A Computational Study of Speech Acts in Social Media

Speech acts are expressed by humans in daily communication that perform an action (e.g. requesting, suggesting, promising, apologizing). Modeling speech acts is important for improving natural language understanding (i.e. human-computer interaction through computers’ comprehension of human language) and developing other natural language processing (NLP) tasks such as question answering and machine translation. Analyzing speech acts on large scale using computational methods could benefit linguists and social scientists in getting insights into human language and behavior.

Speech acts such as suggesting, questioning and irony have aroused great attention in previous NLP research. However, two common speech acts, complaining and bragging, have remained under explored. Complaints are used to express a mismatch between reality and expectations towards an entity or event. Previous research has only focused on binary complaint identification (i.e. whether a social media post contains a complaint or not) using traditional machine learning models with feature engineering. Bragging is one of the most common ways of self-presentation, which aims to create a favorable image by disclosing positive statements about speakers or their in-group. Previous studies on bragging have been limited to manual analyses of small data sets, e.g. fewer than 300 posts.

The main aim of this thesis is to enrich the study of speech acts in computational linguistics. First, we introduce the task of classifying complaint severity levels and propose a method for injecting external linguistic information into novel pretrained neural language models (e.g. BERT). We show that incorporating linguistic features is beneficial to complaint severity classification. We also improve the performance of binary complaint prediction with the help of complaint severity information in multi-task learning settings (i.e. jointly model these two tasks). Second, we introduce the task of identifying bragging and classifying their types as well as a new annotated data set. We analyze linguistic patterns of bragging and their types and present error analysis to identify model limitations. Finally, we examine the relationship between online bragging and a range of common socio-demographic factors including gender, age, education, income and popularity.