Disentangling Aspect and Stance via a Siamese Autoencoder for Aspect Clustering of Vaccination Opinions DOI Creative Commons
Li Zhu, Runcong Zhao, Gabriele Pergola

et al.

Findings of the Association for Computational Linguistics: ACL 2022, Journal Year: 2023, Volume and Issue: unknown, P. 1827 - 1842

Published: Jan. 1, 2023

Mining public opinions about vaccines from social media has been increasingly relevant to analyse trends in debates and provide quick insights policy-makers. However, the application of existing models hindered by wide variety users’ attitudes new aspects continuously arising debate. Existing approaches, frequently framed via well-known tasks, such as aspect classification or text span detection, make direct usage supervision information constraining predefined classes, while still not distinguishing those stances. As a result, this significantly dynamic integration aspects. We thus propose model, namely Disentangled Opinion Clustering (DOC), for vaccination opinion mining media. DOC is able disentangle stances disentangling attention mechanism Swapping-Autoencoder, designed process unseen categories clustering approach, leveraging clustering-friendly representations induced out-of-the-box Sentence-BERT encodings mechanisms. conduct thorough experimental assessment demonstrating benefit mechanisms cluster-based approach on both quality clusters generalization across categories, outperforming methodologies aspect-based mining.

Language: Английский

Enhancing Health Information Retrieval with RAG by prioritizing topical relevance and factual accuracy DOI Creative Commons
Rishabh Upadhyay, Marco Viviani

Deleted Journal, Journal Year: 2025, Volume and Issue: 28(1)

Published: April 1, 2025

Abstract The exponential surge in online health information, coupled with its increasing use by non-experts, highlights the pressing need for advanced Health Information Retrieval (HIR) models that consider not only topical relevance but also factual accuracy of retrieved given potential risks associated misinformation. To this aim, paper introduces a solution driven Retrieval-Augmented Generation (RAG), which leverages capabilities generative Large Language Models (LLMs) to enhance retrieval health-related documents grounded scientific evidence. In particular, we propose three-stage model: first stage, user’s query is employed retrieve topically relevant passages references from knowledge base constituted literature. second these passages, alongside initial query, are processed LLMs generate contextually rich text (GenText). last be evaluated and ranked both point view means their comparison GenText, either through stance detection or semantic similarity. addition calculating accuracy, GenText can offer layer explainability it, aiding users understanding reasoning behind retrieval. Experimental evaluation our model on benchmark datasets against baseline demonstrates effectiveness enhancing factually accurate thus presenting significant step forward misinformation mitigation problem.

Language: Английский

Citations

1
Disentangling Aspect and Stance via a Siamese Autoencoder for Aspect Clustering of Vaccination Opinions DOI Creative Commons
Li Zhu, Runcong Zhao, Gabriele Pergola

et al.

Findings of the Association for Computational Linguistics: ACL 2022, Journal Year: 2023, Volume and Issue: unknown, P. 1827 - 1842

Published: Jan. 1, 2023

Mining public opinions about vaccines from social media has been increasingly relevant to analyse trends in debates and provide quick insights policy-makers. However, the application of existing models hindered by wide variety users’ attitudes new aspects continuously arising debate. Existing approaches, frequently framed via well-known tasks, such as aspect classification or text span detection, make direct usage supervision information constraining predefined classes, while still not distinguishing those stances. As a result, this significantly dynamic integration aspects. We thus propose model, namely Disentangled Opinion Clustering (DOC), for vaccination opinion mining media. DOC is able disentangle stances disentangling attention mechanism Swapping-Autoencoder, designed process unseen categories clustering approach, leveraging clustering-friendly representations induced out-of-the-box Sentence-BERT encodings mechanisms. conduct thorough experimental assessment demonstrating benefit mechanisms cluster-based approach on both quality clusters generalization across categories, outperforming methodologies aspect-based mining.

Language: Английский

Citations

1