LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection

Stance detection is a critical task in natural language processing that determines an author’s viewpoint toward a specific target, playing a pivotal role in social science research and various applications. Traditional approaches incorporating Wikipedia-sourced data into small language models (SLMs)...

Full description

Saved in:
Bibliographic Details
Main Authors: Woojin Lee, Jaewook Lee, Harksoo Kim
Format: Article
Language:English
Published: PeerJ Inc. 2024-12-01
Series:PeerJ Computer Science
Subjects:
Online Access:https://peerj.com/articles/cs-2585.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1850245680184950784
author Woojin Lee
Jaewook Lee
Harksoo Kim
author_facet Woojin Lee
Jaewook Lee
Harksoo Kim
author_sort Woojin Lee
collection DOAJ
description Stance detection is a critical task in natural language processing that determines an author’s viewpoint toward a specific target, playing a pivotal role in social science research and various applications. Traditional approaches incorporating Wikipedia-sourced data into small language models (SLMs) to compensate for limited target knowledge often suffer from inconsistencies in article quality and length due to the diverse pool of Wikipedia contributors. To address these limitations, we utilize large language models (LLMs) pretrained on expansive datasets to generate accurate and contextually relevant target knowledge. By providing concise, real-world insights tailored to the stance detection task, this approach surpasses the limitations of Wikipedia-based information. Despite their superior reasoning capabilities, LLMs are computationally intensive and challenging to deploy on smaller devices. To mitigate these drawbacks, we introduce a reasoning distillation methodology that transfers the reasoning capabilities of LLMs to more compact SLMs, enhancing their efficiency while maintaining robust performance. Our stance detection model, LOGIC (LLM-Originated Guidance for Internal Cognitive improvement of small language models in stance detection), is built on Bidirectional and Auto-Regressive Transformer (BART) and fine-tuned with auxiliary learning tasks, including reasoning distillation. By incorporating LLM-generated target knowledge into the inference process, LOGIC achieves state-of-the-art performance on the VAried Stance Topics (VAST) dataset, outperforming advanced models like GPT-3.5 Turbo and GPT-4 Turbo in stance detection tasks.
format Article
id doaj-art-c4c869dd3e8841dca394e8e49e3741d4
institution OA Journals
issn 2376-5992
language English
publishDate 2024-12-01
publisher PeerJ Inc.
record_format Article
series PeerJ Computer Science
spelling doaj-art-c4c869dd3e8841dca394e8e49e3741d42025-08-20T01:59:22ZengPeerJ Inc.PeerJ Computer Science2376-59922024-12-0110e258510.7717/peerj-cs.2585LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detectionWoojin Lee0Jaewook Lee1Harksoo Kim2Department of Artificial Intelligence, Konkuk University, Seoul, Republic of South KoreaDepartment of Artificial Intelligence, Konkuk University, Seoul, Republic of South KoreaDepartment of Computer Science and Engineering, Konkuk University, Seoul, Republic of South KoreaStance detection is a critical task in natural language processing that determines an author’s viewpoint toward a specific target, playing a pivotal role in social science research and various applications. Traditional approaches incorporating Wikipedia-sourced data into small language models (SLMs) to compensate for limited target knowledge often suffer from inconsistencies in article quality and length due to the diverse pool of Wikipedia contributors. To address these limitations, we utilize large language models (LLMs) pretrained on expansive datasets to generate accurate and contextually relevant target knowledge. By providing concise, real-world insights tailored to the stance detection task, this approach surpasses the limitations of Wikipedia-based information. Despite their superior reasoning capabilities, LLMs are computationally intensive and challenging to deploy on smaller devices. To mitigate these drawbacks, we introduce a reasoning distillation methodology that transfers the reasoning capabilities of LLMs to more compact SLMs, enhancing their efficiency while maintaining robust performance. Our stance detection model, LOGIC (LLM-Originated Guidance for Internal Cognitive improvement of small language models in stance detection), is built on Bidirectional and Auto-Regressive Transformer (BART) and fine-tuned with auxiliary learning tasks, including reasoning distillation. By incorporating LLM-generated target knowledge into the inference process, LOGIC achieves state-of-the-art performance on the VAried Stance Topics (VAST) dataset, outperforming advanced models like GPT-3.5 Turbo and GPT-4 Turbo in stance detection tasks.https://peerj.com/articles/cs-2585.pdfNatural language processingStance detectionReasoning distillation
spellingShingle Woojin Lee
Jaewook Lee
Harksoo Kim
LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
PeerJ Computer Science
Natural language processing
Stance detection
Reasoning distillation
title LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
title_full LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
title_fullStr LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
title_full_unstemmed LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
title_short LOGIC: LLM-originated guidance for internal cognitive improvement of small language models in stance detection
title_sort logic llm originated guidance for internal cognitive improvement of small language models in stance detection
topic Natural language processing
Stance detection
Reasoning distillation
url https://peerj.com/articles/cs-2585.pdf
work_keys_str_mv AT woojinlee logicllmoriginatedguidanceforinternalcognitiveimprovementofsmalllanguagemodelsinstancedetection
AT jaewooklee logicllmoriginatedguidanceforinternalcognitiveimprovementofsmalllanguagemodelsinstancedetection
AT harksookim logicllmoriginatedguidanceforinternalcognitiveimprovementofsmalllanguagemodelsinstancedetection