Task and Sentiment Adaptation for Appraisal Tagging

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 1960–1970 May 2-6, 2023 ©2023 Association for Computational Linguistics Task and Sentiment Adaptation for Appraisal Tagging Lin Tian RMIT University, Australia lin.tian2@student.rmit.edu.au Xiuzhen Zhang∗ RMIT University, Australia xiuzhen.zhang@rmit.edu.au Maria Myung-Hee Kim Defence Science and Technology Group Australia myung.kim@defence.gov.au Jennifer Biggs Defence Science and Technology Group Australia jennifer.biggs@defence.gov.au Abstract Sentiment analysis and opinion mining of the opinion-bearing text are important tasks in NLP. The Appraisal framework in systemic functional linguistics is a theory for analysing the linguistic patterns for expressing emotion and opinion. Manual annotation of appraisals however, requires linguistic expertise, and is costly and time-consuming. In this paper, we study how to automatically identify and tag appraisal text segments. We formulate the problem as a sequence tagging problem and propose a novel approach, Adaptive Ap- praisal (A2), which employs task and sentiment adapters on pre-trained language models for se- quence appraisal tagging. Experiments on user comments, blogs and microblogs show that A2 outperforms baseline models and achieves good performance for cross-domain and cross- lingual settings. Source code for A2 is available at: https://github.com/ltian678/AA-code.git 1 Introduction With the development of the Web technology, opinion-bearing user generated texts such as re- views, users comments, blogs and microblogs are widespread. Sentiment analysis and opinion min- ing on such texts are prominent NLP tasks that have attracted extensive research studies in the liter- ature (Liu, 2022). On the other hand, the Appraisal framework (Martin and White, 2003) is a systemic functional linguistic theory describing how lan- guage is used by writers or speakers to express emotion and opinion. The Appraisal framework consists of three subsystems: 1) Attitude, which includes personal emotion, judgement and evalu- ation of entities; 2) Engagement, which regards one’s own opinions or with respect to others; and 3) Graduation, which describes strength of the atti- tude and engagement expressed. Appraisal annotated resources have been used for deeper sentiment and emotion analy- ∗Corresponding author sis (Whitelaw et al., 2005) than simple sentiment classification, but building such resources manually requires significant linguistic expertise and is time- consuming (Read and Carroll, 2012; Kolhatkar et al., 2020). Automated appraisal tagging would be extremely beneficial to support annotation and analysis efforts by expert linguists. To our best knowledge, the only publicly available appraisal an- notated corpus is the Simon Fraser University Opin- ion and Comments Corpus (SOCC) (Kolhatkar et al., 2020)1, which is based on news comments. In creating the SOCC corpus, 663,173 user com- ments were collected, and expert linguists manually annotated 1,043 comments. To the best of our knowledge, there has been no prior work leveraging machine learning for au- tomatically tagging appraisals. We fill this gap in this paper. We especially target domains with voluminous opinionated texts but zero or very lim- ited appraisal annotation resources, such as blogs and microblogs. Our research focuses on sequence tagging for cross-domain and cross-lingual texts with low resources. In the literature, various ap- proaches for sequence tagging tasks have been …

Content truncated to 500 words