Papers by Hen-Hsen Huang

Chinese Discourse Parsing: Model and Evaluation (2020.lrec-1)

Challenge:	This tutorial will provide an overview of financial opinion mining and provide research directions.
Approach:	This tutorial will introduce financial opinion mining and examine possible research directions.
Outcome:	This tutorial aims to provide an overview of financial opinion mining and figure out research directions.

Copied to clipboard

Lin Chuan-An, Shyh-Shiun Hung, Hen-Hsen Huang, Hsin-Hsi Chen

Unveiling Selection Biases: Exploring Order and Token Sensitivity in Large Language Models (2024.findings-acl)

Challenge:	Chinese discourse parsing has not yet a consistent evaluation metric . micro vs. macro F1 scores, binary v. multiway ground truth, and left-heavy v . right-heaviness binarization are important for Chinese discourses .
Approach:	They propose a neural network model that unifies a pre-trained transformer and a CKY-like algorithm and compare it with previous models with different evaluation scenarios.
Outcome:	The proposed model outperforms the previous models with different evaluation scenarios.

Copied to clipboard

Sheng-Lun Wei, Cheng-Kuang Wu, Hen-Hsen Huang, Hsin-Hsi Chen

MESAQA: A Dataset for Multi-Span Contextual and Evidence-Grounded Question Answering (2025.coling-main)

Challenge:	Using zero-shot or few-shot prompting, Large Language Models have been widely adopted in downstream applications.
Approach:	They propose to quantify the impact of option order and token usage on LLMs and propose mitigation strategies to enhance model performance.
Outcome:	The proposed mitigation strategies improve model performance and reduce the impact of token and order sensitivity on LLMs.

Copied to clipboard

Jui-I Wang, Hen-Hsen Huang, Hsin-Hsi Chen

Question Answering

Challenge:	Existing question answering systems focus on extracting answers from single spans, but real-world scenarios require synthesizing information from multiple spans.
Approach:	They propose a dataset that leverages the MASH-QA dataset and large language models (LLMs) to ensure that each Q/A pair requires considering all selected spans.
Outcome:	The proposed method enables the model to answer multiple Q/A pairs in a single span, while ensuring that all selected spans are considered.

Self-Adapted Utterance Selection for Suicidal Ideation Detection in Lifeline Conversations (2023.eacl-main)

Copied to clipboard

Zhong-Ling Wang, Po-Hsien Huang, Wen-Yau Hsu, Hen-Hsen Huang

Dialogue Systems & Conversational Agents

Challenge:	Existing methods for identifying suicidal ideation in phone conversations are difficult to use because of their long duration and noisy nature.
Approach:	They propose a self-adaptive approach that identifies the most critical utterances that the NLP model can more easily distinguish.
Outcome:	The proposed approach outperforms the baseline models in overall performance with an F score of 66.01% and significantly higher F-score in detecting the most dangerous cases.

Issues and Perspectives from 10,000 Annotated Financial Social Media Data (2020.lrec-1)

Copied to clipboard

Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Argument-Based Sentiment Analysis on Forward-Looking Statements (2024.findings-acl)

Challenge:	In the NLP community, many researchers have begun to use machine learning on financial and economic data.
Approach:	They present a dataset with 10,000 financial tweets annotated by experts from the front desk and the middle desk in a bank’s treasury.
Outcome:	The annotated financial tweets of a bank's front desk and middle desk are compared against a general sentiment dictionary and a domain-specific dictionary.

Copied to clipboard

Chin-Yi Lin, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Semantics-Preserved Data Augmentation for Aspect-Based Sentiment Analysis (2021.emnlp-main)

Challenge:	Existing models for argument mining are limited in interpreting future-oriented arguments.
Approach:	They propose a categorization of argument units into claims, premises, and scenarios coupled with a unique sentiment analysis framework.
Outcome:	The proposed framework outperforms existing models in most tasks and is more efficient than existing methods.

Copied to clipboard

Ting-Wei Hsu, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Overcoming Copyright Barriers in Corpus Distribution Through Non-Reversible Hashing (2026.acl-long)

Challenge:	Existing methods for data augmentation address data deficiencies and semantic consistency, but they ignore the second issue.
Approach:	They propose a semantics-preserving data augmentation approach that preserves the semantics of a textual sequence.
Outcome:	The proposed method achieves better performance on publicly available datasets and stock price/risk movement prediction scenarios.

Copied to clipboard

Arthur Amalvy, Vincent Labatut, Xavier Bost, Hen-Hsen Huang

Challenge:	Annotated corpora are crucial in the field of natural language processing, but are difficult to exchange among researchers.
Approach:	They propose a method to lawfully share the annotations of any sequential copyrighted corpus.
Outcome:	The proposed method is robust to reasonable divergences in the version of the copyrighted data owned by the user.

Bias in the Ear of the Listener: Assessing Sensitivity in Audio Language Models Across Linguistic, Demographic, and Positional Variations (2026.findings-eacl)

Copied to clipboard

Sheng-Lun Wei, Yu-Ling Liao, Yen-Hua Chang, Hen-Hsen Huang, Hsin-Hsi Chen

Transfer of Frames from English FrameNet to Construct Chinese FrameNet: A Bilingual Corpus-Based Approach (L18-1)

Challenge:	Recent advances extend language understanding beyond text to speech, enabling unified reasoning across modalities.
Approach:	They construct and release a speech-augmented benchmark based on Global MMLU Lite and a data set spanning English, Chinese, and Korean.
Outcome:	The proposed model is robust to demographic factors but sensitive to language and option order, suggesting that speech can amplify structural biases.

Copied to clipboard

Tsung-Han Yang, Hen-Hsen Huang, An-Zi Yen, Hsin-Hsi Chen

Multilinguality

Challenge:	Current publicly available Chinese FrameNet has a relatively low coverage of frames and lexical units compared with other languages.
Approach:	They propose an automatic way to construct Chinese FrameNet using a sentence-aligned English-Chinese bilingual corpus.
Outcome:	The proposed resource can provide frame recommendations acceptable by annotators.

Using Contextually Aligned Online Reviews to Measure LLMs’ Performance Disparities Across Language Varieties (2025.naacl-short)

Copied to clipboard

Zixin Tang, Chieh-Yang Huang, Tsung-che Li, Ho Yin Sam Ng, Hen-Hsen Huang, Ting-Hao Kenneth Huang

Challenge:	Of the world's 7,000 languages, sixty (60) million people speak British English, 23 million speak Taiwan Mandarin, and 10 million speak European Portuguese.
Approach:	They propose a contextually aligned dataset that captures comments in different languages from real-world scenarios.
Outcome:	The proposed approach shows that large language models underperform in Taiwan Mandarin in a sentiment analysis task.

Do Before You Judge: Self-Reference as a Pathway to Better LLM Evaluation (2025.findings-emnlp)

Copied to clipboard

Wei-Hsiang Lin, Sheng-Lun Wei, Hen-Hsen Huang, Hsin-Hsi Chen

Challenge:	LLM-as-Judge frameworks are increasingly popular for AI evaluation, yet research findings on the relationship between models’ generation and judgment abilities remain inconsistent.
Approach:	They propose a self-reference-guided evaluation strategy that leverages a model’s own answers as references to strengthen the correlation between generation and judgment abilities.
Outcome:	The proposed approach strengthens the correlation between model generation and judgment abilities and provides a reliable proxy for model selection in evaluation tasks.

SEEN: Structured Event Enhancement Network for Explainable Need Detection of Information Recall Assistance (2022.emnlp-main)

Copied to clipboard

You-En Lin, An-Zi Yen, Hen-Hsen Huang, Hsin-Hsi Chen

Self-Augmented Preference Alignment for Sycophancy Reduction in LLMs (2025.emnlp-main)

Challenge:	Existing work on information recall focuses on reactively retrieving life events . but, proactively detecting the need for information recall services is rarely discussed .
Approach:	They propose a human-annotated life experience retelling dataset to detect the right time to trigger an information recall service.
Outcome:	The proposed system detects life event inconsistency, additional information in life events, and forgotten events.

Copied to clipboard

Chien Hung Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Natural Language Interfaces

Challenge:	Sycophantic behavior in models can erode user trust by creating a perception of dishonesty or bias.
Approach:	They propose to assess the user’s expected answer rather than ignore it and introduce self-augmented preference alignment to reduce sycophancy.
Outcome:	The proposed methods significantly reduce sycophancy across tasks and improve models' assessment ability.

A Chinese Writing Correction System for Learning Chinese as a Foreign Language (C18-2)

Copied to clipboard

Yow-Ting Shiue, Hen-Hsen Huang, Hsin-Hsi Chen

Machine Translation

Challenge:	a new writing correction system for Chinese learners is available for learning as a second language . a classification approach to English GEC does not require exact recognition of error types . however, there is no general model that handles all types of Chinese writing errors.
Approach:	They propose a Chinese writing correction system that takes a wrong input sentence and generates correction suggestions.
Outcome:	The proposed system generates correction suggestions for Chinese sentences with English translations, helping users understand correct usages of certain grammar patterns.

GenSense: A Generalized Sense Retrofitting Model (C18-1)

Copied to clipboard

Yang-Yin Lee, Ting-Yu Yen, Hen-Hsen Huang, Yow-Ting Shiue, Hsin-Hsi Chen

ZARA: Improving Few-Shot Self-Rationalization for Small Language Models (2023.findings-emnlp)

Challenge:	Existing word embedding models use only one vector to represent a word, which is problematic in some natural language processing tasks that require sense level representation.
Approach:	They propose a generalized sense embedding learning framework that integrates with the semantic relations between the senses, the relation strength and the semantic strength.
Outcome:	The proposed model outperforms previous models in three types of experiments: semantic relatedness, contextual word similarity and semantic difference.

Copied to clipboard

Wei-Lin Chen, An-Zi Yen, Cheng-Kuang Wu, Hen-Hsen Huang, Hsin-Hsi Chen

Learning to Generate Explanation from e-Hospital Services for Medical Suggestion (2022.coling-1)

Challenge:	Recent studies demonstrate great performance gain for self-rationalization by few-shot prompting LMs with rationale-augmented exemplars.
Approach:	They propose to leverage explanations for small LMs to improve few-shot self-rationalization by reducing the problem of plausibility judgement to natural language inference.
Outcome:	The proposed approach achieves SOTA performance on the FEB benchmark, for both the task accuracy and the explanation metric.

Copied to clipboard

Wei-Lin Chen, An-Zi Yen, Hen-Hsen Huang, Hsin-Hsi Chen

Learning to Map Natural Language Statements into Knowledge Base Representations for Knowledge Base Construction (L18-1)

Challenge:	Neural models have shown remarkable success in various tasks, however, simply offering the predictions may not satisfy the requirement of end-users.
Approach:	They propose a novel model which generates a medical suggestion and provides an explanation as the outline of the reasoning.
Outcome:	The proposed model achieves promising performances in both quantitative and human evaluation.

Copied to clipboard

Chin-Ho Lin, Hen-Hsen Huang, Hsin-Hsi Chen

A Complete Shift-Reduce Chinese Discourse Parser with Robust Dynamic Oracle (2020.acl-main)

Challenge:	Currently, the construction and updating of knowledge bases rely on human labor.
Approach:	They propose to map relational phrases in triples from natural language to knowledge base predicate format.
Outcome:	The proposed mapping results show high quality and promising coverage on relational phrases compared to previous research.

Copied to clipboard

Shyh-Shiun Hung, Hen-Hsen Huang, Hsin-Hsi Chen

Robustness in NLP

Challenge:	Existing work on hierarchical discourse parsing in English is based on the RST-style one.
Approach:	They propose a Chinese discourse parser that integrates pre-trained text encoder and employs novel training strategies to improve rhetorical relation recognition.
Outcome:	The proposed system achieves state-of-the-art performance in Chinese discourse parsing.

LED: A Dataset for Life Event Extraction from Dialogs (2023.findings-eacl)

Copied to clipboard

Yi-Pei Chen, An-Zi Yen, Hen-Hsen Huang, Hideki Nakayama, Hsin-Hsi Chen

Information Extraction & Text Mining

Challenge:	Existing methods for extracting life events from conversations are limited.
Approach:	They propose a dataset containing fine-grained life event annotations on conversational data.
Outcome:	The proposed dataset combines three information extraction frameworks to extract life events from conversations.

Enhancing Society-Undermining Disinformation Detection through Fine-Grained Sentiment Analysis Pre-Finetuning (2024.findings-eacl)

Copied to clipboard

Tsung-Hsuan Pan, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Language Models

Challenge:	a new method for disinformation detection is needed to address the issue of disinformation, authors argue . a series of rigorous experiments establishes a notable connection between disinformation and fine-grained sentiment labels .
Approach:	They propose a method leveraging pre-finetuning concept for efficient detection and removal of disinformation that may undermine society.
Outcome:	The proposed method improves performance across languages and languages, showing promising results.

Disambiguating False-Alarm Hashtag Usages in Tweets for Irony Detection (P18-2)

Copied to clipboard

Hen-Hsen Huang, Chiao-Chen Chen, Hsin-Hsi Chen

Efficient Beam Search for Large Language Models Using Trie-Based Decoding (2025.emnlp-main)

Challenge:	Existing methods to collect self-labeled data for irony detection are based on false-alarm hashtags.
Approach:	They propose a neural network-based model which disambiguates hashtag usages and prunes the self-labeled training data.
Outcome:	The proposed model outperforms the models trained on the less but cleaner training instances.

Copied to clipboard

Brian J Chan, Mao-xun Huang, Jui-Hung Cheng, Chao-Ting Chen, Hen-Hsen Huang

Language Models

Challenge:	Large language models (LLMs) face memorybound performance bottlenecks due to their high memory requirements.
Approach:	They propose a trie-based parallel decoding method that shares a single KV cache across beams with common prefixes to dramatically reduce memory usage and enables efficient decoding.
Outcome:	The proposed method significantly reduces memory usage and enables efficient decoding without compromising generation quality.

Strategy-Induct: Task-Level Strategy Induction for Instruction Generation (2026.findings-acl)

Copied to clipboard

Po-Chun Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Low-Resource NLP

Challenge:	Existing methods for task-level instruction generation rely on input-output pairs . obtaining labeled answers can be difficult or costly, limiting generalization across architectures.
Approach:	They propose a framework that derives task-level instructions solely from a small set of example questions without requiring labeled answers.
Outcome:	The proposed framework outperforms state-of-the-art methods in question-only settings.

NumHG: A Dataset for Number-Focused Headline Generation (2024.lrec-main)

Copied to clipboard

Jian-Tao Huang, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Text Generation

Challenge:	a lack of fine-grained annotations for accurate numeral generation in headlines is a major roadblock . a new dataset, the NumHG, provides over 27,000 annotated numeral-rich news articles for detailed investigation .
Approach:	They propose a dataset that provides annotated numerals for headline generation . they evaluate five well-performing headline-generation models using human evaluation .
Outcome:	The proposed dataset provides annotated numeral-rich news articles for detailed investigation.

Dynamic Graph Transformer for Implicit Tag Recognition (2021.eacl-main)

Copied to clipboard

Yi-Ting Liou, Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen

Structured Data in NLP

Challenge:	Existing studies focus on using explicit information in articles and do not consider the implicit information.
Approach:	They propose a dynamic graph transformer that distills the textual information and the entity relations on the fly.
Outcome:	The proposed model can extract the textual information and the entity relations on the fly.

No One Fits All: From Fixed Prompting to Learned Routing in Multilingual LLMs (2026.findings-acl)

Copied to clipboard

Wei-Chi Wu, Sheng-Lun Wei, Hen-Hsen Huang, Hsin-Hsi Chen

Multilinguality

Challenge:	Existing studies show that translation-based prompting is not universally optimal for multilingual LLMs.
Approach:	They evaluate translation-based prompting across ten languages and four benchmarks . they propose a lightweight classifier that predicts whether native or translation- based prompts are optimal .
Outcome:	The proposed classifiers achieve statistically significant improvements over fixed prompting strategies across ten languages and four benchmarks.

Correcting Chinese Word Usage Errors for Learning Chinese as a Second Language (C18-1)

Copied to clipboard

Yow-Ting Shiue, Hen-Hsen Huang, Hsin-Hsi Chen

Representation Learning

Challenge:	a word usage error is the most common error type in Chinese, according to the HSK dynamic composition corpus . a system that considers both target erroneous token and context can generate a correction vector .
Approach:	They propose a neural network model that considers target erroneous token and context to generate a correction vector and compare it against a candidate vocabulary to propose suitable corrections.
Outcome:	The proposed model can detect 91% of the cases and propose suitable corrections within a list of five candidates.

A Unified RvNN Framework for End-to-End Chinese Discourse Parsing (C18-2)

Copied to clipboard

Lin Chuan-An, Hen-Hsen Huang, Zi-Yuan Chen, Hsin-Hsi Chen

Induct-Learn: Short Phrase Prompting with Instruction Induction (2024.emnlp-main)

Challenge:	Existing work on Chinese discourse parser relies on external packages to extract linguistic features from free text.
Approach:	They propose an end-to-end Chinese discourse parser based on recursive neural network to jointly model the subtasks including elementary discourse unit segmentation, tree structure construction, center labeling, and sense labeling.
Outcome:	The proposed framework achieves state-of-the-art in the Chinese Discourse Treebank dataset.

Copied to clipboard

Po-Chun Chen, Sheng-Lun Wei, Hen-Hsen Huang, Hsin-Hsi Chen