Диссертации, представленные на защиту и подготовленные в НИУ ВШЭ

Metaphor is recognized to be one of the most powerful cognitive tools with which humans conceptualize (Lakoff & Johnson, 1980). Hense, metaphor forms a fundamental part of the language system, and is truly ubiquitous in discourse. As a form of indirect expression, metaphor identification and interpretation pose a substantial technical challenge to a wide range of natural language processing (NLP) applications. This thesis presents an experiment in computational identification of linguistic metaphor in Russian texts. The experiment is designed in terms of binary classification, the unit of classification is a sentence (metaphoric if contains a metaphorically used target verb, non-metaphoric if otherwise); the sentences represent unrestricted non-curated contexts drawn from a large corpus. The central goal of the thesis is to provide data-driven linguistic analysis of context features which can be utilized in order to automatically differentiate utterances which contain linguistic metaphor from non-metaphoric ones. We start by collecting and annotating a corpus of contexts containing linguistic metaphor, as well as non-metaphoric ones, and evaluate the quality of metaphor annotation. Then, we suggest methods of feature engineering for identification of linguistic metaphor at the sentence level; the following types of features are implemented: distributional semantic similarity; lexical co-occurrence; morphosyntactic co-occurrence; concreteness indexes; and occurrences of flag words (lexical signals of metaphoricity) and quotation marks. Following the machine learning experiment, we evaluate the performance of the models and their generalizability, and conduct an in-depth linguistic analysis of the contextual factors that promote the success or failure of features.