The Automatic Identification Modelling Of Body Metaphors In English

Research on the automatic identification of metaphors started in the 1980s. In the beginning, rule-based method was the main research method. In recent years, as more sophisticated machine learning algorithms are employed, the accuracy of metaphor identification has improved. However, most of these studies are restricted in the disciplines of computational linguistics, focusing on increasing the optimization of algorithms. Meanwhile, linguistic patterns existent in linguistic metaphor expressions have not been fully explored. This study attempts to explore the syntagmatic and paradigmatic linguistic features of metaphoric expressions, so as to construct a more linguistically-informed model of automatic metaphor identification. To keep the study on a more manageable scale, this study focuses on the automatic identification of English body-part metaphors, with body-part metaphors and metonyms included.In this study, body-part metaphors are categorized into distinctive classes; mainwhile, linguistic features of body-part metaphors discussed in previous studies are fully employed to construct a more linguistically-informed metaphor identification model. This research addresses two questions:(1) How are English body-part metaphors distributed across different metaphorical meanings and linguistic patterns? (2) How could an automatic metaphor identification model be constructed on the basis of the linguistic features and how is the performance of the model?The research process reported in this study can be roughly divided into three stages: 1) Data collection and manual annotation.49 body-part lexical items, representing the body domain, were chosen from WordNet. Then, these body-part words were searched for in the British National Corpus, and 3,000 snetences were randomly selected from the query results. These sentences were then manually annotated with the type of metaphorical meaning.2) Lingsuitic analysis of the annotated sentences. This involves finding the common semantic categories and linguistic features of the metaphors in the annotated dataset.3) Constructing an automatic metaphor identification model and validating the performance of the model. The construction of the model involved:a) applying a knowledge base built on the basis of the linguistic analysis done in the previous stage; and b) applying machine learning algorithms. The validation of the model was conducted on a validatation dataset, and both precision and recall rates are reported.The major findings of this study are as follows:(1) Conventional metaphor is the most frequent body-part metaphor, characterized by fixed and semi-fixed expressions.(2) A model built on machine learning algorithm alone performs badly, with low recall rate. When supplemented with a knowledge base, however, the performance of the model can be greatly improved.(3) The statistically based model, when enhanced with a knowledge base, performs satisfactorily, with a precision rate of 0.984, a recall rate of 0.755, and an F value of 0.854. Compared with models reported in previous studies, this model is superior in terms of precision and the stability of recall rate.