Papers by Zhouhan Lin

19 papers
Syntax-guided Localized Self-attention by Constituency Syntactic Distance (2022.findings-emnlp)

Copied to clipboard

Challenge: Recent studies have shown that Transformers is implicitly learning syntactic information from data, albeit is highly dependent on the quality and scale of the training data.
Approach: They propose a syntax-guided localized self-attention model that allows directly incorporating grammar structures from an external constituency parser.
Outcome: The proposed model improves translation performance on a variety of datasets, from small to large datasets and with different source languages.
Mirror-Consistency: Harnessing Inconsistency in Majority Voting (2024.findings-emnlp)

Copied to clipboard

Challenge: Large Language Models (LLMs) are a widely-used decoding strategy that relies on the plurality voting rule, which focuses on the most frequent answer while overlooking all other minority responses.
Approach: They propose to incorporate a ‘reflective mirror’ into the self-ensemble decoding process and enables LLMs to critically examine inconsistencies among multiple generations.
Outcome: The proposed method incorporates a ‘reflective mirror’ into the self-ensemble decoding process and enables LLMs to critically examine inconsistencies among multiple generations.
SH2: Self-Highlighted Hesitation Helps You Decode More Truthfully (2024.findings-emnlp)

Copied to clipboard

Challenge: Large language models (LLMs) have made great progress in text generation but suffer from hallucinations during reasoning and generation.
Approach: They propose an inference-time method to help LLMs decode truthfully by selecting tokens with the lowest probabilities and concatenating them to the original context.
Outcome: The proposed method improves LLaMA-7b, LLama2-7b and Mistral-7b on hallucination tasks.
DiVISe: Direct Visual-Input Speech Synthesis Preserving Speaker Characteristics And Intelligibility (2025.findings-naacl)

Copied to clipboard

Challenge: Video-to-speech (V2S) synthesis requires acoustic hints to accurately reconstruct both speech content and speaker characteristics from video clips alone.
Approach: They propose a video-to-speech (V2S) model that predicts Mel-spectrograms directly from video frames.
Outcome: The proposed model outperforms existing models in acoustic intelligibility and preserves speaker-specific characteristics.
Leveraging Grammar Induction for Language Understanding and Generation (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing grammar induction methods do not provide sufficient performance in downstream tasks.
Approach: They propose an unsupervised grammar induction method for language understanding and generation using a grammar parser and a syntactic mask.
Outcome: The proposed method performs better on from-scratch and pre-trained scenarios.
PaD: Program-aided Distillation Can Teach Small Models Reasoning Better than Chain-of-thought Fine-tuning (2024.naacl-long)

Copied to clipboard

Challenge: Large language models excel in various tasks, but their huge size and inaccessibility of parameters present challenges for practical deployment.
Approach: They propose to use CoT data to distill task-specific ability from large language models to smaller models . they use reasoning programs to suppress errors in distilled data and improve distillation quality .
Outcome: The proposed model outperforms LLMs on arithmetic reasoning, symbolic reasoning, and general ability.
Gumbel Reranking: Differentiable End-to-End Reranker Optimization (2025.acl-long)

Copied to clipboard

Challenge: Existing distillation-based approaches suffer from training-inference misalignment and fail to capture interdependencies among candidate documents.
Approach: They propose a method to optimize rerankers by learning a stochastic, document-wise Top-k attention mask using the Gumbel Trick and Relaxed Top-K Sampling.
Outcome: The proposed framework minimizes the overall language loss and improves recall on hotpotQA.
Fourier Transformer: Fast Long Range Modeling by Removing Sequence Redundancy with FFT Operator (2023.findings-acl)

Copied to clipboard

Challenge: Existing transformer models are computationally demanding and prohibitively costly for long sequences due to the quadratic complexity of its selfattention module.
Approach: They propose a transformer-based model that inherits weights from large pretrained models by removing redundancies in hidden sequences using the ready-made Fast Fourier Transform operator.
Outcome: The proposed model outperforms the standard BART model on the long-range modeling benchmark LRA with significant improvements in speed and space.
Interactive Language Learning by Question Answering (D19-1)

Copied to clipboard

Challenge: Existing machine reading comprehension tasks lack interactive information-seeking component of comprehension.
Approach: They propose a question-asking task that asks questions in a text-based environment . they propose QAit, which uses a game generator to build models that include deep reinforcement learning agents.
Outcome: The proposed task poses questions about existence, location, and attributes of objects found in environment.
Extracting Financial Events from Raw Texts via Matrix Chunking (2024.lrec-main)

Copied to clipboard

Challenge: Event Extraction (EE) is widely used in the Chinese financial field to provide valuable structured information.
Approach: They propose a task which extracts financial events from raw texts and an efficient method called MACK.
Outcome: The proposed method is fault-tolerant and can visualize interactions among text components.
Training LLMs to be Better Text Embedders through Bidirectional Reconstruction (2025.emnlp-main)

Copied to clipboard

Challenge: Existing text embedding approaches often leverage the embeddment of the final token, typically a reserved special token such as ‘[EOS]‘.
Approach: They propose to add a new training stage before contrastive learning to enrich the semantics of the final token embedding.
Outcome: The proposed training stage improves performance on the Massive Text Embedding Benchmark (MTEB), achieving new state-of-the-art results across different LLM base models and scales.
Unsupervised Graph-Text Mutual Conversion with a Unified Pretrained Language Model (2023.acl-long)

Copied to clipboard

Challenge: Existing unsupervised approaches for learning knowledge graphs require multiple modules and require entity information or relation type for training.
Approach: They propose a method that uses a unified pretrained language model to achieve fully unsupervised graph-text mutual conversion for the first time.
Outcome: The proposed method outperforms state-of-the-art methods for G2T and T2G tasks by fine-tuning only one pretrained model.
Straight to the Tree: Constituency Parsing with Neural Syntactic Distance (P18-1)

Copied to clipboard

Challenge: Compared to traditional shift-reduce parsing schemes, our approach is free from the potentially disastrous compounding error.
Approach: They propose a model that predicts a scalar for each split position in a sentence and then determines the topology of grammar tree based on syntactic distances.
Outcome: The proposed model achieves the state-of-the-art single model F1 score of 92.1 on PTB and 86.4 on CTB dataset, surpassing the previous single model results by a large margin.
Exploiting Syntactic Structure for Better Language Modeling: A Syntactic Distance Approach (2020.acl-main)

Copied to clipboard

Challenge: incorporating syntactic structure into language models has been a challenge since the 1990s.
Approach: They propose to use syntactic information to integrate syntastic structure into neural language models by providing ground truth parse trees as additional training signals.
Outcome: The proposed model achieves lower perplexity and better quality when ground truth parse trees are provided as training signals.
Enable Fast Sampling for Seq2Seq Text Diffusion (2024.findings-emnlp)

Copied to clipboard

Challenge: Existing text generation methods use autoregressive (AR) methods, which generate tokens one by one, but are time-consuming.
Approach: They propose an efficient model FMSeq which utilizes flow matching to straighten the generation path, thereby enabling fast sampling for diffusion-based seq2seq text generation.
Outcome: The proposed model generates comparable quality to the SOTA diffusion-based DiffuSeq in just 10 steps, achieving a 200-fold speedup.
Text Editing as Imitation Game (2022.findings-emnlp)

Copied to clipboard

Challenge: Text editing is an important domain of processing tasks to edit the text in a localized fashion, such as text simplification.
Approach: They propose a nonautoregressive decoder for state-to-action demonstrations that parallels the decoding while retaining the dependencies between tokens.
Outcome: The proposed model outperforms the autoregressive baselines on a suite of Arithmetic Equation benchmarks in terms of performance, efficiency, and robustness.
Transkimmer: Transformer Learns to Layer-wise Skim (2022.acl-long)

Copied to clipboard

Challenge: Prior work has proposed to augment Transformer model with the capability of skimming tokens to improve its computational efficiency.
Approach: They propose to add a parameterized predictor before each layer that learns to make the skimming decision.
Outcome: The proposed model achieves 10.97x speedup on GLUE benchmark compared with BERT-base baseline with less than 1% accuracy degradation.
RASAT: Integrating Relational Structures into Pretrained Seq2Seq Model for Text-to-SQL (2022.emnlp-main)

Copied to clipboard

Challenge: Experimental results show RASAT can leverage a variety of relational structures while inheriting the pretrained parameters from the T5 model.
Approach: They propose a Transformer seq2seq architecture augmented with relation-aware self-attention that leverages relational structures while inheriting pretrained parameters from the T5 model.
Outcome: The proposed model can leverage relational structures while inheriting pretrained parameters from the T5 model effectively.
Leveraging Unimodal Self-Supervised Learning for Multimodal Audio-Visual Speech Recognition (2022.acl-long)

Copied to clipboard

Challenge: Existing methods for audio-visual speech recognition use extra data to increase performance . a recent study shows that the use of unimodal self-supervised learning improves performance on multimodal tasks.
Approach: They propose to use unimodal self-supervised learning to train AVSR models on unlabelled unilateral data.
Outcome: The proposed model improves on lip reading sentences 2 by 30% even without an external language model.

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