Computational Modeling of Metaphor in Discourse

 Metaphor is used as a language resource/tool to better represent one’s point in communication. It can help achieving social goals such as illustrating attitudes indirectly. This thesis aims to understand metaphor from this social perspective in order to capture how metaphor is used in a discourse and identify a broad spectrum of predictors from the discourse context that contribute towards its detection. We build computational models for metaphor detection that adopt the notion of framing in discourse, a well-known approach for conceptualizing discourse processes. I claim that developing computational models based on this view paves the way for metaphor processing at the discourse level such as extended metaphor detection, and ultimately contribute to modeling people’s use of metaphor in interaction. 

In order to model metaphor from this social perspective, we begin with corpus studies to observe people’s use of metaphor in three distinct domains where people use different metaphors for different purposes. This foundational work reveals how the layperson conception of metaphor differs from the technical operationalization of linguists from past work. The focus of our subsequent work is on metaphorical language that is recognizable as such by laypersons. 

Next, we perform two case studies, which illuminate the value of metaphor detection in discourse, to explore situational factors that affect people’s use of metaphor. The first study investigates inner situational factors. We build logistic regression models to discover whether metaphor usage is influenced by three psychological distress conditions including PTSD, depression, and anxiety. Our annotation scheme allows separating effects on language choices of the three factors: contextual expectations, content of the message, and framing. Separating these factors gives us deeper insight into understanding people’s metaphor choice, and necessitates consideration of these factors in our next studies. The second study examines external situational factors. We investigate the influence of stressful cancer events on people’s use of metaphor. This study verifies the association between the cancer events and metaphor usage, and the effectiveness of the situational factor as a new type of predictor for metaphor detection. 

 Then, we build computational models for detecting metaphors that can be around related metaphors, not restricted in their syntactic positions. These models find topical patterns by leveraging lexical context, to explore how a metaphorical frame switch is distinguished from a literal one. We design, implement, and evaluate computational models of three kinds: (1) features of frame contrast, which capture lexical contrast around metaphorical frames; (2) features of frame transition, which capture topic transition patterns occurring around metaphorical frames; and (3) features of frame facets, which capture frame facet patterns occurring around metaphorical frames. We demonstrate that these three features in a nonlinear machine learning model are effective in metaphor detection, and discuss the mechanism through which the frame information enables more accurate metaphor detection in discourse.