Papers by Fangzhao Wu
Effective and Efficient Query-aware Snippet Extraction for Web Search (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to extract webpage snippets ignore contextual information of webpages, which may be sub-optimal. |
| Approach: | They propose a query-aware webpage snippet extraction method called DeepQSE that captures contextual information of webpages. |
| Outcome: | The proposed method can significantly improve the performance of DeepQSE without affecting its performance. |
Privacy-Preserving News Recommendation Model Learning (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing news recommendation methods rely on centralized storage of user behavior data for model training, which may lead to privacy concerns and risks due to the privacy-sensitive nature of user behaviors. |
| Approach: | They propose a privacy-preserving method where user behavior data is locally stored on user devices to train accurate news recommendation models. |
| Outcome: | The proposed method can train accurate news recommendation models without centralized storage of user behavior data. |
Defending against Indirect Prompt Injection by Instruction Detection (2025.findings-emnlp)
Copied to clipboard
Tongyu Wen, Chenglong Wang, Xiyuan Yang, Haoyu Tang, Yueqi Xie, Lingjuan Lyu, Zhicheng Dou, Fangzhao Wu
| Challenge: | Indirect Prompt Injection attacks can be exploited by LLMs that are embedded with external data. |
| Approach: | They propose a detection-based approach that leverages the behavioral states of LLMs to identify potential IPI attacks. |
| Outcome: | The proposed approach reduces the success rate of attacks to 0.03% on the BIPIA benchmark. |
On the Vulnerability of Safety Alignment in Open-Access LLMs (2024.findings-acl)
Copied to clipboard
Jingwei Yi, Rui Ye, Qisi Chen, Bin Zhu, Siheng Chen, Defu Lian, Guangzhong Sun, Xing Xie, Fangzhao Wu
| Challenge: | Large language models (LLMs) are susceptible to malicious exploitation, but are often rejected and limited harmfulness is limited. |
| Approach: | They propose two types of reverse alignment techniques: reverse supervised fine-tuning (RSFT) and reverse preference optimization (RPO). |
| Outcome: | The proposed methods can significantly enhance the success rate and harmfulness of jailbreak attacks, but they face high rejection rates and limited harmfulness. |
SGT: Securing Open-Source LLMs Against Malicious Fine-tuning via Safety Guidance Trigger (2026.acl-long)
Copied to clipboard
| Challenge: | open-weight large language models increase vulnerability to malicious fine-tuning . despite these advantages, open-source LLMs increase the potential for misuse . |
| Approach: | They propose a safety guide for open-weight large language models that guides fine-tuning toward the safety manifold to preserve alignment. |
| Outcome: | The proposed safety guidance trigger significantly improves robustness against malicious fine-tuning. |
Uni-FedRec: A Unified Privacy-Preserving News Recommendation Framework for Model Training and Online Serving (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing news recommendation methods rely on user behavior data to model user interests and user interests. |
| Approach: | They propose a unified news recommendation framework that uses user data locally stored in user clients to train models and serve users in a privacy-preserving way. |
| Outcome: | The proposed framework outperforms baseline methods and effectively protects user privacy. |
Tiny-NewsRec: Effective and Efficient PLM-based News Recommendation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing work fine tunes the PLM with the news recommendation task, which can cause a domain shift problem. |
| Approach: | They propose a self-supervised method to adapt general PLM to news domain with a contrastive matching task between news titles and news bodies. |
| Outcome: | The proposed method can improve both the effectiveness and efficiency of the large PLM-based news recommendation model while maintaining its performance. |
Reviews Meet Graphs: Enhancing User and Item Representations for Recommendation with Hierarchical Attentive Graph Neural Network (D19-1)
Copied to clipboard
| Challenge: | Existing methods to learn user and item representations from review texts do not take into account the user-user and item-item relatedness of the user. |
| Approach: | They propose to use review content and user-item graphs to integrate them as different views. |
| Outcome: | The proposed approach can learn user and item representations from review content and user-item graphs. |
Hierarchical User and Item Representation with Three-Tier Attention for Recommendation (N19-1)
Copied to clipboard
| Challenge: | Existing methods to learn user and item representations from reviews are limited . existing methods learn user representations based on ratings given by users . |
| Approach: | They propose a hierarchical user and item representation model with three-tier attention to learn user and items from reviews for recommendation. |
| Outcome: | The proposed model can learn user and item representations from reviews on four benchmark datasets. |
Neural News Recommendation with Heterogeneous User Behavior (D19-1)
Copied to clipboard
| Challenge: | Existing news recommendation methods rely on news click history to model user interest, but data sparsity is a problem . other kinds of user behaviors such as webpage browsing and search queries can provide useful clues of users’ news reading interest. |
| Approach: | They propose to exploit heterogeneous user behaviors to learn news representations from their titles via CNN networks and apply attention networks to select important words. |
| Outcome: | The proposed approach exploits heterogeneous user behaviors on a real-world dataset. |
NoisyTune: A Little Noise Can Help You Finetune Pretrained Language Models Better (2022.acl-short)
Copied to clipboard
| Challenge: | Existing methods for finetuning pretrained language models (PLMs) have risks in overfitting the pretraining tasks and data, which may lead to suboptimal performance. |
| Approach: | They propose a method which adds noise to parameters of PLMs before fine-tuning. |
| Outcome: | The proposed method can be used on GLUE English and XTREME multilingual benchmarks. |
Neural News Recommendation with Long- and Short-term User Representations (P19-1)
Copied to clipboard
| Challenge: | Existing news recommendation methods learn a single representation for each user, which may be insufficient. |
| Approach: | They propose a neural news recommendation approach which can learn both long- and short-term user representations by using a news encoder and a user encoder. |
| Outcome: | The proposed approach can learn both long- and short-term user representations on a real-world dataset. |
DA-Transformer: Distance-aware Transformer (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing models that capture token distances are not optimal for modeling the orders and relations of contexts. |
| Approach: | They propose a distance-aware Transformer that can exploit the real distances between tokens to re-scale the raw self-attention weights. |
| Outcome: | The proposed model outperforms the existing Transformer and its variants on five benchmark datasets and can improve the performance of many tasks. |
Are You Copying My Model? Protecting the Copyright of Large Language Models for EaaS via Backdoor Watermark (2023.acl-long)
Copied to clipboard
Wenjun Peng, Jingwei Yi, Fangzhao Wu, Shangxi Wu, Bin Bin Zhu, Lingjuan Lyu, Binxing Jiao, Tong Xu, Guangzhong Sun, Xing Xie
| Challenge: | Large language models (LLMs) have demonstrated exceptional abilities in both text understanding and generation. |
| Approach: | They propose an Embedding Watermark method that implants backdoors on embeddings to protect copyright of large language models. |
| Outcome: | The proposed method protects the copyright of large language models without compromising service quality while minimizing the adverse impact on the original embeddings’ utility. |
Exploring Sequence-to-Sequence Learning in Aspect Term Extraction (P19-1)
Copied to clipboard
| Challenge: | Aspect term extraction (ATE) aims at identifying all aspect terms in a sentence . sequence labeling based methods cannot make full use of overall meaning of sentence if they have dependencies between labels. |
| Approach: | They propose to formalize ATE as a sequence-to-sequence (Seq2Seque) learning task . they propose gated unit networks and position-aware attention mechanism to make it suit to ATE . |
| Outcome: | The proposed learning task is effective when labels correspond to words one by one . the proposed learning system is gated unit networks and position-aware attention mechanism . |
Attentive Pooling with Learnable Norms for Text Representation (2020.acl-main)
Copied to clipboard
| Challenge: | Existing pooling methods that use fixed pooling norms may not be optimal for learning text representations in different tasks. |
| Approach: | They propose to learn pooling norms in an end-to-end manner to automatically find the optimal ones for text representation in different tasks. |
| Outcome: | The proposed approach improves on four benchmark datasets on a neural NLP model. |
One Teacher is Enough? Pre-trained Language Model Distillation from Multiple Teachers (2021.findings-acl)
Copied to clipboard
| Challenge: | Pre-trained language models (PLMs) have huge model sizes and computational complexity, making it difficult to deploy them to low-latency and high-concurrence online systems. |
| Approach: | They propose a multi-teacher knowledge distillation framework for pre-trained language model compression that can train high-quality student model from multiple teacher PLMs. |
| Outcome: | The proposed framework can train high-quality student model from multiple teacher PLMs with shared pooling and prediction layers to align output space for better collaborative teaching. |
HieRec: Hierarchical User Interest Modeling for Personalized News Recommendation (2021.acl-long)
Copied to clipboard
| Challenge: | Existing news recommendation methods learn a single user embedding for each user from their previous behaviors to represent their overall interest. Existing methods only learn 'one' embeddable representation vectors to model user interest. |
| Approach: | They propose a news recommendation method with hierarchical user interest modeling that captures user interest in news rather than a single user embedding. |
| Outcome: | The proposed method can better capture multi-grained user interest in news. |
Fine-grained Interest Matching for Neural News Recommendation (2020.acl-main)
Copied to clipboard
| Challenge: | Existing studies represent each user as a single vector and then match the candidate news vector, which may lose fine-grained information for recommendation. |
| Approach: | They propose a Fine-grained interest matching method for neural news recommendation based on multi-level representations and fine-grain matching between segment pairs of each browsed news and the candidate news at each semantic level. |
| Outcome: | The proposed model can capture more fine-grained interest matching signals by performing interactions between each pair of news at multi-level semantic granularities. |
RLSeek: Evidence-Grounded Reasoning for RAG Hallucination Detection (2026.acl-long)
Copied to clipboard
Zhaoheng Huang, Dacheng Wen, Yutao Zhu, Xiaoying Lian, Yushi Liang, Kai Hao, Nan Li, Liangjie Zhang, Qi Zhang, Ji-Rong Wen, Zhicheng Dou, Fangzhao Wu
| Challenge: | Recent work addresses this problem by training span-level hallucination detectors using reinforcement learning and chain-of-thought reasoning. |
| Approach: | They propose a framework that explicitly enforces active evidence seeking during CoT reasoning by requiring quotation of relevant source segments at each verification step. |
| Outcome: | The proposed framework improves hallucination span detection performance with limited reasoning overhead and improved robustness in out-of-domain settings. |
Hi-Transformer: Hierarchical Interactive Transformer for Efficient and Effective Long Document Modeling (2021.acl-short)
Copied to clipboard
| Challenge: | Existing approaches to model long documents are difficult due to the quadratic complexity of text length. |
| Approach: | They propose a hierarchical interactive Transformer for efficient long document modeling. |
| Outcome: | Extensive experiments on three benchmark datasets validate the efficiency and effectiveness of Hi-Transformer in long document modeling. |
Efficient-FedRec: Efficient Federated Learning Framework for Privacy-Preserving News Recommendation (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing news recommendation methods rely on centralized storage of user click behavior data, which may lead to privacy concerns and hazards. |
| Approach: | They propose a federated learning framework for privacy-preserving news recommendation . they propose aggregation of news representations and user model by a client . |
| Outcome: | The proposed framework reduces computation and communication cost on clients while keeping promising model performance. |
Two Birds with One Stone: Unified Model Learning for Both Recall and Ranking in News Recommendation (2022.findings-acl)
Copied to clipboard
| Challenge: | Existing news recommender systems conduct news recall and ranking separately with different models, but maintaining multiple models leads to high computational cost and high latency. |
| Approach: | They propose a unified method for recall and ranking in news recommendation that uses historical news click behaviors to extract user embeddings for ranking from the user's attention query. |
| Outcome: | The proposed method improves recall and ranking efficiency and effectiveness on a benchmark dataset. |
PTUM: Pre-training User Model from Unlabeled User Behaviors via Self-supervision (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Existing methods for user modeling cannot exploit useful information in unlabeled data . Existing models only model task-specific user information and do not exploit universal user information encoded in user behaviors. |
| Approach: | They propose to pre-train user models from large-scale unlabeled user behavior data. |
| Outcome: | The proposed method can model relatedness between historical and future behaviors on two real-world datasets. |
PP-Rec: News Recommendation with Personalized User Interest and Time-aware News Popularity (2021.acl-long)
Copied to clipboard
| Challenge: | Existing personalized news recommendation methods have difficulties in making accurate recommendations to cold-start users. |
| Approach: | They propose to incorporate news popularity information to improve cold-start recommendations . they propose to use a popularity-aware user encoder to eliminate popularity bias . |
| Outcome: | The proposed method improves accuracy and diversity of personalized news recommendation on two real-world datasets. |
SentiRec: Sentiment Diversity-aware Neural News Recommendation (2020.aacl-main)
Copied to clipboard
| Challenge: | Existing news recommendation methods rank candidate news based on relevance to users’ historical browsed news, but if browsed data is dominated by certain kinds of sentiment, the model may recommend news with the same sentiment. |
| Approach: | They propose a sentiment diversity-aware neural news recommendation approach which can recommend news with more diverse sentiment without performance sacrifices. |
| Outcome: | The proposed approach can improve the sentiment diversity in news recommendation without performance sacrifice. |
Neural News Recommendation with Multi-Head Self-Attention (D19-1)
Copied to clipboard
| Challenge: | Precisely modeling news and users is critical for news recommendation, and capturing the contexts of words and news is important to learn news and user representations. |
| Approach: | They propose a neural news recommendation approach with multi-head self-attention to model the interactions between words and news and use multi-headed self- attention to capture relatedness between the news. |
| Outcome: | The proposed approach can learn representations from news titles by modeling the interactions between words and users and capture relatedness between the news. |
Measuring Human Contribution in AI-Assisted Content Generation (2026.acl-long)
Copied to clipboard
Yueqi Xie, Tao Qi, Jingwei Yi, Xiyuan Yang, Ryan Whalen, Junming Huang, Qian Ding, Yu Xie, Xing Xie, Fangzhao Wu
| Challenge: | generative AI has created a new way to generate content with humans . varying degrees of human contribution in content generation poses significant challenges for the delineation of originality . |
| Approach: | They propose a framework to measure human contribution in AI-assisted content generation by calculating mutual information between human input and AI-aided output relative to self-information of AI-assist output. |
| Outcome: | The proposed measure discriminates between varying degrees of human contribution across multiple creative domains and is validated in real-world applications. |
Neural News Recommendation with Topic-Aware News Representation (P19-1)
Copied to clipboard
| Challenge: | Existing methods for learning accurate news representations do not consider topic information in news. |
| Approach: | They propose a neural news recommendation approach with topic-aware news representations using CNN networks and attention networks to select important words. |
| Outcome: | The proposed approach is based on a topic-aware news encoder and user encoder. |
MIND: A Large-scale Dataset for News Recommendation (2020.acl-main)
Copied to clipboard
Fangzhao Wu, Ying Qiao, Jiun-Hung Chen, Chuhan Wu, Tao Qi, Jianxun Lian, Danyang Liu, Xing Xie, Jianfeng Gao, Winnie Wu, Ming Zhou
| Challenge: | Personalized news recommendation is an important technique for personalized news service. |
| Approach: | They propose to build a large-scale news recommendation dataset from Microsoft News . they demonstrate that news recommendation relies on the quality of news content understanding . |
| Outcome: | The proposed dataset contains 1 million users and more than 160k English news articles, each of which has rich textual content such as title, abstract and body. |
DebiasGAN: Eliminating Position Bias in News Recommendation with Adversarial Learning (2022.findings-emnlp)
Copied to clipboard
| Challenge: | Existing news recommendation methods use click behaviors for interest inference and model training, but position biases can be inaccurate in targeting user interest. |
| Approach: | They propose a news recommendation method that eliminates position biases by adversarial learning by a candidate-aware click model and a bias-invariant click model. |
| Outcome: | The proposed method can effectively alleviate position biases on click behaviors and capture unbiased user interest. |
NewsBERT: Distilling Pre-trained Language Model for Intelligent News Application (2021.findings-emnlp)
Copied to clipboard
| Challenge: | Existing language models are pre-trained and distilled on general corpus like Wikipedia, which has gaps with the news domain and may be suboptimal for news intelligence. |
| Approach: | They propose a method to distill existing language models on Wikipedia to enable efficient news intelligence. |
| Outcome: | The proposed model can be used to build and test a news intelligence application on Wikipedia and Wikipedia. |