Publications

2026

  1. EMNLP
    ask_answer.png
    Ask or Answer: A Decision Framework for Multi-Turn Health Misinformation Intervention
    Xiaoying Song, Anirban Saha Anik, Jinyu Liu, Qitao Tan, Geng Yuan, and Lingzi Hong
    In Conference on Empirical Methods in Natural Language Processing, 2026
    Proposes RO-PnR (Reward-Optimized Probe-and-Respond), which decides at each turn whether to probe for more information or commit to a final correction, guided by a turn-level reward weighing a probe’s expected information gain against its interaction cost. Achieves the highest cost-adjusted utility across three health-misinformation datasets and three base models while using  30% fewer turns than always-probe baselines.
    multi-turn dialogue clarifying questions reward optimization
  2. Preprint
    cognitive_survey.png
    Towards Human-Centered Natural Language Processing: A Survey of Cognitive Signal Integration
    Xiaoying Song, Anirban Saha Anik, Jinyu Liu, Qitao Tan, Geng Yuan, Song Fu, Eduardo Blanco, and Lingzi Hong
    arXiv preprint, 2026
    Surveys cognition-aware NLP: the datasets of cognitive signals such as eye-tracking traces and EEG, the preprocessing pipelines, the tasks they are applied to, and how they are integrated with language models. Summarizes open challenges and future opportunities for more interpretable, user-centered human-AI interaction.
    cognitive signals eye tracking EEG survey human-centered NLP
  3. Preprint
    st_crisis.png
    ST-Crisis: A Diagnostic Benchmark for Spatiotemporal Grounding in Crisis Communication
    Xiaoying Song, Thi Bich Ngan Do, Nikhitha Rayabarapu, Anirban Saha Anik, Jinyu Liu, Qitao Tan, Vanessa Frias-Martinez, and Lingzi Hong
    arXiv preprint, 2026
    Introduces ST-Crisis, a benchmark pairing situational awareness posts with crisis need queries to evaluate spatiotemporal grounding and reasoning. A diagnostic evaluation of vanilla LLMs and spatiotemporally grounded RAG models exposes where current models fail to ground evidence in location and time during crisis response.
    spatiotemporal reasoning benchmark RAG crisis communication
  4. Preprint
    PALETTE.png
    Palette: A Modular, Controllable, and Efficient Framework for On-demand Authorized Safety Alignment Relaxation in LLMs
    Qitao Tan, Xiaoying Song, Arman Akbari, Arash Akbari, Yanzhi Wang, Xiaoming Zhai, Lingzi Hong, Zhen Xiang, Jin Lu, and Geng Yuan
    arXiv preprint arXiv:2605.24154, 2026 PDF
    Proposes Palette, a framework that selectively relaxes LLM refusal behavior in authorized professional domains while preserving safety elsewhere, combining refusal-direction search, lightweight adaptation, and parameter merging for multi-domain composition without retraining.
    safety alignment refusal behavior model merging controllability
  5. EMNLP
    Q.png
    Q-realign: Piggybacking Realignment on Quantization for Safe and Efficient LLM Deployment
    Qitao Tan, Xiaoying Song, Ningxi Cheng, Ninghao Liu, Xiaoming Zhai, Lingzi Hong, Yanzhi Wang, Zhen Xiang, and Geng Yuan
    In Conference on Empirical Methods in Natural Language Processing, 2026 PDF
    Introduces Q-realign, a post-training defense that restores safety alignment eroded by task-specific fine-tuning by piggybacking realignment on quantization. It reduces unsafe outputs while preserving task performance, realigning a 7B model on consumer hardware in about 40 minutes.
    safety realignment quantization fine-tuning risks efficient deployment
  6. EMNLP
    ZeroQAT.png
    End-to-End On-Device Quantization-Aware Training for LLMs at Inference Cost
    Qitao Tan, Xiaoying Song, Jin Lu, Guoming Li, Jun Liu, Lingzi Hong, Caiwen Ding, Jundong Li, Xiaoming Zhai, Shaoyi Huang, Wei Niu, and Geng Yuan
    In Conference on Empirical Methods in Natural Language Processing, 2026 PDF
    Proposes ZeroQAT, a quantization-aware training framework that uses zeroth-order, forward-only gradient estimation to avoid backpropagation memory costs, enabling 2-4 bit fine-tuning of a 13B model on an 8GB GPU and on-smartphone training.
    quantization zeroth-order optimization on-device training
  7. JIS
    LCC.png
    Comparative analysis of large language models’ performance in book classification tasks using Library of Congress Classification system
    Xiaoying Song, Pengcheng Luo, Jason Thomale, Oksana Zavalina, and Lingzi Hong
    Journal of Information Science, 2026 Paper
    Compares LLMs on classifying books into the Library of Congress Classification system, varying input data, data size, model, and class granularity. Llama3 performs best; diverse inputs help, but models struggle with long text and fine-grained categories.
    LLM evaluation book classification library cataloging
  8. JOIS
    JOIS.png
    Academic collaboration recommendation based on graph neural network and multi-attribute embedding
    Guanghui Ye, Jinyu Wei, Qitao Tan, Chuan Wu, Xiaoying Song, and Songye Li
    Journal of Information Science, 2026 Paper
    Proposes MAH, a heterogeneous network representation learning method for recommending academic collaborators that combines co-authorship and citation networks with multi-dimensional researcher attributes, improving recommendation performance on real-world datasets.
    collaboration recommendation graph neural networks network embedding

2025

  1. EMNLP
    Speak.png
    Speaking at the Right Level: Literacy-Controlled Counterspeech Generation with RAG-RL
    Xiaoying Song, Anirban Saha Anik, Dibakar Barua, Pengcheng Luo, Junhua Ding, and Lingzi Hong
    In Conference on Empirical Methods in Natural Language Processing, 2025 PDF
    Proposes Controlled-Literacy, a RAG plus reinforcement learning framework that generates counterspeech to health misinformation tailored to the audience’s health literacy level, producing more accessible and preferred counterspeech than baselines.
    readability alignment health misinformation RAG reinforcement learning
  2. EMNLP
    Crisis.png
    A Dynamic Fusion Model for Consistent Crisis Response
    Xiaoying Song, Anirban Saha Anik, Eduardo Blanco, Vanessa Frias-Martinez, and Lingzi Hong
    In Conference on Empirical Methods in Natural Language Processing, 2025 PDF
    Introduces a metric for style consistency in automated crisis responses and a two-stage fusion approach that scores candidate responses and integrates them at the instance level, improving both response quality and stylistic uniformity across datasets.
    crisis communication style consistency text generation model fusion
  3. COLM
    multi_agent.png
    Multi-Agent Retrieval-Augmented Framework for Evidence-Based Counterspeech Against Health Misinformation
    Xiaoying Song*, Anirban Saha Anik*, Elliott Wang, Bryan Wang, Bengisu Yarimbas, and Lingzi Hong
    In Conference on Language Modeling, 2025 PDF
    Presents a multi-agent retrieval-augmented framework that combines static and dynamic evidence to generate counterspeech against health misinformation, outperforming baselines in politeness, relevance, informativeness, and factual accuracy.
    dialogue system health misinformation multi-agent systems RAG
  4. NAACL
    hatespeech.png
    Echoes of Discord: Forecasting Hater Reactions to Counterspeech
    Xiaoying Song, Sharon Lisseth Perez, Xinchen Yu, Eduardo Blanco, and Lingzi Hong
    In Nations of the Americas Chapter of the Association for Computational Linguistics, 2025 PDF
    Studies whether counterspeech leads haters to reenter conversations and whether reentry stays hateful, introducing ReEco, a dataset of triple-turn Reddit conversations, and linking counterspeech language to hater reactions.
    response effectiveness hate speech user behavior prediction
  5. COLING
    humanlike.png
    Assessing the Human Likeness of AI-Generated Counterspeech
    Xiaoying Song, Sujana Mamidisetty, Eduardo Blanco, and Lingzi Hong
    In International Conference on Computational Linguistics, 2025 PDF
    Evaluates how human-like AI-generated counterspeech is, testing several LLM generation strategies. Both classifiers and human judges easily distinguish AI-generated from human-written counterspeech, with differences in linguistic style, politeness, and specificity.
    counterspeech LLM evaluation human likeness hate speech
  6. ISCRAM
    iscram.png
    Dynamic Fusion of Large Language Models for Crisis Communication
    Xiaoying Song, Lingzi Hong, Anirban Saha Anik, and Vanessa Frias-Martinez
    In Proceedings of the International ISCRAM Conference, 2025 PDF
    Explores instruction prompting, retrieval-augmented generation, and dynamic fusion of multiple LLMs to answer crisis-related information needs on social media. Dynamic fusion produces better responses in professionalism, actionability, empathy, and relevance.
    crisis communication LLM fusion RAG social media
  7. ASIST
    hybrid.png
    A Hybrid Framework for Subject Analysis: Integrating Embedding-Based Regression Models with Large Language Models
    Jinyu Liu, Xiaoying Song, Diana Zhang, Jason Thomale, Daqing He, and Lingzi Hong
    In Association for Information Science and Technology, 2025 Paper
    Proposes a hybrid framework for library subject analysis in which embedding-based models predict the number of Library of Congress Subject Headings and post-edit LLM outputs, yielding more controlled, vocabulary-aligned predictions with fewer hallucinations.
    subject analysis LCSH multi-label classification embeddings
  8. iConference
    flagging.png
    Analyzing the language of rejection: a study of user flagging responses to hate speech on Reddit
    Sharon Lisseth Perez, Xiaoying Song, and Lingzi Hong
    In Information Research an international electronic journal, 2025 PDF
    Builds a dataset of Reddit user flagging messages paired with hateful comments and applies transformer models to classify them. Linguistic analysis reveals distinct patterns in flagging content, suggesting user flags can improve hate speech moderation.
    hate speech content moderation Reddit user flagging

2024

  1. EMNLP
    outcome.png
    Outcome-Constrained Large Language Models for Countering Hate Speech
    Lingzi Hong, Pengcheng Luo, Eduardo Blanco, and Xiaoying Song
    In Conference on Empirical Methods in Natural Language Processing, 2024 PDF
    Generates counterspeech constrained by desired conversation outcomes, low incivility and non-hateful hater reentry, using prompting, fine-tuning, and reinforcement learning with LLMs. The methods steer generation toward those outcomes.
    counterspeech hate speech reinforcement learning controlled generation

2023

  1. MLJ
    ML.png
    An effective negative sampling approach for contrastive learning of sentence embedding
    Qitao Tan, Xiaoying Song, Guanghui Ye, and Chuan Wu
    Machine Learning, 2023 Paper
    Proposes nmCSE, a negative mining method for unsupervised contrastive sentence embedding learning that selects negatives via distance-based weighting and grid sampling instead of random sampling, improving performance on seven semantic textual similarity datasets.
    sentence embeddings contrastive learning negative sampling
  2. JOI
    JOI.png
    Research frontier detection and analysis based on research grants information: A case study on health informatics in the US
    Guanghui Ye, Cancan Wang, Chuan Wu, Ze Peng, Jinyu Wei, Xiaoying Song, Qitao Tan, and Lanqi Wu
    Journal of Informetrics, 2023 Paper
    Detects research frontiers from research grant information rather than papers or patents, using adapted indicators and path mining to trace topic evolution. A case study on US health informatics identifies five frontier topics among 47 clustered topics.
    research frontier detection grant analysis topic evolution