Papers by Danushka Bollegala

Do not let the history haunt you: Mitigating Compounding Errors in Conversational Question Answering (2020.lrec-1)

Challenge:	Abstractive summarization is a novel method for opinionated texts . it uses a recursive Gaussian mixture to generate topic sentences .
Approach:	They propose an unsupervised abstractive summarization method for opinionated texts . they alternate the unimodal Gaussian prior with a recursive Gausssian mixture .
Outcome:	The proposed method generates topic sentences with tree-structured topic guidance, which are more informative and cover more input contents than the current model.

Copied to clipboard

Angrosh Mandya, James O’ Neill, Danushka Bollegala, Frans Coenen

Natural Language Interfaces

Challenge:	Existing approaches employ human-written ground-truth answers for answering conversational questions at test time, but in a realistic scenario, the CoQA model will not have access to ground-Truth answers.
Approach:	They propose a sampling strategy that dynamically selects between target answers and model predictions during training, closely simulating the situation at test time.
Outcome:	The proposed sampling strategy closely simulates the situation at test time and significantly lowers the performance of CoQA systems.

Learning Word Meta-Embeddings by Autoencoding (C18-1)

Copied to clipboard

Danushka Bollegala, Cong Bao

Dictionary-based Debiasing of Pre-trained Word Embeddings (2021.eacl-main)

Challenge:	Existing word embeddings have shown superior performance in numerous Natural Language Processing (NLP) tasks, however, their performances vary significantly across different tasks.
Approach:	They propose to combine distributed word embeddings to produce more accurate and complete meta-embeddings of words.
Outcome:	The proposed meta-embeddings outperform the state-of-the-art in multiple tasks.

Copied to clipboard

An Empirical Study on Fine-Grained Named Entity Recognition (C18-1)

Challenge:	Existing methods for learning word embeddings using dictionaries do not require access to training resources or knowledge regarding the word embeds used.
Approach:	They propose a method for debiasing pre-trained word embeddings using dictionaries . they learn constraints that must be satisfied by unbiased word embeds from dictionary definitions .
Outcome:	The proposed method removes unfair biases encoded in pre-trained word embeddings while preserving useful semantics.

Copied to clipboard

Khai Mai, Thai-Hoang Pham, Minh Trung Nguyen, Tuan Duc Nguyen, Danushka Bollegala, Ryohei Sasano, Satoshi Sekine

Can Word Sense Distribution Detect Semantic Changes of Words? (2023.findings-emnlp)

Challenge:	Named entity recognition (NER) is a well studied topic in natural language processing.
Approach:	They propose to remove the CNN layer and use dictionary and category embeddings to improve Japanese FG-NER performance.
Outcome:	The proposed method improves Japanese FG-NER F-score from 66.76% to 75.18%.

Copied to clipboard

Xiaohang Tang, Yi Zhou, Taichi Aida, Procheta Sen, Danushka Bollegala

A Semantic Distance Metric Learning approach for Lexical Semantic Change Detection (2024.findings-acl)

Challenge:	Existing methods to detect semantic variations of words are not accurate for time-sensitive predictions.
Approach:	They propose to use pretrained static sense embeddings to annotate a word's occurrence with a sense id to compare its distributions.
Outcome:	The proposed method compares word sense distributions across two corpora to predict meaning change . the results show that pretrained LLMs can detect changes in words over time .

Copied to clipboard

Evaluating Unsupervised Dimensionality Reduction Methods for Pretrained Sentence Embeddings (2024.lrec-main)

Challenge:	Existing Word-in-Context (WiC) datasets are used to detect temporal semantic changes of words.
Approach:	They propose a supervised two-staged SCD method that uses existing Word-in-Context (WiC) datasets to predict temporal semantic changes of words.
Outcome:	The proposed method achieves strong performance in multiple languages and significant improvements on WiC benchmarks.

Copied to clipboard

Gaifan Zhang, Yi Zhou, Danushka Bollegala

A Multilingual Social Bias Benchmark Incorporating Thinking Processes (2026.acl-long)

Challenge:	Sentence embeddings produced by pretrained language models are high dimensional (ca. 1024-4096) this is problematic when representing large numbers of sentences in memory- or compute-constrained devices.
Approach:	They propose to use Principal Component Analysis to reduce the dimensionality of sentence embeddings produced by pretrained language models to reduce their complexity.
Outcome:	The proposed methods reduce the dimensionality of sentence embeddings by 50% without incurring significant loss in performance in multiple downstream tasks.

Copied to clipboard

Masahiro Kaneko, Danushka Bollegala, Timothy Baldwin

In-Contextual Gender Bias Suppression for Large Language Models (2024.findings-eacl)

Challenge:	Large Language Models (LLMs) can learn useful knowledge and harmful stereotypes, making bias evaluation essential.
Approach:	They propose a multilingual social bias benchmark that incorporates human-generated reasoning as part of the thinking process.
Outcome:	The proposed method demonstrates superior performance over LLM-generated methods . human-generated thinking yields higher-quality evaluations than template-based approaches .

Copied to clipboard

Daisuke Oba, Masahiro Kaneko, Danushka Bollegala

Sentiment-Stance-Specificity (SSS) Dataset: Identifying Support-based Entailment among Opinions. (L18-1)

Challenge:	Prior work has proposed debiasing methods that require human labelled examples, data augmentation and fine-tuning of LLMs, which are computationally expensive.
Approach:	They propose to suppress gender biases by providing textual preambles from manually designed templates and real-world statistics without accessing model parameters.
Outcome:	The proposed methods suppress gender biases in English LLMs using a CrowsPairs dataset without accessing model parameters.

Copied to clipboard

Pavithra Rajendran, Danushka Bollegala, Simon Parsons

Map of Encoders – Mapping Sentence Encoders using Quantum Relative Entropy (2026.acl-long)

Challenge:	Argument mining is a method for extracting argument components and structures from natural language texts.
Approach:	They propose to model arguments as a set of premises that either support each other or collectively support a conclusion.
Outcome:	The proposed rules give an overall accuracy of 0.83 for the three datasets.

Copied to clipboard

Gaifan Zhang, Danushka Bollegala

Annotating Training Data for Conditional Semantic Textual Similarity Measurement using Large Language Models (2025.emnlp-main)

Challenge:	a method to compare and visualise sentence encoders at scale is proposed . we map encoder LLMs using QRE-based feature vectors, which are then projected to 2D .
Approach:	They propose a method to compare and visualise sentence encoders at scale by creating a map of encoder . they construct a QRE-based map of sentences covering 1101 publicly available sentence encoded sentences .
Outcome:	The proposed method compares sentence encoders at scale by creating a map of encoder models . it shows that the map accurately reflects relationships between encoder and unit base encoder .

Copied to clipboard

Gaifan Zhang, Yi Zhou, Danushka Bollegala

Unsupervised Semantic Variation Prediction using the Distribution of Sibling Embeddings (2023.findings-acl)

Challenge:	Semantic similarity between two sentences depends on the aspects considered between those sentences.
Approach:	They propose a Conditional Semantic Textual Similarity task which measures the similarity between two sentences under a specified condition.
Outcome:	The proposed method improves Spearman correlation by 5.4% by training a supervised model on the re-annotated dataset.

Copied to clipboard

Graph Convolution over Multiple Dependency Sub-graphs for Relation Extraction (2020.coling-main)

Challenge:	Existing work on semantic variation prediction has focused on comparing an averaged contextualised representation of a word . however, some of the previously associated meanings of . a target word can become obsolete over time, while novel usages of existing words are observed.
Approach:	They propose a method that uses the entire cohort of contextualised embeddings of a target word to detect the semantic variation of words.
Outcome:	The proposed method outperforms existing methods on a SemEval-2020 benchmark dataset and is comparable to the state-of-the-art.

Copied to clipboard

Angrosh Mandya, Danushka Bollegala, Frans Coenen

Multimodality

Challenge:	Existing graph convolutional networks use pruned dependency trees to filter irrelevant nodes from sentence graphs.
Approach:	They propose to construct multiple sub-graphs from shortest dependency path and words linked to entities in the dependency parse to obtain more informative features useful for relation extraction.
Outcome:	The proposed method achieves state-of-the-art performance on a sentence-level relation extraction dataset and the SemEval 2010 Task 8 sentence- level relation extraction data.

Multi-Source Attention for Unsupervised Domain Adaptation (2020.aacl-main)

Copied to clipboard

Xia Cui, Danushka Bollegala

Low-Resource NLP

Challenge:	Existing approaches for domain adaptation (UDA) focus on adapting to a domain from a single source domain, but labelled instances are not available for the target domain.
Approach:	They propose to model source-selection in unsupervised domain adaptation as an attention-learning problem, where attention is learned over the sources per given target instance.
Outcome:	The proposed method outperforms previous proposed methods on two cross-domain sentiment classification datasets and is able to explain the predictions.

Gender Bias in Masked Language Models for Multiple Languages (2022.naacl-main)

Copied to clipboard

Masahiro Kaneko, Aizhan Imankulova, Danushka Bollegala, Naoaki Okazaki

SCDTour: Embedding Axis Ordering and Merging for Interpretable Semantic Change Detection (2025.findings-emnlp)

Challenge:	Masked Language Models (MLMs) pre-trained by predicting masked tokens on large corpora have been used successfully in natural language processing tasks for a variety of languages.
Approach:	They propose to use English attribute word lists to evaluate bias in eight languages without manually annotating data.
Outcome:	The proposed model significantly correlates with the existing English datasets for gender bias.

Copied to clipboard

Explainability & Interpretability in NLP

Challenge:	Existing methods to improve interpretability of SCD often lead to performance degradation . agglomerating axes produces a more refined set of word senses, which improves performance .
Approach:	They propose a method that orders and merges interpretable axes to improve SCD performance.
Outcome:	The proposed method preserves performance while maintaining high interpretability . it produces a more refined set of word senses, which improves performance .

Learning Dynamic Contextualised Word Embeddings via Template-based Temporal Adaptation (2023.acl-long)

Copied to clipboard

Xiaohang Tang, Yi Zhou, Danushka Bollegala

Query Obfuscation by Semantic Decomposition (2022.lrec-1)

Challenge:	Existing methods for learning dynamic contextualised word embeddings do not capture temporal semantic variations of words.
Approach:	They propose a method for learning DCWEs by time-adapting a pretrained Masked Language Model using time-sensitive templates.
Outcome:	The proposed method significantly reduces the perplexity of test sentences in C2 outperforming the current state-of-the-art.

Copied to clipboard

Danushka Bollegala, Tomoya Machide, Ken-ichi Kawarabayashi

Learning to Borrow– Relation Representation for Without-Mention Entity-Pairs for Knowledge Graph Completion (2022.naacl-main)

Challenge:	Existing methods to protect privacy of search engine users by decomposing queries using semantically related and unrelated distractor terms are not available to most web search engines.
Approach:	They propose a method to protect the privacy of search engine users by decomposing queries using semantically related and unrelated distractor terms.
Outcome:	The proposed method can reconstruct search results relevant to the original query term without compromising the privacy of the search engine users.

Copied to clipboard

Huda Hakami, Mona Hakami, Angrosh Mandya, Danushka Bollegala

Solving Cosine Similarity Underestimation between High Frequency Words by ℓ2 Norm Discounting (2023.findings-acl)

Challenge:	Existing methods to integrate text corpora with knowledge graphs (KGs) have been effective in various NLP tasks such as analyzing and predicting relationships between entities.
Approach:	They propose a method that borrows LDPs from entities that co-occur in sentences to represent entities that do not co-exist in a single sentence.
Outcome:	The proposed method improves the performance of prior methods such as TransE, DistMult, ComplEx and RotatE.

Copied to clipboard

Saeth Wannasuphoprasit, Yi Zhou, Danushka Bollegala

RelWalk - A Latent Variable Model Approach to Knowledge Graph Embedding (2021.eacl-main)

Challenge:	Existing methods to estimate cosine similarity between words have been proposed, but no solution has been proposed.
Approach:	They propose a method to discount the l2 norm of contextualised word embedding by the frequency of that word in a corpus when measuring the cosine similarities between words.
Outcome:	The proposed method underestimates the similarity scores between two words when used with contextualised token embeddings from masked language models such as BERT.

Copied to clipboard

Danushka Bollegala, Huda Hakami, Yuichi Yoshida, Ken-ichi Kawarabayashi

Evaluating the Robustness of Discrete Prompts (2023.eacl-main)

Challenge:	Existing methods for embedding entities and relations in knowledge graphs are heuristically motivated and theoretical understanding of such embeddables is underdeveloped.
Approach:	They extend the random walk model of word embeddings to Knowledge Graph Embeddings (KGEs) they propose a learning objective motivated by the theoretical analysis to learn KGEs from a given knowledge graph.
Outcome:	The proposed learning objective is motivated by the theoretical analysis to learn KGEs from a given knowledge graph.

Copied to clipboard

Yoichi Ishibashi, Danushka Bollegala, Katsuhito Sudoh, Satoshi Nakamura

Debiasing Pre-trained Contextualised Embeddings (2021.eacl-main)

Challenge:	Existing methods that generate discrete prompts from a small set of training instances have reported superior performance, but manual writing prompts that generalize well is challenging due to several reasons.
Approach:	They propose to use discrete prompts to learn lexical constructs that would not be encountered in manually-written prompts.
Outcome:	The proposed method is robust against perturbations to NLI inputs but sensitive to other types of perturbations such as shuffling and deletion of prompt tokens.

Copied to clipboard

Investigating the Contextualised Word Embedding Dimensions Specified for Contextual and Temporal Semantic Changes (2025.coling-main)

Challenge:	a study of contextualised word embeddings shows discriminative biases are encoded in contextualised embeddables.
Approach:	They propose a fine-tuning method that can be applied at token- or sentence-levels to debias pre-trained contextualised embeddings.
Outcome:	The proposed method can be applied at token- or sentence-levels to debias pre-trained models without requiring retrains.

Copied to clipboard

Evaluating the Evaluation of Diversity in Commonsense Generation (2025.acl-long)

Challenge:	Existing studies on the meaning of contextualised word embeddings (SCWEs) have not shown how meaning changes are encoded in the embeddable space.
Approach:	They compare pre-trained and fine-tuned contextualised word embeddings on contextual and temporal semantic change detection benchmarks.
Outcome:	The pre-trained and fine-tuned versions of (SCWE) and their fine- tuned versions on contextual and temporal semantic change detection benchmarks show that they represent semantic changes across all dimensions when fine--and that they are more efficient than ICA.

Copied to clipboard

Tianhui Zhang, Bei Peng, Danushka Bollegala

Improving Diversity of Commonsense Generation by Large Language Models via In-Context Learning (2024.findings-emnlp)

Challenge:	Existing evaluation metrics for commonsense generation are unclear on which metrics are best suited for evaluating the diversity of outputs.
Approach:	They propose to use a large language model to analyze commonsense generation data to determine which diversity metrics are best suited for commonsensing.
Outcome:	The proposed metrics outperform form-based metrics and show high correlations with the LLM-based ratings.

Copied to clipboard

Tianhui Zhang, Bei Peng, Danushka Bollegala

Is Something Better than Nothing? Automatically Predicting Stance-based Arguments Using Deep Learning and Small Labelled Dataset (N18-2)

Challenge:	Large Language Models (LLMs) have shown proficiency in enhancing the generation quality across various tasks without the need for any fine-tuning.
Approach:	They propose a method that diversifies the LLM generations while preserving their quality.
Outcome:	The proposed method can be used as training data to improve diversity in existing commonsense generators.

Copied to clipboard

Pavithra Rajendran, Danushka Bollegala, Simon Parsons

Text Classification

Challenge:	Argument mining is a subset of NLP that deals with extracting arguments from user-based content.
Approach:	They propose to use weakly supervised and semi-supervised methods to automatically annotate reviews and provide large annotated datasets.
Outcome:	The proposed methods can be used to learn better models for implicit/explicit opinion classification.

Tree-Structured Neural Topic Model (2020.acl-main)

Copied to clipboard

Masaru Isonuma, Junichiro Mori, Danushka Bollegala, Ichiro Sakata

Debiasing Isn’t Enough! – on the Effectiveness of Debiasing MLMs and Their Social Biases in Downstream Tasks (2022.coling-1)

Challenge:	Existing topic models do not organize topics into coherent groups or hierarchies.
Approach:	They propose a tree-structured neural topic model with an infinite number of branches and a topic distribution over a forest.
Outcome:	The proposed model improves data scalability and competitive performance when inducing latent topics and tree structures.

Copied to clipboard

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Together We Make Sense–Learning Meta-Sense Embeddings (2023.findings-acl)

Challenge:	Existing measures for social bias evaluation are inadequate for MLMs to accurately evaluate the social biases in their systems.
Approach:	They propose task-agnostic intrinsic and task-specific extrinsic social bias evaluation measures for MLMs that use different methods to re-learn social biases during fine-tuning on downstream tasks.
Outcome:	The findings highlight the limitations of existing MLM bias evaluation measures and raise concerns on the deployment of MLMs in downstream applications using those measures.

Copied to clipboard

Haochen Luo, Yi Zhou, Danushka Bollegala

Evaluating Short-Term Temporal Fluctuations of Social Biases in Social Media Data and Masked Language Models (2024.emnlp-main)

Challenge:	Existing sense embeddings do not cover all senses of ambiguous words equally well due to discrepancies in their training resources.
Approach:	They propose a meta-sense embedding method that preserves sense neighbourhoods by combining multiple independently trained source sense embeddables.
Outcome:	The proposed method outperforms several baselines on Word Sense Disambiguation and Word-in-Context tasks.

Copied to clipboard

Yi Zhou, Danushka Bollegala, Jose Camacho-Collados

Language-Independent Tokenisation Rivals Language-Specific Tokenisation for Word Similarity Prediction (2020.lrec-1)

Challenge:	Social biases such as gender or racial biase are reported in language models . a recent study has shown that MLMs encode discriminatory social biase .
Approach:	They analyse temporal corpora of MLMs trained on chronologically ordered temporal snapshots . they find that gender and racial biases are encoded in MLM models .
Outcome:	The proposed model identifies gender biases in MLMs but most remain stable over time . gender bias is associated with higher likelihood scores in some demographic groups .

Copied to clipboard

Danushka Bollegala, Ryuichi Kiryo, Kosuke Tsujino, Haruki Yukawa

The Gaps between Fine Tuning and In-context Learning in Bias Evaluation and Debiasing (2025.coling-main)

Challenge:	Language-independent tokenisation (LIT) methods that do not require labelled language resources or lexicons have gained popularity because of their compactness and ability to handle unseen or rare words.
Approach:	They empirically compare language-independent tokenisation methods with language-specific tokenisation (LST) methods using carefully created lexicons and training resources.
Outcome:	The proposed methods outperform LIT and LST on evaluation tasks across eight languages.

Copied to clipboard

Masahiro Kaneko, Danushka Bollegala, Timothy Baldwin

Swap and Predict – Predicting the Semantic Changes in Words across Corpora by Context Swapping (2023.findings-emnlp)

Challenge:	FT-based debiasing methods cause a performance degradation in downstream tasks . FT works by updating some or all parameters, while ICL uses prompts without modifying the model parameters.
Approach:	They propose to use ICL to customize PLMs to downstream tasks without parameter updates.
Outcome:	The proposed method lowers the performance degradation of FT-based debiasing methods compared to FT models . the proposed method improves performance on large datasets while allowing for smaller changes to PLMs .

Copied to clipboard

Gender-preserving Debiasing for Pre-trained Word Embeddings (P19-1)

Challenge:	Detecting semantic changes of words is an important task for various NLP applications that must make time-sensitive predictions.
Approach:	They propose a method that randomly swaps contexts between two different corpora to detect whether a given word changes its meaning . they then use a pretrained masked language model to generate contextualised word embeddings of w, which are then used to predict the semantic changes of words in four languages .
Outcome:	The proposed method achieves significant performance improvements compared to baselines for the English semantic change prediction task.

Copied to clipboard

Bias Mitigation or Cultural Commonsense? Evaluating LLMs with a Japanese Dataset (2025.emnlp-main)

Challenge:	Existing methods for debiasing word embeddings have shown discriminative biases . word embeds learnt from social media have shown to encode racist, offensive and discriminative language usage.
Approach:	They propose a method that preserves gender-related information while removing stereotypical gender biases from pre-trained word embeddings.
Outcome:	The proposed method preserves gender-related information while removing stereotypical discriminative gender biases from pre-trained word embeddings.

Copied to clipboard

Taisei Yamamoto, Ryoma Kumon, Danushka Bollegala, Hitomi Yanaka

Gender Bias in Meta-Embeddings (2022.findings-emnlp)

Challenge:	Recent studies have demonstrated that large language models exhibit social biases . however, debiasing methods may degrade the capabilities of LLMs if they are not properly evaluated .
Approach:	They propose a Japanese benchmark to evaluate social biases and cultural commonsense in large language models in a unified format.
Outcome:	The proposed method degrades the performance of the LLMs on the cultural commonsense task by 75%.

Copied to clipboard

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Joint Learning of Sense and Word Embeddings (L18-1)

Challenge:	Existing methods to develop meta-embeddings from source embeddings contain unfair gender-related biases, and how these influence the meta-bedding has not been studied yet.
Approach:	They propose to use multiple debiasing methods on a single source embedding to create a gender-based meta-embedding.
Outcome:	The proposed method amplifies gender biases compared to input source embeddings.

Copied to clipboard

Mohammed Alsuhaibani, Danushka Bollegala

Representation Learning

Challenge:	Existing methods for learning lower-dimensional representations of words using unlabelled data learn a single representation for a word, ignoring the different senses of that word (polysemy).
Approach:	They propose a method that jointly learns sense-aware word embeddings using both unlabelled and sense-tagged text corpora.
Outcome:	The proposed method outperforms competing methods on word similarity and short-text classification benchmark datasets.

Sense Embeddings are also Biased – Evaluating Social Biases in Static and Contextualised Sense Embeddings (2022.acl-long)

Copied to clipboard

Yi Zhou, Masahiro Kaneko, Danushka Bollegala

Representation Learning

Challenge:	Existing studies have evaluated social biases in word embeddings, but they are understudied.
Approach:	They propose to evaluate the social biases in sense embeddings using a benchmark dataset for word embedders.
Outcome:	The proposed measures show that even when no biases are found at word-level, there are still worrying levels of social biase at sense-level which are often ignored by the word- level bias evaluation measures.

Detect and Classify – Joint Span Detection and Classification for Health Outcomes (2021.emnlp-main)

Copied to clipboard

Micheal Abaho, Danushka Bollegala, Paula Williamson, Susanna Dodd

Autoencoding Improves Pre-trained Word Embeddings (2020.coling-main)

Challenge:	Existing methods for detecting health outcomes from text ignore global structural correspondences between sentence-level and word-level information present in a given text.
Approach:	They propose a method that uses both word-level and sentence-level information to perform outcome span detection and outcome type classification.
Outcome:	The proposed method consistently outperforms decoupled methods, reporting competitive results.

Copied to clipboard

Synthetic Data Generation for Training Diversified Commonsense Reasoning Models (2026.acl-long)

Challenge:	Existing work has shown that word embeddings are distributed in a narrow cone and that centering and projection can improve the accuracy of pre-trained word embeds without requiring additional training data.
Approach:	They propose to remove the top principal components from pre-trained word embeddings and center and project them onto principal component vectors to reinstate isotropy in the embeddable space.
Outcome:	The proposed method is equivalent to applying a linear autoencoder to minimize the squared L2 reconstruction error.

Copied to clipboard

Tianhui Zhang, Bei Peng, Danushka Bollegala

A Dataset for Inter-Sentence Relation Extraction using Distant Supervision (L18-1)

Challenge:	Existing Generative Commonsense Reasoning datasets are created using a small number of human annotators, covering only a narrow set of commonsense scenarios.
Approach:	They propose to use a synthetic dataset to train diverse commonsense generators.
Outcome:	The proposed model improves both generation diversity and quality compared with vanilla models and human-crafted datasets across different size Large Language Models (LLMs).

Copied to clipboard

Angrosh Mandya, Danushka Bollegala, Frans Coenen, Katie Atkinson

I Wish I Would Have Loved This One, But I Didn’t – A Multilingual Dataset for Counterfactual Detection in Product Review (2021.emnlp-main)

Challenge:	Existing methods for intra-sentence relation extraction use a distance supervision method to extract relations between entities.
Approach:	They propose a benchmark dataset for the task of inter-sentence relation extraction using relations previously used for intra-sentent relation extraction.
Outcome:	The proposed dataset is compared with baseline models and recurrent neural network models on the developed dataset.

Copied to clipboard

James O’Neill, Polina Rozenshtein, Ryuichi Kiryo, Motoko Kubota, Danushka Bollegala

Multilinguality

Challenge:	Using machine translation, counterfactual statements are often found in natural languages.
Approach:	They annotate a multilingual CFD dataset from Amazon product reviews covering counterfactuals written in English, German, and Japanese languages.
Outcome:	The proposed dataset is robust against selection biases due to cue phrase-based sentence selection.

Why does PairDiff work? - A Mathematical Analysis of Bilinear Relational Compositional Operators for Analogy Detection (C18-1)

Copied to clipboard

Huda Hakami, Kohei Hayashi, Danushka Bollegala

A Predictive Factor Analysis of Social Biases and Task-Performance in Pretrained Masked Language Models (2023.emnlp-main)

Challenge:	Existing methods for representing semantic relations between words are unclear . identifying relations between word and entity is important for NLP applications .
Approach:	They propose to compute the vector offset between word embeddings to represent relation between two words . they show that PairDiff is an uncorrelated bilinear operator that can be simplified to a linear form .
Outcome:	The proposed method is surprisingly accurate and can be used on multiple word embeddings.

Copied to clipboard

Yi Zhou, Jose Camacho-Collados, Danushka Bollegala

Language Models

Challenge:	Various types of social biases have been reported with pretrained Masked Language Models (MLMs) in prior work.
Approach:	They conduct a comprehensive study on 39 pretrained MLMs to examine their model factors and their social biases.
Outcome:	The proposed model factors influence social biases learned by an MLM and their downstream task performance.

Improving Pre-trained Language Model Sensitivity via Mask Specific losses: A case study on Biomedical NER (2024.naacl-long)

Copied to clipboard

Micheal Abaho, Danushka Bollegala, Gary Leeming, Dan Joyce, Iain Buchan

Language Models

Challenge:	Fine-tuning is the prevailing practice for adapting language models (LMs) to new domains.
Approach:	They propose a mask specific language model that weights the importance of domain-specific terms during fine-tuning to avoid insensitivity.
Outcome:	The proposed approach outperforms advanced masking strategies such as span- and PMI-based masking.

On the Curious Case of l2 norm of Sense Embeddings (2022.findings-emnlp)

Copied to clipboard

Yi Zhou, Danushka Bollegala

Frustratingly Easy Meta-Embedding – Computing Meta-Embeddings by Averaging Source Word Embeddings (N18-2)

Challenge:	l2 norm of sense embeddings encodes information related to frequency of that sense in the training corpus . l2-normal feature is useful for word-in-context (WiC) and word sense disambiguation (WSD)
Approach:	They propose to include the l2 norm of a sense embedding as a feature in a classifier to improve word sense learning methods that use static sense embeds.
Outcome:	The l2 norm of sense embeddings is a surprisingly effective feature for word sense related tasks such as word-in-context (WiC) and word sense disambiguation (WSD).

Copied to clipboard

Joshua Coates, Danushka Bollegala

Representation Learning

Challenge:	Existing methods for producing word embeddings have shown to produce accurate meta-embeddings from pre-trained source embeddables.
Approach:	They propose to use arithmetic mean of two distinct word embedding sets to produce an accurate meta-embedding.
Outcome:	The proposed method produces meta-embeddings comparable or better than more complex methods.

Comparing Intrinsic Gender Bias Evaluation Measures without using Human Annotated Examples (2023.eacl-main)

Copied to clipboard

Masahiro Kaneko, Danushka Bollegala, Naoaki Okazaki

Unsupervised Parsing by Searching for Frequent Word Sequences among Sentences with Equivalent Predicate-Argument Structures (2024.findings-acl)

Challenge:	Existing methods to evaluate gender biases in pre-trained language models have been limited by the cost and difficulties of recruiting human annotators.
Approach:	They propose a method to compare intrinsic gender bias evaluation measures without relying on human annotated examples.
Outcome:	The proposed method compares gender-based gender bias evaluation measures without human annotators without human input.

Copied to clipboard

Junjie Chen, Xiangheng He, Danushka Bollegala, Yusuke Miyao

Evaluating the Effect of Retrieval Augmentation on Social Biases (2026.eacl-long)

Challenge:	Unsupervised constituency parsing focuses on identifying word sequences that form a syntactic unit (i.e., constituents) in target sentences.
Approach:	They propose a frequency-based parser that computes the span-overlap score as the word sequence’s frequency in the PAS-equivalent sentence set and identifies the constituent structure by finding a constituent tree with the maximum span- overlap score.
Outcome:	The proposed method outperforms existing unsupervised parsers in eight out of ten languages and is more accurate than previous methods.

Copied to clipboard

Tianhui Zhang, Yi Zhou, Danushka Bollegala

Information Retrieval

Challenge:	RAG is a popular method for injecting up-to-date knowledge into LLMs.
Approach:	They examine how RAG modulates social biases across three languages and four categories . they find that biased documents are amplified even when base LLM has low-level of intrinsic bias .
Outcome:	The proposed method can enhance factual accuracy but its effect on social biases is not well understood.

Unsupervised Attention-based Sentence-Level Meta-Embeddings from Contextualised Language Models (2022.lrec-1)

Copied to clipboard

Keigo Takahashi, Danushka Bollegala