 110M-70B parameters. The confound carries into sparse autoencoders (18-36% of features blend senses), sits in <=1% of activation dimensions, and hurts downstream tasks: filtering it out improves word sense disambiguation and makes knowledge edits more selective (p = 0.002).

Tazin Afrin, Omid Kashefi, Christopher Olshefski et al.

We present the design and evaluation of a web-based intelligent writing assistant that helps students recognize their revisions of argumentative essays. To understand how our revision assistant can best support students, we have implemented four versions of our system with differences in the unit span (sentence versus sub-sentence) of revision analysis and the level of feedback provided (none, binary, or detailed revision purpose categorization). We first discuss the design decisions behind relevant components of the system, then analyze the efficacy of the different versions through a Wizard of Oz study with university students. Our results show that while a simple interface with no revision feedback is easier to use, an interface that provides a detailed categorization of sentence-level revisions is the most helpful based on user survey data, as well as the most effective based on improvement in writing outcomes.

Mingda Zhang, Chun-Te Chu, Andrey Zhmoginov et al.

In this work, we present BasisNet which combines recent advancements in efficient neural network architectures, conditional computation, and early termination in a simple new form. Our approach incorporates a lightweight model to preview the input and generate input-dependent combination coefficients, which later controls the synthesis of a more accurate specialist model to make final prediction. The two-stage model synthesis strategy can be applied to any network architectures and both stages are jointly trained. We also show that proper training recipes are critical for increasing generalizability for such high capacity neural networks. On ImageNet classification benchmark, our BasisNet with MobileNets as backbone demonstrated clear advantage on accuracy-efficiency trade-off over several strong baselines. Specifically, BasisNet-MobileNetV3 obtained 80.3% top-1 accuracy with only 290M Multiply-Add operations, halving the computational cost of previous state-of-the-art without sacrificing accuracy. With early termination, the average cost can be further reduced to 198M MAdds while maintaining accuracy of 80.0% on ImageNet.

Mingda Zhang, Tristan Maidment, Ahmad Diab et al.

The observation that computer vision methods overfit to dataset specifics has inspired diverse attempts to make object recognition models robust to domain shifts. However, similar work on domain-robust visual question answering methods is very limited. Domain adaptation for VQA differs from adaptation for object recognition due to additional complexity: VQA models handle multimodal inputs, methods contain multiple steps with diverse modules resulting in complex optimization, and answer spaces in different datasets are vastly different. To tackle these challenges, we first quantify domain shifts between popular VQA datasets, in both visual and textual space. To disentangle shifts between datasets arising from different modalities, we also construct synthetic shifts in the image and question domains separately. Second, we test the robustness of different families of VQA methods (classic two-stream, transformer, and neuro-symbolic methods) to these shifts. Third, we test the applicability of existing domain adaptation methods and devise a new one to bridge VQA domain gaps, adjusted to specific VQA models. To emulate the setting of real-world generalization, we focus on unsupervised domain adaptation and the open-ended classification task formulation.

Meiqi Guo, Rebecca Hwa, Yu-Ru Lin et al.

We investigate the impact of political ideology biases in training data. Through a set of comparison studies, we examine the propagation of biases in several widely-used NLP models and its effect on the overall retrieval accuracy. Our work highlights the susceptibility of large, complex models to propagating the biases from human-selected input, which may lead to a deterioration of retrieval accuracy, and the importance of controlling for these biases. Finally, as a way to mitigate the bias, we propose to learn a text representation that is invariant to political ideology while still judging topic relevance.

Mingda Zhang, Rebecca Hwa, Adriana Kovashka

Images and text in advertisements interact in complex, non-literal ways. The two channels are usually complementary, with each channel telling a different part of the story. Current approaches, such as image captioning methods, only examine literal, redundant relationships, where image and text show exactly the same content. To understand more complex relationships, we first collect a dataset of advertisement interpretations for whether the image and slogan in the same visual advertisement form a parallel (conveying the same message without literally saying the same thing) or non-parallel relationship, with the help of workers recruited on Amazon Mechanical Turk. We develop a variety of features that capture the creativity of images and the specificity or ambiguity of text, as well as methods that analyze the semantics within and across channels. We show that our method outperforms standard image-text alignment approaches on predicting the parallel/non-parallel relationship between image and text.

Gaurav Trivedi, Phuong Pham, Wendy Chapman et al.

Natural Language Processing (NLP) systems often make use of machine learning techniques that are unfamiliar to end-users who are interested in analyzing clinical records. Although NLP has been widely used in extracting information from clinical text, current systems generally do not support model revision based on feedback from domain experts. We present a prototype tool that allows end users to visualize and review the outputs of an NLP system that extracts binary variables from clinical text. Our tool combines multiple visualizations to help the users understand these results and make any necessary corrections, thus forming a feedback loop and helping improve the accuracy of the NLP models. We have tested our prototype in a formative think-aloud user study with clinicians and researchers involved in colonoscopy research. Results from semi-structured interviews and a System Usability Scale (SUS) analysis show that the users are able to quickly start refining NLP models, despite having very little or no experience with machine learning. Observations from these sessions suggest revisions to the interface to better support review workflow and interpretation of results.