Using Aspect-Based Sentiment Analysis to Classify Attitude-bearing Words

Using Aspect-Based Sentiment Analysis to Classify ATTITUDE-bearing Words Pradeesh Parameswaran, Andrew Trotman, Veronica Liesaputra, David Eyers Department of Computer Science University of Otago New Zealand [pradeesh,andrew,veronica,dme]@cs.otago.ac.nz Abstract APPRAISAL is widely used by linguists to study how people judge things or people. Au- tomating APPRAISAL could be beneficial for use cases such as moderating online com- ments. In 2020, the Australasian Language Technology Association (ALTA) organised a shared task to classify a branch APPRAISAL, which involves how humans judge other hu- mans (JUDGEMENT). It proved to be a dif- ficult task as the best performing system ob- tained an F1 = 0.155. In this work, we hypoth- esise that JUDGEMENT and APPRECIATION branches in APPRAISAL are similar to opinion in Aspect-Based Sentiment Analysis (ABSA) tasks, as such we can leverage on ABSA opin- ion extraction techniques to further improve the performance of automated approaches for identifying JUDGEMENT and APPRECIATION. We evaluated the performance of six different ABSA models on two publicly available AP- PRAISAL data sets (biographies and psycho- logical evaluation) by training them on exist- ing ABSA SemEval data sets. Our results show that there is an overlap between opinion- extraction and APPRAISAL task, as we ob- tained F1 = 0.623 on biographies data set and 0.414 on psychological evaluation data set. However, we cannot be certain if our findings can be extended across other APPRAISAL data sets due to the challenges in annotating and the availability of these data sets. 1 Introduction In 2020, ALTA organised a shared task challenge aimed to classify how humans judge other humans using a well-known linguistic taxonomy known as APPRAISAL (Martin and White, 2005) auto- matically. APPRAISAL allows linguists to evalu- ate language in a social context such as identify- ing and understanding how people make judge- ments about people and objects (ATTITUDE) (Mar- tin and White, 2005). The taxonomy is commonly used by Systemic Functional linguists to analyse the language choices and attitudes used by writ- ers and speakers (Chen, 2022) in various mediums (Starfield et al., 2015; Ross and Caldwell, 2020; Su and Hunston, 2019). Identifying ATTITUDE-bearing words can help to reduce the workload of moderators such as in online forums and Facebook by analysing the lan- guage that is being used and flagging it to moder- ators to be reviewed if there are any legal implica- tions based on the APPRAISAL taxonomy (Steiger et al., 2021). Although, there were two winners declared for the ALTA 2020 Shared Task challenge, the task proved to be difficult as the best-performing team only obtained an F1 score of 0.155 (Moll´a, 2020). The main reason for poor scores was the size of data set (N = 300): too small for automated meth- ods to generalise from properly. A lot of the larger APPRAISAL data set is not publicly released, thus making it difficult for automated approaches to be be built. However, there might be a solution for us to tackle this problem without the need of a large data set. Recently, Su and Hunston …

Content truncated to 500 words