Aspect-Based Sentiment Analysis (ABSA) has seen significant advancements with the introduction of Transformer-based models, which have reshaped the landscape of Natural Language Processing (NLP) tasks. This paper introduces enhancements to the Instruct-DeBERTa model which is one of the leading ABSA models for ABSA. It takes a hybrid approach combining the strengths of InstructABSA for Aspect Term Extraction (ATE) and DeBERTa-V3-baseabsa-V1 for Aspect Sentiment classification (ASC). In this work, we enhance Instruct-DeBERTa by introducing category classification through a cosine similarity-based method, comparing aspect embeddings with predefined categories. Also for InstructABSA and DeBERTa-V3-baseabsa-V1, we investigate different configurations by adding a linear layer followed by ReLU activation, incorporation of regularization and optimization of attention heads. These modifications were tailored specifically for the data sets in the hospitality domain. Our empirical evaluations, run on diverse datasets, have shown that these enhancements significantly raise the performance of Instruct-DeBERTa for hospitality domain datasets.
Keywords: Natural Language Processing | Aspect-based Sentiment Analysis | Aspect Extraction | Information Extraction |
@inproceedings{jayakody2024enhanced,
title={{Enhanced Aspect-Based Sentiment Analysis with Integrated Category Extraction for Instruct-DeBERTa}},
author={Jayakody, Dineth and Isuranda, Koshila and Malkith, A V A and de Silva, Nisansa and Ponnamperuma, Sachintha Rajith and Sandamali, G G N and Sudheera, K L K and Sarathchandra, Kashnika Gimhani},
booktitle={Proceedings of the 38th Pacific Asia Conference on Language, Information and Computation},
pages={665--674},
year={2024},
abstract={Aspect-Based Sentiment Analysis (ABSA) has seen significant advancements with the introduction of Transformer-based models, which have reshaped the landscape of Natural Language Processing (NLP) tasks. This paper introduces enhancements to the Instruct-DeBERTa model which is one of the leading ABSA models for ABSA. It takes a hybrid approach combining the strengths of InstructABSA for Aspect Term Extraction (ATE) and DeBERTa-V3-baseabsa-V1 for Aspect Sentiment classification (ASC). In this work, we enhance Instruct-DeBERTa by introducing category classification through a cosine similarity-based method, comparing aspect embeddings with predefined categories. Also for InstructABSA and DeBERTa-V3-baseabsa-V1, we investigate different configurations by adding a linear layer followed by ReLU activation, incorporation of regularization and optimization of attention heads. These modifications were tailored specifically for the data sets in the hospitality domain. Our empirical evaluations, run on diverse datasets, have shown that these enhancements significantly raise the performance of Instruct-DeBERTa for hospitality domain datasets.},
misc={https://aclanthology.org/2024.paclic-1.63/,NLP,https://goo.gl/iY6aTr}
}
ActivePI: Nisansa de Silva
We propose to advance aspect-based sentiment analysis (ABSA) by developing enhanced models, datasets, and evaluation methods that improve the accuracy and applicability of sentiment classification across multiple domains.
ActivePI: Nisansa de Silva
We propose to advance natural language processing for low-resource settings by leveraging data augmentation, transfer learning, and model adaptation techniques to overcome the scarcity of annotated data and linguistic resources.