Research - ConscientAI

Adapter-based Multi-document Summarisation: Opinion Summarisation Use Case

Kushan Hewapathirana, Nisansa de Silva, C.D. Athuraliya

International Conference on Agents and Artificial Intelligence (ICAART)

Mar 2026

Abstract

This study explores adapter-based fine-tuning to enhance the PRIMERA model for opinion summarisation. PRIMERA, a state-of-the-art multi-document summarisation (MDS) model, exhibits strong transfer potential owing to its pre-training on large-scale MDS corpora. Leveraging adapter architectures, this work demonstrates substantial improvements when extending PRIMERA to opinion summarisation through parameter-efficient fine-tuning. In addition, an LLM-based evaluation paradigm is introduced using the DeepEval framework, enabling semantic and sentiment-aware assessment beyond lexical-overlap metrics such as ROUGE. To improve training efficiency, an agentic optimisation framework is proposed, where evaluation reasoning guides iterative adapter configuration, reducing fine-tuning cycles while maintaining performance. Results show that adapter-augmented PRIMERA surpasses the opinion summarisation baseline ADASUM, establishing a reproducible, interpretable, and computationally efficient path for low-resource MDS. Overall, this work highlights how adapter-based fine-tuning and reasoning-guided optimisation together advance both performance and applicability in opinion summarisation.

Domain Adaptation for Multi-document Summarisation: A Case Study in the Medical Research Domain

Kushan Hewapathirana, Nisansa de Silva, C.D. Athuraliya, Piumi Kandanaarachchi

Pacific Asia Conference on Language, Information and Computation (PACLIC)

Dec 2025

Abstract

Effectively summarising medical research is critical for supporting evidence-based decision-making in healthcare. While fine-tuning task-specific models on domain data is established practice, the comparative advantages over increasingly capable general-purpose LLMs remain an open question. This study systematically evaluates domain-adapted PRIMERA against several open-source large language models (LLaMA 3.2 3B, Mistral 7B, OpenChat 7B, and Gemma 7B) in zero-shot settings using the MS^2 dataset, which includes 20,000 systematic reviews summarising over 470,000 medical studies. Fine-tuning leads to notable improvements in ROUGE scores—ROUGE-1 from 12.8 to 33.0, ROUGE-2 from 2.0 to 6.5, and ROUGE-L from 8.1 to 22.6. Comparative evaluation indicates that the fine-tuned model consistently achieves stronger performance across all three ROUGE metrics, human evaluations, and LLM-as-a-judge assessments. These results suggest that domain-adapted models can offer advantages over general-purpose LLMs in specialised settings, particularly where factual accuracy and coverage are critical, though at the cost of reduced flexibility across domains.

Adapter-based Fine-tuning for PRIMERA

Kushan Hewapathirana, Nisansa de Silva, C.D. Athuraliya

Applied Data Science & Artificial Intelligence Conference (ADScAI)

Apr 2025

Abstract

Multi-document summarisation (MDS) involves generating concise summaries from clusters of related documents. PRIMERA (Pyramid-based Masked Sentence Pre-training for Multi-document Summarisation) is a pre-trained model specifically designed for MDS, utilizing the LED architecture to handle long sequences effectively. Despite its capabilities, fine-tuning PRIMERA for specific tasks remains resourceintensive. To mitigate this, we explore the integration of adapter modules—small, trainable components inserted within transformer layers—that allow models to adapt to new tasks by updating only a fraction of the parameters, thereby reducing computational requirements.

M2DS: Multilingual Dataset for Multi-document Summarisation

Kushan Hewapathirana, Nisansa de Silva, C.D. Athuraliya

International Conference on Computational Collective Intelligence (ICCCI)

Sep 2024

Abstract

In the rapidly evolving digital era, there is an increasing demand for concise information as individuals seek to distil key insights from various sources. Recent attention from researchers on Multi-document Summarisation (MDS) has resulted in diverse datasets covering customer reviews, academic papers, medical and legal documents, and news articles. However, the English-centric nature of these datasets has created a conspicuous void for multilingual datasets in today’s globalised digital landscape, where linguistic diversity is celebrated. Media platforms such as British Broadcasting Corporation (BBC) have disseminated news in 20+ languages for decades. With only 380 million people speaking English natively as their first language, accounting for less than 5% of the global population, the vast majority primarily relies on other languages. These facts underscore the need for inclusivity in MDS research, utilising resources from diverse languages. Recognising this gap, we present the Multilingual Dataset for Multi-document Summarisation (M2DS), which, to the best of our knowledge, is the first dataset of its kind. It includes document-summary pairs in five languages from BBC articles published during the 2010–2023 period. This paper introduces M2DS, emphasising its unique multilingual aspect, and includes baseline scores from state-of-the-art MDS models evaluated on our dataset.

Multi-Document Summarization: A Comparative Evaluation

Kushan Hewapathirana, Nisansa de Silva, C.D. Athuraliya

International Conference on Industrial and Information Systems (ICIIS)

Aug 2023

Abstract

This paper is aimed at evaluating state-of-the-art models for Multi-document Summarization (MDS) on different types of datasets in various domains and investigating the limitations of existing models to determine future research directions. To address this gap, we conducted an extensive literature review to identify state-of-the-art models and datasets. We analyzed the performance of PRIMERA and PEGASUS models on BigSurvey-MDS and MS^2 datasets, which posed unique challenges due to their varied domains. Our findings show that the General-Purpose Pretrained Model LED outperforms PRIMERA and PEGASUS on the MS^2 dataset. We used the ROUGE score as a performance metric to evaluate the identified models on different datasets. Our study provides valuable insights into the models' strengths and weaknesses, as well as their applicability in different domains. This work serves as a reference for future MDS research and contributes to the development of accurate and robust models which can be utilized on demanding datasets with academically and/or scientifically complex data as well as generalized, relatively simple datasets.

AI for Mitigating Effects of Climate and Weather Changes in Agriculture

Narmada Balasooriya, C.D. Athuraliya, Janak Gunatilleke

AI for Social Good Workshop, International Conference on Machine Learning (ICML)

June 2019

Abstract

In recent years, floods, landslides and droughts have become an annual occurrence in Sri Lanka. Despite the efforts made by the government and other entities, these natural disasters remain challenging mainly to the people who live in high risk areas. It is also crucial to predict such disasters early on to facilitate evacuation of people living in these areas. Furthermore, Sri Lankan economy largely depends on agriculture, yet this sector still remains untouched by recent advancements of AI and other predictive analytics techniques. There is an increased tendency amongst Sri Lankan youth to refrain from agriculture sector due to the lack of technology adoption and risks associated with it. The work by Peiris et.al demonstrates how seasonal climate data is used to predict coconut yield in Sri Lanka. Another Sri Lankan tech company has initiated the project AiGrow to increase the use of state-of-the-art technology in agriculture sector by utilizing AI, smart sensor systems and fertigation systems (automated fertilizer delivery machine). Regionally, Pulse Lab Jakarta has developed a visualization tool to track the impact of forest and peatland fires in Indonesia.

Visualizing the Consequences of Climate Change Using Cycle-Consistent Adversarial Networks

Victor Schmidt, Alexandra Luccioni, S. Karthik Mukkavilli, Kris Sankaran, Yoshua Bengio, Narmada Balasooriya, Jennifer Chayes

AI for Social Good Workshop, International Conference on Learning Representations (ICLR)

May 2019

Abstract

We present a project that aims to generate images that depict accurate, vivid, and personalized outcomes of climate change using Cycle-Consistent Adversarial Networks (CycleGANs). By training our CycleGAN model on street-view images of houses before and after extreme weather events (e.g. floods, forest fires, etc.), we learn a mapping that can then be applied to images of locations that have not yet experienced these events. This visual transformation is paired with climate model predictions to assess likelihood and type of climate-related events in the long term (50 years) in order to bring the future closer in the viewer’s mind. The eventual goal of our project is to enable individuals to make more informed choices about their climate future by creating a more visceral understanding of the effects of climate change, while maintaining scientific credibility by drawing on climate model projections.

Context-aware Capsule Network for Multi-label Classification

Sameera Ramasinghe, C.D. Athuraliya, Salman H. Khan

Workshop on Brain-Driven Computer Vision, European Conference on Computer Vision (ECCV)

Sep 2018

Abstract

Recently proposed Capsule Network is a brain inspired architecture that brings a new paradigm to deep learning by modelling input domain variations through vector based representations. Despite being a seminal contribution, CapsNet does not explicitly model structured relationships between the detected entities and among the capsule features for related inputs. Motivated by the working of cortical network in HVS, we seek to resolve CapsNet limitations by proposing several intuitive modifications to the CapsNet architecture. We introduce, (1) a novel routing weight initialization technique, (2) an improved CapsNet design that exploits semantic relationships between the primary capsule activations using a densely connected Conditional Random Field and (3) a Cholesky transformation based correlation module to learn a general priority scheme. Our proposed design allows CapsNet to scale better to more complex problems, such as the multi-label classification task, where semantically related categories co-exist with various interdependencies. We present theoretical bases for our extensions and demonstrate significant improvements on ADE20K scene dataset.