Papers by Hwee Tou Ng

On the Robustness of Question Rewriting Systems to Questions of Varying Hardness (2022.acl-long)

Challenge:	Existing quality estimation models are not good enough to distinguish good corrections from bad ones, resulting in low F0.5 scores when used for system combination.
Approach:	They propose a new quality estimation model that gives a better estimate of the quality of a corrected sentence.
Outcome:	The proposed model outperforms the state-of-the-art on the CoNLL-2014 and BEA-2019 test sets, and achieves the highest F0.5 scores published to date.

Copied to clipboard

Hai Ye, Hwee Tou Ng, Wenjuan Han

Robustness in NLP

Challenge:	entailment : absence of questions classified based on their rewriting hardness or difficulty . enactment of QR system to rewrite context-dependent questions in CQA requires context knowledge .
Approach:	They propose a heuristic method to automatically classify questions into subsets of varying hardness . they then conduct a human evaluation to annotate the rewriting hardness of questions .
Outcome:	The proposed learning framework improves the overall performance compared to baselines.

From Moments to Milestones: Incremental Timeline Summarization Leveraging Large Language Models (2024.acl-long)

Copied to clipboard

Qisheng Hu, Geonsik Moon, Hwee Tou Ng

Preference-Guided Reflective Sampling for Aligning Language Models (2024.emnlp-main)

Challenge:	Prior work on timeline summarization has neglected the potential synergy between the two forms of timelines.
Approach:	They propose a timeline summarization approach that leverages large language models to generate both event and topic timelines.
Outcome:	The proposed approach outperforms the best existing approaches in four TLS benchmarks.

Copied to clipboard

Hai Ye, Hwee Tou Ng

Domain Generalization for Text Classification with Memory-Based Supervised Contrastive Learning (2022.coling-1)

Challenge:	Repeated random sampling is a widely used method that independently queries the model multiple times to generate outputs.
Approach:	They propose a more efficient method for iterative data generation and model re-training that leverages tree-based tree-derived generation framework to enable more efficient sampling.
Outcome:	The proposed method significantly outperforms repeated random sampling in best-of-N sampling on AlpacaEval and Arena-Hard.

Copied to clipboard

Qingyu Tan, Ruidan He, Lidong Bing, Hwee Tou Ng

Revisiting DocRED - Addressing the False Negative Problem in Relation Extraction (2022.emnlp-main)

Challenge:	Existing approaches to cross-domain text classification focus on one-to-one domain adaptation.
Approach:	They propose a framework for domain generalization that uses contrastive learning with a memory-saving queue.
Outcome:	The proposed framework outperforms state-of-the-art methods on Amazon review sentiment datasets and rumour detection datasets.

Copied to clipboard

Qingyu Tan, Lu Xu, Lidong Bing, Hwee Tou Ng, Sharifah Mahani Aljunied

Are Decoder-Only Language Models Better than Encoder-Only Language Models in Understanding Word Meaning? (2024.findings-acl)

Challenge:	Using incomplete annotations, we find that false negative samples are prevalent in the DocRED dataset . we reannotate 4,053 documents in the dataset by adding the missed relation triples back to the original DocRED.
Approach:	They propose to re-annotate 4,053 documents in the document-level relation extraction dataset by adding missing relation triples back to the original DocRED.
Outcome:	The proposed dataset improves on the existing DocRED dataset by 13 F1 points.

Copied to clipboard

Muhammad Qorib, Geonsik Moon, Hwee Tou Ng

A Co-Attentive Cross-Lingual Neural Model for Dialogue Breakdown Detection (2020.coling-main)

Challenge:	Large language models are highly effective tools for solving different kinds of problems in natural language processing.
Approach:	They propose to use large language models to solve a myriad of problems.
Outcome:	The proposed model performs worse on word meaning comprehension than an encoder-only model with vastly fewer parameters.

Copied to clipboard

Qian Lin, Souvik Kundu, Hwee Tou Ng

Document-Level Relation Extraction with Adaptive Focal Loss and Knowledge Distillation (2022.findings-acl)

Challenge:	Existing models for dialogue breakdown detection do not focus on preventing dialogue breakdowns.
Approach:	They propose a model that integrates a pretrained cross-lingual language model and a co-attention network for dialogue breakdown detection.
Outcome:	The proposed model outperforms all previous approaches on evaluation metrics in Japanese and English tracks in Dialogue Breakdown Detection Challenge 4 .

Copied to clipboard

Qingyu Tan, Ruidan He, Lidong Bing, Hwee Tou Ng

Efficient and Interpretable Grammatical Error Correction with Mixture of Experts (2024.findings-emnlp)

Challenge:	Document-level relation extraction (DocRE) is a more challenging task than sentence-level one.
Approach:	They propose a semi-supervised framework for document-level relation extraction with three components . they use an axial attention module for learning the interdependency among entity-pairs .
Outcome:	The proposed model outperforms baseline models on two DocRE datasets and outperformed previous models on human annotated data and distantly supervised data.

Copied to clipboard

Muhammad Qorib, Alham Aji, Hwee Tou Ng

Responsible & Trustworthy NLP

Challenge:	Error type information has been widely used to improve the performance of grammatical error correction models.
Approach:	They propose a mixture-of-experts model for grammatical error correction that uses error type information to generate corrections and combine models.
Outcome:	The proposed model achieves the performance of T5-XL with three times fewer effective parameters and produces interpretable corrections by also identifying the error type during inference.

Think&Cite: Improving Attributed Text Generation with Self-Guided Tree Search and Progress Reward Modeling (2025.acl-long)

Copied to clipboard

Junyi Li, Hwee Tou Ng

Text Generation

Challenge:	Large language models (LLMs) are prone to hallucinations and producing factually incorrect information.
Approach:	They propose a framework that allows LLMs to generate citations that provide evidence for any statement.
Outcome:	The proposed framework outperforms baseline approaches on three datasets and significantly outperformed baseline approaches.

Mind the Biases: Quantifying Cognitive Biases in Language Model Prompting (2023.findings-acl)

Copied to clipboard

Ruixi Lin, Hwee Tou Ng

Responsible & Trustworthy NLP

Challenge:	Cognitive biases in the human decision making process can lead to flawed responses when we are under uncertainty.
Approach:	They propose to expose cognitive biases on results of language model prompting which display bias modes resembling cognitive bias.
Outcome:	The proposed methods show that a toning-down transformation of the drug-drug description in a prompt can elicit a bias similar to the framing effect, warning users to distrust when prompting language models for answers.

A Nil-Aware Answer Extraction Framework for Question Answering (D18-1)

Copied to clipboard

Souvik Kundu, Hwee Tou Ng

Class-Adaptive Self-Training for Relation Extraction with Incompletely Annotated Training Data (2023.findings-acl)

Challenge:	Recent research suggests that reading comprehension-based question answering systems assume that every question has a valid answer in the associated passage.
Approach:	They propose a novel nil-aware answer span extraction framework that can return Nil or a text span from the associated passage as an answer in a single step.
Outcome:	The proposed framework outperforms baseline approaches on a newsQA dataset.

Copied to clipboard

Qingyu Tan, Lu Xu, Lidong Bing, Hwee Tou Ng

Neural Quality Estimation of Grammatical Error Correction (D18-1)

Challenge:	Existing relation extraction models rely on supervised machine learning, but many datasets are incompletely annotated, causing false negatives and errors during inference stage.
Approach:	They propose a class-adaptive re-sampling self-training framework that favored the pseudo-labels of classes with high precision and low recall scores.
Outcome:	The proposed framework outperforms existing methods on the Re-DocRED and ChemDisgene datasets when the training data are incompletely annotated.

Copied to clipboard

Shamil Chollampatt, Hwee Tou Ng

Mitigating Exposure Bias in Grammatical Error Correction with Data Augmentation and Reweighting (2023.eacl-main)

Challenge:	Grammatical error correction systems are expected to correct most learners’ writing errors, but in practice they often produce spurious corrections and fail to correct many errors, thereby misleading learners.
Approach:	They propose to use supervised learning to estimate the quality of GEC output sentences to help instructors decide whether to correct the errors or ignore them altogether.
Outcome:	The proposed model improves on a feature-based baseline and shows that the state-of-the-art system can be improved when quality scores are used as features for re-ranking the N-best candidates.

Copied to clipboard

Hannan Cao, Wenmian Yang, Hwee Tou Ng

Low-Resource NLP

Challenge:	Existing approaches to grammatical error correction (GEC) use sequence-to-sequence models, but there is an exposure bias problem.
Approach:	They propose a data manipulation approach to overcome the exposure bias problem in seq2seq GEC . they propose augmentation methods to mimic decoder input and reweighting methods to automatically balance the importance of each kind of augmented samples.
Outcome:	The proposed method improves on benchmark GEC datasets.

Multi-Source Test-Time Adaptation as Dueling Bandits for Extractive Question Answering (2023.acl-long)

Copied to clipboard

Hai Ye, Qizhe Xie, Hwee Tou Ng

A Reassessment of Reference-Based Grammatical Error Correction Metrics (C18-1)

Challenge:	Recent research on test-time adaptation suggests a possible way to improve the generalization ability of LLMs.
Approach:	They propose to use multi-armed bandit learning and multi-arm dueling bandits to solve a multi-source test-time model adaptation problem from user feedback.
Outcome:	The proposed model is more effective than other strong baselines on extractive question answering datasets.

Copied to clipboard

Shamil Chollampatt, Hwee Tou Ng

Towards Benchmarking and Improving the Temporal Reasoning Capability of Large Language Models (2023.acl-long)

Challenge:	Existing studies on the correlation of GEC metrics with human judgments were inconclusive . a recent study found that GLEU produces counter-intuitive scores in common test examples .
Approach:	They propose to use GLEU to evaluate grammatical error correction (GEC) systems . they also use statistical significance tests to assess their agreement with human judgments .
Outcome:	The proposed metrics show no significant advantage over MaxMatch (GLEU) the results contradict previous studies that claim GLEU superior .

Copied to clipboard

Qingyu Tan, Hwee Tou Ng, Lidong Bing

Just Go Parallel: Improving the Multilingual Capabilities of Large Language Models (2025.acl-long)

Challenge:	Recent time-dependent question answering datasets tend to be biased in either their coverage of time spans or question types.
Approach:	They propose a temporal reasoning framework based on temporal span extraction and time-sensitive reinforcement learning to improve the temporal ability of large language models.
Outcome:	The proposed framework improves the temporal reasoning capability of large language models by using temporal span extraction and time-sensitive reinforcement learning.

Copied to clipboard

Muhammad Reza Qorib, Junyi Li, Hwee Tou Ng

Multilinguality

Challenge:	Large language models (LLMs) have impressive translation capabilities even without being explicitly trained on parallel data.
Approach:	They propose to add parallel data to enhance multilingual encoder-based and encoder decoder language models by focusing on translation and multilingual common-sense reasoning.
Outcome:	The proposed methods show that adding parallel data can significantly improve LLMs’ multilingual capabilities.

Parametric Knowledge is Not All You Need: Toward Honest Large Language Models via Retrieval of Pretraining Data (2026.findings-acl)

Copied to clipboard

Christopher Adrian Kusuma, Muhammad Reza Qorib, Hwee Tou Ng

Grammatical Error Correction with Contrastive Learning in Low Error Density Domains (2021.findings-emnlp)

Challenge:	Large language models are highly capable of answering questions, but they are often unaware of their own knowledge boundary, i.e., knowing what they know and what they don’t know.
Approach:	They propose a method to evaluate LLM honesty using Pythia with publicly available pretraining data.
Outcome:	The proposed method is based on Pythia, a truly open LLM with publicly available pretraining data.

Copied to clipboard

Hannan Cao, Wenmian Yang, Hwee Tou Ng

Improved Word Sense Disambiguation Using Pre-Trained Contextualized Word Representations (D19-1)

Challenge:	grammatical error correction (GEC) is a text generation task . performance on low error density domains where texts written by native speakers can be improved.
Approach:	They propose a contrastive learning approach to encourage the GEC model to assign a higher probability to a correct sentence while reducing the probability of incorrect sentences that the model tends to generate.
Outcome:	The proposed approach significantly improves the performance of GEC models in low error density domains.

Copied to clipboard

Christian Hadiwinoto, Hwee Tou Ng, Wee Chung Gan

Rationalize and Align: Enhancing Writing Assistance with Rationale via Self-Training for Improved Alignment (2025.findings-acl)

Challenge:	Contextualized word representations are effective in downstream tasks such as question answering, named entity recognition, and sentiment analysis.
Approach:	They propose to integrate pre-trained contextualized word representations into a neural network that captures the whole sentence and the word representation in the sentence.
Outcome:	The proposed approach outperforms the state-of-the-art approach that makes use of non-contextualized word embeddings on multiple benchmark WSD datasets.

Copied to clipboard

Hannan Cao, Hai Ye, Hwee Tou Ng

Challenge:	Existing writing assistants rely on supervised fine-tuning to optimize models for multiple revisions.
Approach:	They propose a framework that enhances WA performance with rationale and alignment.
Outcome:	The proposed framework outperforms state-of-the-art WAs and the closed-source GPT-4o by 3.9 and 7.1 points on average across eight well-established writing-related test sets.

Towards Robust Temporal Reasoning of Large Language Models via a Multi-Hop QA Dataset and Pseudo-Instruction Tuning (2024.findings-acl)

Copied to clipboard

Qingyu Tan, Hwee Tou Ng, Lidong Bing

Upping the Ante: Towards a Better Benchmark for Chinese-to-English Machine Translation (L18-1)

Challenge:	Existing LLMs lack the ability to deal with temporal knowledge.
Approach:	They propose a temporal question-answering dataset Complex-TR that focuses on multi-answered and multi-hop temporal reasoning and propose augmentation strategy to improve LLMs' performance.
Outcome:	The proposed dataset improves LLMs’ performance on temporal QA benchmarks by significant margins.

Copied to clipboard

Christian Hadiwinoto, Hwee Tou Ng

Machine Translation

Challenge:	Currently, there is no widely accepted standard for evaluation of machine translation (MT) for Chinese-to-English translation, there are no standard for standardized training sets, development sets, and test sets.
Approach:	They propose to use Chinese-to-English machine translation as a benchmark . they build a highly competitive state-of-the-art MT system that outperforms reported results .
Outcome:	The proposed system outperforms reported results on NIST OpenMT test sets in almost all papers published in major conferences and journals in computational linguistics and artificial intelligence in the past 11 years.

Learning to Identify Follow-Up Questions in Conversational Question Answering (2020.acl-main)

Copied to clipboard

Souvik Kundu, Qian Lin, Hwee Tou Ng

Finding the Sweet Spot: Preference Data Construction for Scaling Preference Optimization (2025.acl-long)

Challenge:	Recent work on conversational question answering does not focus on follow-up questions . a practical conversational QA system must understand the conversation history well .
Approach:	They propose a three-way attentive pooling network that determines suitability of a follow-up question by capturing pair-wise interactions between associated passage, conversation history, and a candidate follow- up question.
Outcome:	The proposed model outperforms baseline systems by significant margins in the follow-up question identification task.

Copied to clipboard

Yao Xiao, Hai Ye, Linyao Chen, Hwee Tou Ng, Lidong Bing, Xiaoli Li, Roy Ka-Wei Lee

Frustratingly Easy System Combination for Grammatical Error Correction (2022.naacl-main)

Challenge:	Large language models generate unintended outputs due to their unsupervised nature.
Approach:	They propose a method to construct preference pairs of selected and rejected LLMs by repeated random sampling to improve alignment performance.
Outcome:	The proposed method improves performance as the sample size increases.

Copied to clipboard

Muhammad Reza Qorib, Seung-Hoon Na, Hwee Tou Ng

Unsupervised Grammatical Error Correction Rivaling Supervised Methods (2023.emnlp-main)

Challenge:	Using a simple logistic regression algorithm, we combine GEC models for binary classification.
Approach:	They propose a logistic regression algorithm that can combine GEC models with binary classification.
Outcome:	The proposed method outperforms the state-of-the-art by 4.2 points on the CoNLL-2014 and 7.2 points on BEA-2019 test sets.

Copied to clipboard

Hannan Cao, Liping Yuan, Yuchen Zhang, Hwee Tou Ng

Low-Resource NLP

Challenge:	Current state-of-the-art grammatical error correction systems rely on labeled data . current systems require manual correction and require a large quantity of labeles .
Approach:	They propose an unsupervised method to build a grammatical error correction system using a fixer and a critic.
Outcome:	The proposed system outperforms previous unsupervised systems on English and Chinese GEC.

Grammatical Error Correction: Are We There Yet? (2022.coling-1)

Copied to clipboard

Muhammad Reza Qorib, Hwee Tou Ng

Adaptive Semi-supervised Learning for Cross-domain Sentiment Classification (D18-1)

Challenge:	grammatical error correction (GEC) systems outperform humans on the CoNLL-2014 test set, but there are still classes of errors that they fail to correct.
Approach:	They found that state-of-the-art GEC systems outperform humans by a wide margin on the CoNLL-2014 test set . however, they found that there are still classes of errors that they fail to correct .
Outcome:	The F0.5 evaluation metric outperforms the CoNLL-2014 test set, but there are still classes of errors that they fail to correct.

Copied to clipboard

Ruidan He, Wee Sun Lee, Hwee Tou Ng, Daniel Dahlmeier

Information Retrieval

Challenge:	Existing methods for cross-domain sentiment classification are difficult and costly . domain adaptation is difficult because data in source and target domains are drawn from different distributions.
Approach:	They propose a semi-supervised learning approach that minimizes the distance between source and target instances in embedded feature space.
Outcome:	The proposed approach can improve on baseline methods in various settings.

Exploiting Document Knowledge for Aspect-level Sentiment Classification (P18-2)

Copied to clipboard

Ruidan He, Wee Sun Lee, Hwee Tou Ng, Daniel Dahlmeier

Sentiment Analysis

Challenge:	Existing public aspect-level datasets for aspect-based sentiment classification are small . existing methods for aspect level sentiment classification require annotation of all opinion targets .
Approach:	They propose two approaches that transfer knowledge from document-level data to improve aspect-level sentiment classification.
Outcome:	The proposed methods improve aspect-level sentiment classification on 4 public datasets.

Improved Word Sense Disambiguation with Enhanced Sense Representations (2021.findings-emnlp)

Copied to clipboard

Yang Song, Xin Cai Ong, Hwee Tou Ng, Qian Lin

A Survey of Unsupervised Dependency Parsing (2020.coling-main)

Challenge:	Existing supervised word sense disambiguation systems do not provide enough information about word senses.
Approach:	They propose to incorporate synonyms, example phrases or sentences showing usage of word senses and sense gloss of hypernyms into the sense representations.
Outcome:	The proposed system achieves an F1 score of 82.0% on the standard benchmark test dataset of the English all-words WSD task.

Copied to clipboard

Wenjuan Han, Yong Jiang, Hwee Tou Ng, Kewei Tu

Responsible & Trustworthy NLP

Challenge:	Syntactic dependency parsing is an important task in natural language processing . unsupervised learning of dependency parses requires training sentences to be manually annotated with their correct parse trees.
Approach:	They propose to survey existing approaches to unsupervised dependency parsing . they identify two major classes of approaches and discuss recent trends .
Outcome:	The proposed methods can be used in semantic parsing, machine translation, relation extraction, and many other tasks.

Improving the Robustness of Question Answering Systems to Question Paraphrasing (P19-1)

Copied to clipboard

Wee Chung Gan, Hwee Tou Ng

DynaQuest: A Dynamic Question Answering Dataset Reflecting Real-World Knowledge Updates (2025.findings-acl)

Challenge:	Despite advancement of question answering systems, generalizability of QA models is a topic of concern.
Approach:	They propose to use a neural paraphrasing model to generate multiple paraphrased questions for a given source question and a set of paraphrase suggestions to re-train the models.
Outcome:	The proposed approach requires no human intervention to re-train the models for improved robustness to question paraphrasing.

Copied to clipboard

Qian Lin, Junyi Li, Hwee Tou Ng

Question Answering

Challenge:	Large language models (LLMs) are typically trained on static datasets, preventing them from integrating real-time updates.
Approach:	They propose a dynamic question-answer answering dataset reflecting real-world knowledge updates that are automatically compared between Wikipedia versions and generating question-anchor pairs based on these updates.
Outcome:	The proposed framework improves LLMs' performance on time-sensitive question answering by maintaining a dynamic knowledge updating process.

Feature Adaptation of Pre-Trained Language Models across Languages and Domains with Robust Self-Training (2020.emnlp-main)

Copied to clipboard

Hai Ye, Qingyu Tan, Ruidan He, Juntao Li, Hwee Tou Ng, Lidong Bing

Do Multi-Hop Question Answering Systems Know How to Answer the Single-Hop Sub-Questions? (2021.eacl-main)

Challenge:	Adapting pre-trained language models (PrLMs) to new domains has gained much attention . Adaptation of PrLMs to newdomains is important, but requires fine-tuning .
Approach:	They propose to use PrLMs to adapt to new domains without fine-tuning . they use class-aware feature self-distillation to learn discriminative features .
Outcome:	The proposed model can learn discriminative features from pre-trained language models without fine-tuning.

Copied to clipboard

Yixuan Tang, Hwee Tou Ng, Anthony Tung

Question Answering

Challenge:	Existing models fail to answer a large portion of sub-questions . Existing systems have achieved super-human performance .
Approach:	They propose to use a neural decomposition model to generate sub-questions for a multi-hop question and extract the corresponding sub-answers.
Outcome:	The proposed model is based on a hotpotQA dataset with a multi-hop question and sub-answers.

Effective Attention Modeling for Aspect-Level Sentiment Classification (C18-1)

Copied to clipboard

Ruidan He, Wee Sun Lee, Hwee Tou Ng, Daniel Dahlmeier

Sentiment Analysis

Challenge:	Aspect-level sentiment classification aims to determine sentiment polarity of review sentence towards opinion target . main challenge is to separate different opinion contexts for different targets .
Approach:	They propose a method that captures the semantic meaning of the opinion target and a model that incorporates syntactic information into the attention mechanism.
Outcome:	The proposed method captures the semantic meaning of the opinion target and incorporates syntactic information into the attention mechanism.

Robust Question Answering against Distribution Shifts with Test-Time Adaption: An Empirical Study (2022.findings-emnlp)

Copied to clipboard

Hai Ye, Yuyang Ding, Juntao Li, Hwee Tou Ng

Robustness in NLP

Challenge:	Existing work on robustness tuning (RT) methods has found that QA models fail when the test data has a distribution shift compared to the training data.
Approach:	They propose to use test-time adaptation methods to improve QA models after deployment to evaluate their model against text corruption and changes in language and domain.
Outcome:	The proposed method improves TTA to be more robust to variation in hyper-parameters and test distributions over time.

Does BERT Know that the IS-A Relation Is Transitive? (2022.acl-short)

Copied to clipboard

Ruixi Lin, Hwee Tou Ng

An Interactive Multi-Task Learning Network for End-to-End Aspect-Based Sentiment Analysis (P19-1)

Challenge:	Recent studies suggest pre-trained BERT can capture lexico-semantic clues from words in context.
Approach:	They examine word senses and the transitive property of IS-A relation . they aim to quantify how much BERT agrees with transitivity property .
Outcome:	The proposed model can capture lexico-semantic clues from words in context . but to what extent it captures transitive nature of some lexical relations is unclear .

Copied to clipboard

Ruidan He, Wee Sun Lee, Hwee Tou Ng, Daniel Dahlmeier

Sentiment Analysis

Challenge:	Aspect-based sentiment analysis produces a list of aspect terms and their corresponding sentiments for a sentence.
Approach:	They propose an interactive multi-task learning network which can learn multiple tasks simultaneously . they use a shared set of latent variables to iteratively pass information between tasks .
Outcome:	The proposed method outperforms existing methods on three benchmark datasets.

ALLECS: A Lightweight Language Error Correction System (2023.eacl-demo)

Copied to clipboard

Muhammad Reza Qorib, Geonsik Moon, Hwee Tou Ng