AMEX AI Labs at SemEval-2022 Task 10: Contextualized fine-tuning of BERT for Structured Sentiment Analysis

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1296 - 1304 July 14-15, 2022 ©2022 Association for Computational Linguistics AMEX AI Labs at SemEval-2022 Task 10: Contextualized fine-tuning of BERT for Structured Sentiment Analysis Pratyush Sarangi, Shamika Ganesan, Piyush Arora, Salil Rajeev Joshi American Express AI Labs, Bangalore, India {pratyush.k.sarangi|shamika.ganesan|piyush.arora1|salilrajeev.joshi}@aexp.com Abstract We describe the work carried out by AMEX AI Labs on the structured sentiment analysis task at SemEval-2022. This task focuses on extract- ing fine grained information w.r.t. to source, target and polar expressions in a given text. We propose a BERT based encoder, which utilizes a novel concatenation mechanism for combin- ing syntactic and pretrained embeddings with BERT embeddings. Our system achieved an average rank of 14/32 systems, based on the average scores across seven datasets for five languages provided for the monolingual task. The proposed BERT based approaches outper- formed BiLSTM based approaches used for structured sentiment extraction problem. We provide an in-depth analysis based on our post submission analysis. 1 Introduction In this paper we present the work done by AMEX AI Labs on the structured sentiment analysis mono- lingual task at SemEval-2022 (Barnes et al., 2022). Structured sentiment analysis (SSA) focuses on per- forming fine grained analysis and extracting opin- ion tuples from a given input text (Barnes et al., 2021). An example is presented in Fig. 1. Input sentence: Some others give the new UMUC 5 stars - don’t believe them Expected outputs are two tuples: • Tuple1 - ("source": "Some others", "target": "the new UMUC", "polar expression": "5 stars") • Tuple2 - ("source": "", "target": "them", "po- lar expression": "don’t believe"). Aspect based sentiment analysis is a popular and widely researched topic in the NLP commu- nity (Wagner et al., 2014; Pontiki et al., 2015; Schouten and Frasincar, 2015). However, the SSA task goes beyond traditional aspect based sentiment Figure 1: Sample sentence, source: (Barnes et al., 2021) mining. The main challenges in SSA task are as follows: • Focus is on correct span detection of source, targets, and polar expression elements as shown in Fig. 1. • Correctly identifying relationships between source, target, and polar expression elements, as illustrated in the example above. • Building a robust, generalized approach that is applicable across domains as well as across languages. Most of the state-of-the-art (SOTA) approaches have explored joint learning, span based multi- task learning, BiLSTM encoders or CRF based approaches for SSA related tasks (Barnes et al., 2021; Li et al., 2019; Zhao et al., 2020; Chen et al., 2020). We propose a contextualized BERT based approach called ‘CBERT’, for capturing context information effectively and address the main chal- lenges for SSA task as discussed above. The main contributions of our work are: • We propose a BERT based encoding approach and combine syntactic and pre-trained embed- dings using a novel concatenation approach to extract and find opinion tuples effectively. • We performed a detailed error analysis high- lighting the main challenges of SSA task and suggest future directions …

Content truncated to 500 words