SPDB Innovation Lab at SemEval-2022 Task 10: A Novel End-to-End Structured Sentiment Analysis Model based on the ERNIE-M

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1401 - 1405 July 14-15, 2022 ©2022 Association for Computational Linguistics SPDB Innovation Lab at SemEval-2022 Task 10: A Novel End-to-End Structured Sentiment Analysis Model based on the ERNIE-M Yalong Jia, Zhenghui Ou, Yang Yang Shanghai Pudong Development Bank {jiayl2, ouzh, yangy103}@spdb.com.cn Abstract Sentiment analysis is a classical problem of nat- ural language processing. SemEval 2022 sets a problem on the structured sentiment analysis in task 10, which is also a study-worthy topic in research area. In this paper, we propose a method which can predict structured sentiment information on multiple languages with lim- ited data. The ERNIE-M pretrained language model is employed as a lingual feature extrac- tor which works well on multiple language pro- cessing, followed by a graph parser as a opinion extractor. The method can predict structured sentiment information with high interpretabil- ity. We apply data augmentation as the given datasets are so small. Furthermore, we use K-fold cross-validation and DeBERTaV3 pre- trained model as extra English embedding gen- erator to train multiple models as our ensemble strategies. Experimental results show that the proposed model has considerable performance on both monolingual and cross-lingual tasks. 1 Introduction Sentiment analysis (Liu, 2012) is widely used in many aspects of computer science nowadays, such as human computer interaction, lingual feature ex- traction, etc. Affective computing techniques en- able us to explore the sentiment message, which depicts the preference, emotion or even idea of people, behind the sentence itself. Sentiment analysis task can be classified into text level, sentence level, entity level and opinion tuple level. The goal of text or sentence level senti- ment analysis is to predict sentiments of given doc- uments or sentences. On the other hand, entity level sentiment analysis needs to consider sentiments be- tween each entity in given sentence. Furthermore, a sentence may contain multiple entities with dif- ferent sentiments (positive or negative). Therefore, in comparison to entity level sentiment analysis, opinion tuple level sentiment analysis requires ad- ditional extraction of relations between entities and opinions. Structured sentiment analysis (Barnes et al., 2022) is a task to predict a sentiment graph for given sentences. It can be theoretically cast as an information extraction problem in which one attempts to find all of the opinion tuples O = O1, ..., On in a text. As we can see in Figure 1, each opinion Oi is a tuple (h, t, e, p) where h is a holder who expresses a polarity p towards a target t through a sentiment expression e, implicitly defin- ing pairwise relationships between elements of the same tuple. The structure of the paper is as follows. Sec- tion 2 briefly reviews recent works on similar tasks; Section 3 describes our model structure in detail. Section 4 shows the analysis of the given datasets; Section 5 introduces our experimental setting and results. And finally in Section 6, we make a con- clusion and give the ideas about future works. 2 Background Dividing …

Content truncated to 500 words