Improving Aspect-Based Sentiment with End-to-End Semantic Role Labeling Model

Proceedings of Recent Advances in Natural Language Processing, pages 888–897 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_096 888 Improving Aspect-Based Sentiment with End-to-End Semantic Role Labeling Model Pavel Pˇrib´aˇn* and Ondˇrej Praˇz´ak* University of West Bohemia, Faculty of Applied Sciences, Czech Republic 1Department of Computer Science and Engineering, 1NTIS – New Technologies for the Information Society, {pribanp,ondfa}@kiv.zcu.cz http://nlp.kiv.zcu.cz Abstract This paper presents a series of approaches aimed at enhancing the performance of Aspect- Based Sentiment Analysis (ABSA) by utiliz- ing extracted semantic information from a Se- mantic Role Labeling (SRL) model. We pro- pose a novel end-to-end Semantic Role La- beling model that effectively captures most of the structured semantic information within the Transformer hidden state. We believe that this end-to-end model is well-suited for our newly proposed models that incorporate semantic in- formation. We evaluate the proposed models in two languages, English and Czech, employ- ing ELECTRA-small models. Our combined models improve ABSA performance in both languages. Moreover, we achieved new state- of-the-art results on the Czech ABSA. 1 Introduction In recent years, the pre-trained BERT-like models based on the Transformer (Vaswani et al., 2017) architecture demonstrated their performance supe- riority across various natural language processing (NLP) tasks. In this paper, we study the possibility of a combination of two seemingly unrelated NLP tasks: Aspect-Based Sentiment Analysis (ABSA) and Semantic Role Labeling (SRL). We believe that the structured semantic information of a sen- tence extracted from an SRL model can enhance the performance of an ABSA model. We investi- gate our assumption on the ELECTRA (Clark et al., 2020) model architecture since it is a lighter and smaller alternative to the popular and commonly used models such as BERT (Devlin et al., 2019) or RoBERTa (Liu et al., 2019). Because the ELEC- TRA model is smaller in terms of the number of parameters, it does require less GPU memory and time to be fine-tuned. Sentiment analysis (SA) is an essential part of NLP. The most prevalent SA task is the Sentiment Classification, where the objective is to classify a text fragment (e.g., sentence or review) as positive or negative, eventually as neutral. In this type of task, we assume that there is only one opinion in the text. In reality, as illustrated in Figure 1, this assumption often does not hold true (Liu, 2012). “The burger was excellent but the waitress was unpleasant” CE ⇒ food, service CP ⇒ food:positive, service:negative Figure 1: Example of CE and CP subtasks of ABSA. Aspect-Based Sentiment Analysis (Liu, 2012; Pontiki et al., 2014) focuses on detecting aspects (e.g., food or service in the restaurant reviews do- main) and determining their polarity, enabling more detailed analysis and understating of the expressed sentiment. As shown by Pontiki et al. (2014), the ABSA task can be further divided into four sub- tasks: Aspect term extraction (TE), Aspect term polarity (TP), Aspect category extraction (CE), and Aspect category polarity (CP). We aim at the CE and CP subtasks,1 and we treat them as a single classification task, see …

Content truncated to 500 words