Metaphor Identification with Paragraph and Word Vectorization: An Attention-Based Neural Approach

Metaphor Identification with Paragraph and Word Vectorization: An Attention-Based Neural Approach Timour Igamberdiev Department of Linguistics College of Humanities Seoul National University tigamberdiev@snu.ac.kr Hyopil Shin Department of Linguistics College of Humanities Seoul National University hpshin@snu.ac.kr Abstract The current study investigates approaches to automatic metaphor identification, the compu- tational task of identifying whether a word or phrase in a portion of text is an instance of metaphor. In addition to using the Skip-Gram and Continuous Bag-of-Words algorithms for word-level feature extraction, the Paragraph Vector is utilized for obtaining sentence-level distributional information, being an exten- sion to these two algorithms for blocks of text larger than the word level. With fea- tures extracted using the above models, the performance of several different neural net- work systems are compared against a baseline of logistic regression on the VU Amsterdam Metaphor Corpus, with results showing a sig- nificant improvement and high success rates across the different models. This can be seen as strong evidence for the necessity of using state-of-the-art neural network architectures in supervised metaphor identification, being able to pick up on the various latent patterns pro- vided by the vector space model. 1 Introduction 1.1 Background and Overview The notion of metaphor is very important in lan- guage, providing us a glimpse into the cognitive as- pects of human linguistic knowledge. Being a fun- damental element of our cognition, it allows us to view one particular concept, often more abstract, in terms of another more basic concept, making the former much more accessible to our understanding (Lakoff and Johnson, 1980). Metaphor is ubiqui- tous in arguably any language, appearing in more subtle forms as in the case of conventional and lexicalized metaphors, as well as very clear cases including novel metaphors, especially within po- etry. For example, the following four metaphors, taken from K¨ovecses (2002), exemplify four dif- ferent stages of the metaphor SOCIAL ORGANIZA- TIONS ARE PLANTS: 1. “They had to prune the workforce.” 2. “Employers reaped enormous benefits from cheap foreign labor.” 3. “He works for the local branch of the bank.” 4. “There is a flourishing black market in software there.” As mentioned by Croft and Cruse (2004), each of the above metaphors are at different stages in their life-cycle, with the first (1) being the most obvious instance of metaphor, giving a strong sense of the notion of the plant source domain, while the last (4) being nearly unnoticeable to a native speaker, with the word flourish having come into English around the year 1300 from the French verb florir, having meant ‘to blossom’ or ‘to bloom’. The distinction between these different types of metaphors is very important for automatic metaphor detection, the computational task of identifying whether a word or phrase in a piece of text is a metaphor or not. Since many recent approaches to computational metaphor processing have used context-based features as input to the classification task (Shutova et al., 2012; Jang et al., 2015; Jang et al., 2016), the degree of success will largely …

Content truncated to 500 words