Papers by Haifeng Zhu

16 papers
CORD: Bridging the Audio–Text Reasoning Gap via Weighted On-policy Cross-modal Distillation (2026.findings-acl)

Copied to clipboard

Challenge: Large Audio Language Models (LALMs) exhibit a degradation in knowledge and reasoning capabilities . empirical results show that CORD significantly bridges the audio–text performance gap .
Approach: They propose a framework that performs online cross-modal self-distillation to bridge the acoustic-semantic gap between LALMs and text-based models.
Outcome: The proposed framework bridges the acoustic-semantic gap between LALMs and text-based models . it employs on-policy reverse KL divergence with importance-aware weighting .
Multi-turn Response Selection using Dialogue Dependency Relations (2020.emnlp-main)

Copied to clipboard

Challenge: Existing models for multi-turn response selection ignore the dependencies between the turns.
Approach: They propose a dialogue extraction algorithm to transform a dialog history into threads based on their dependency relations.
Outcome: The proposed model outperforms the state-of-the-art models on DSTC7 and DSTF8* with competitive results on UbuntuV2 .
RepoDebug: Repository-Level Multi-Task and Multi-Language Debugging Evaluation of Large Language Models (2025.findings-emnlp)

Copied to clipboard

Challenge: Large Language Models (LLMs) have exhibited significant proficiency in code debugging, especially in automatic program repair.
Approach: They propose a repository-level code debugging dataset with 22 subtypes of errors that supports 8 commonly used programming languages and 3 debug tasks.
Outcome: The proposed dataset supports 8 commonly used programming languages and 3 debugging tasks.
Learning to Contrast the Counterfactual Samples for Robust Visual Question Answering (2020.emnlp-main)

Copied to clipboard

Challenge: Existing methods of generating counterfactual samples are not fully utilized in the task of Visual Question Answering (VQA).
Approach: They propose a self-supervised contrastive learning mechanism to learn the relationship between original samples, factual samples and counterfactual samples.
Outcome: The proposed method surpasses state-of-the-art models on the VQA-CP dataset, a diagnostic benchmark for assessing the VQ model’s robustness.
Post-Training Dialogue Summarization using Pseudo-Paraphrasing (2022.findings-naacl)

Copied to clipboard

Challenge: Existing approaches to dialogue summarization use dialogue-specific features that require additional knowledge to recognize or make the models harder to tune.
Approach: They propose to post-train pretrained language models to rephrase from dialogue to narratives and fine-tune them as usual.
Outcome: The proposed approach outperforms existing models by summary quality and implementation costs.
Incomplete Utterance Rewriting by A Two-Phase Locate-and-Fill Regime (2023.findings-acl)

Copied to clipboard

Challenge: Existing models with incomplete utterances have too large search space, resulting in poor quality of rewriting results.
Approach: They propose a 2-phase rewriting framework which predicts empty slots in the utterance that need to be completed and generates the part to be filled into each position.
Outcome: The proposed framework achieves state-of-the-art results on several public rewriting datasets.
A Diverse and Effective Retrieval-Based Debt Collection System with Expert Knowledge (2025.naacl-industry)

Copied to clipboard

Challenge: Existing debt collection systems lack script diversity, contextual relevance and coherence due to their complexity.
Approach: They propose a script library based on real-world debt collection conversations and a retrieval based response system for contextual relevance.
Outcome: The proposed system improves script diversity and responds to debtor-collector conversations better through knowledge distillation.
MoE Adapter for Large Audio Language Models: Sparsity, Disentanglement, and Gradient-Conflict-Free (2026.findings-acl)

Copied to clipboard

Challenge: Existing research on Large Language Models (LLMs) limited to textual input modality . acoustic information is intrinsically heterogeneous, entangling attributes such as speech, music, and environmental context.
Approach: They propose a sparse Mixture-of-Experts architecture to decouple acoustic information by routing audio tokens to specialized experts.
Outcome: The proposed architecture outperforms existing models on audio semantic and paralinguistic tasks while retaining shared experts for global context.
Towards Graph-hop Retrieval and Reasoning in Complex Question Answering over Textual Database (2024.lrec-main)

Copied to clipboard

Challenge: Existing benchmarks for textual question answering only focus on single-chain or single-hop retrieval . Existing approaches to answer complex questions have limitations .
Approach: They propose to conduct Graph-Hop, a novel multi-chains and multi-hops retrieval paradigm in complex question answering.
Outcome: The proposed model provides explicit and fine-grained evidence graphs for complex question to support comprehensive and detailed reasoning.
Link Prediction on N-ary Relational Facts: A Graph-based Approach (2021.findings-acl)

Copied to clipboard

Challenge: Existing work on knowledge graphs (KGs) focused on binary relations, but higher-arity relations are ubiquitous in real-world KGs.
Approach: They propose a graph-based approach to link prediction on knowledge graphs using n-ary relational facts and edge-biased fully-connected attention.
Outcome: The proposed approach performs substantially better than current state-of-the-art across a variety of n-ary relational benchmarks.
In-sample Curriculum Learning by Sequence Completion for Natural Language Generation (2023.acl-long)

Copied to clipboard

Challenge: Existing work on curriculum learning rely on task-specific expertise and cannot generalize to different tasks.
Approach: They propose to do in-sample curriculum learning for natural language generation tasks using human-crafted rules and a numeric score for each sample based on domain expertise to rank the model.
Outcome: The proposed learning strategy generalizes well to different tasks and achieves significant improvements over baselines.
DuQM: A Chinese Dataset of Linguistically Perturbed Natural Questions for Evaluating the Robustness of Question Matching Models (2022.emnlp-main)

Copied to clipboard

Challenge: a comprehensive evaluation of QM models should be conducted on natural texts, not on artificial adversarial examples . ral models are often not robust to adversarials, which means they predict unexpected outputs .
Approach: They use a Chinese dataset to evaluate the robustness of QM models . they show that the effect of artificial adversarial examples does not work on natural texts .
Outcome: The proposed model is more robust than other models on natural questions with 32 linguistic perturbations.
ChatMatch: Evaluating Chatbots by Autonomous Chat Tournaments (2022.acl-long)

Copied to clipboard

Challenge: Existing automated evaluation systems of chatbots rely on static chat scripts as ground truth, which is hard to obtain.
Approach: They propose an interactive chatbot evaluation framework that allows chatbots to compete with each other like in a sports tournament.
Outcome: The proposed framework can rank chatbots independently from their model architectures and domains . existing evaluation systems rely on static chat scripts as ground truth .
3DRP-Net: 3D Relative Position-aware Network for 3D Visual Grounding (2023.emnlp-main)

Copied to clipboard

Challenge: 3D visual grounding aims to localize the desired objects in a 3D point cloud by a free-form language description.
Approach: They propose a relation-aware framework which captures relative spatial relationships between objects and enhances object attributes.
Outcome: The proposed framework outperforms state-of-the-art methods on three benchmarks . it captures relative spatial relationships between objects and enhances object attributes .
Reducing Sensitivity on Speaker Names for Text Generation from Dialogues (2023.findings-acl)

Copied to clipboard

Challenge: Pre-trained language models are sensitive to nuances, resulting in unfairness in real-world applications.
Approach: They propose to quantitatively measure a model's sensitivity on speaker names and comprehensively evaluate a number of known methods for reducing speaker name sensitivity.
Outcome: The proposed approach reduces speaker name sensitivity and improves quality of generation.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations