Efficient Hybrid Generation Framework for Aspect-Based Sentiment Analysis

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1007–1018 May 2-6, 2023 ©2023 Association for Computational Linguistics Efficient Hybrid Generation Framework for Aspect-Based Sentiment Analysis Haoran Lv1, Junyi Liu1, Henan Wang1, Yaoming Wang2, Jixiang Luo2, Yaxiao Liu1 1Amazon Web Services, China 2Shanghai Jiao Tong University, China {lvhaoran, liujunyi, henanwan, liuyaxia}@amazon.com {wang_yaoming, ljx123456}@sjtu.edu.cn Abstract Aspect-based sentiment analysis (ABSA) has attracted broad attention due to its commer- cial value. Natural Language Generation-based (NLG) approaches dominate the recent ad- vance in ABSA tasks. However, current NLG practices are inefficient because most of them directly employ an autoregressive generation framework that cannot efficiently generate loca- tion information and semantic representations of ABSA targets. In this paper, we propose a novel framework, namely Efficient Hybrid Generation (EHG) to revolutionize traditions. Specifically, we leverage an Efficient Hybrid Transformer to generate the location and se- mantic information of ABSA targets in parallel. Besides, we design a novel global hybrid loss function in combination with bipartite match- ing to achieve end-to-end model training. Ex- tensive experiments demonstrate that our pro- posed EHG framework greatly improves the efficiency of NLG-based methods and outper- forms the competitive baselines in almost all cases. 1 Introduction Aspect-based sentiment analysis (ABSA) enjoys broad commercial value, e.g., analyzing customer’s opinions through review data to improve products. ABSA consists of four basic sentiment elements, including aspect term, opinion term, aspect cate- gory and sentiment polarity. Illustrated in Fig. 1, the aspect terms a1, a2 in the sentence belong to the aspect categories c1, c2 , respectively, and their corresponding opinion terms are o1, o2 , with senti- ment polarity s1, s2. For simplicity, we name the aspect term and opinion term as sentiment entities (SE), and the aspect category and sentiment polar- ity as sentiment abstractions (SA). The majority of ABSA subtasks utilize sentences as input and the combination of SE and SA as output, so we define the output set of all ABSA subtasks as sen- timent tuples (ST). Fig. 1 shows all the subtasks Sentence: The food is delicious, but the service is not good 𝒂𝟏 ∈ 𝒄𝟏 𝒂𝟐 ∈ 𝒄𝟐 𝒐𝟐 𝒐𝟏 𝒔𝟏 𝒔𝟐 Aspect Term Extraction(AE): Extracting all the aspect terms Opinion Term Extraction(OE): Extracting all the opinion terms Aspect Term Extraction and Sentiment Classification(AESC): Extracting all aspect terms and their corresponding sentiment polarity Aspect Opinion Pair Extraction (AOPE): Extracting all aspect terms and their corresponding opinion terms Aspect Sentiment Triplet Extraction (ASTE): Extracting all aspect terms and their corresponding opinion terms and sentiment polarity Target Aspect Sentiment Detection (TASD): Extracting all aspect terms and their corresponding aspect categories and sentiment polarity Sentence Sentence Sentence Sentence Sentence Sentence 𝒂𝟏, 𝒂𝟐 𝒐𝟏, 𝒐𝟐 𝒂𝟏 − 𝒔𝟏, 𝒂𝟐 − 𝒔𝟐 𝒂𝟏 − 𝒐𝟏, 𝒂𝟐 − 𝒐𝟐 𝒂𝟏 − 𝒐𝟏 − 𝒔𝟏, 𝒂𝟐 − 𝒐𝟐 − 𝒔𝟐 𝒂𝟏 − 𝒄𝟏 − 𝒔𝟏, 𝒂𝟐 − 𝒄𝟐 − 𝒔𝟐 Subtask Output Input Figure 1: The six ABSA subtasks studied in this work, where a1=food, a2=service, c1=food quality, c2=service …

Content truncated to 500 words