Comprehensive analysis of aspect term extraction methods using various text embeddings. (September 2021)

• Record Type: Journal Article
• Authors: Augustyniak, Łukasz; Kajdanowicz, Tomasz; Kazienko, Przemysław
• Is Part Of: Computer speech & language. Volume 69(2021)
• Publication Date: 2021-09
• Subjects: Aspect-based sentiment analysis -- Aspect term extraction -- Word embeddings -- Character embeddings -- LSTM -- BiLSTM -- CRF -- SemEval; Speech processing systems -- Periodicals; Automatic speech recognition -- Periodicals; Computers -- Periodicals; Linguistics -- Periodicals; Speech-Language Pathology -- Periodicals; Traitement automatique de la parole -- Périodiques; Reconnaissance automatique de la parole -- Périodiques; Automatic speech recognition; Speech processing systems; Electronic journals; Periodicals; 006.454
• DOI: 10.1016/j.csl.2021.101217
• Languages: English
• ISSNs: 0885-2308

Abstract

Highlights: A wide comparison with ablation analysis of Aspect Term Extraction methods. A guideline to choose the best Aspect Term Extraction models for particular user needs. We analyzed 88 method combinations, model customizations (LSTM vs. BiLSTM, CRF vs. no CRF) and pre-trained word embeddings. We also evaluated the results of three contextual text representations (BERT, Flair, ELMo) using the BiLSTM-CRF model. We analyzed the influence on the performance of extending the word vectorization step with character-based word embeddings. The experimental results on SemEval datasets revealed that the BiLSTM could be used as a very good predictor. Language model-based model are not always the best and obvious choice for vector representation layer in NLP tasks. Abstract: Recently, a variety of model designs and methods have blossomed in the context of the sentiment analysis domain. However, there is still a lack of comprehensive studies of Aspect-based Sentiment Analysis. We want to fill this gap and propose a comparison with ablation analysis of Aspect Term Extraction using various text embeddings methods. We particularly focused on simple architectures based on long short-term memory (LSTM) with optional conditional random field (CRF) enhancement using different pre-trained word embeddings. Moreover, we analyzed the influence on the performance of extending the word vectorization step with character-based word embeddings. The experimental results on SemEval datasets revealed that bi-directional long short-term memory (BiLSTM) could be used as a very good predictor, even comparing to very sophisticated and complex models using huge word embeddings or language models. We presented a comprehensive analysis of various customizations of LSTM-based architecture and word/character embeddings that could be used as a guideline to choose the best model version for particular user needs.