Improving Neural Metaphor Detection with Visual Datasets

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5928–5933 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5928 Improving Neural Metaphor Detection with Visual Datasets Gitit Kehat and James Pustejovsky Brandeis University Waltham, MA, USA {gititkeh, jamesp}@brandeis.edu Abstract We present new results on Metaphor Detection by using text from visual datasets. Using a straightforward technique for sampling text from Vision-Language datasets, we create a data structure we term a visibility word embedding. We then combine these embeddings in a relatively simple BiLSTM module augmented with contextualized word representations (ELMo), and show improvement over previous state-of-the-art approaches that use more complex neural network architectures and richer linguistic features, for the task of verb classification. Keywords: Metaphor Detection, Visual datasets, Visibility word embeddings, neural models 1. Introduction Metaphors play a special role in human language and thought, as they evoke a complex array of hidden connota- tions, past experiences, feelings, and humor, in the service of helping the speaker convey their message in a way that is easier to relate to. However, by their very nature, metaphors continue to pose a challenge to Natural Language Process- ing (NLP) systems, and their identification is crucial for many tasks, such as Machine Translation, Information Re- trieval, and others. In most cases, metaphor identification is done at the sen- tence level, where the input consists of some or all of the words in the sentence, and the output refers to the metaphoricity of the word(s) in the specific context. Often, Metaphor Identification takes the form of one of two tasks: (1) Sequence Labeling, in which each token in the sentence is classified as either “metaphorical” or “literal” (multiple outputs per sentence), or (2) Classification of a specific tar- get word, usually the main verb (one output per sentence). In this paper, we deal with the second task, which more formally takes a sentence w, ...wn and a verb index i as input, and outputs a label for the target verb wi of either “metaphorical” or “literal”, in relation to its role in the sen- tence (See Figure 1 for examples for non-metaphorical (lit- eral) and metaphorical usages of the same verb in different contexts). In our approach to improve metaphor detection, we follow Black (1979)’s observation that a metaphor is essentially an interaction between two terms, creating an “implication- complex” to resolve two incompatible meanings. Opera- tionally, we follow Turney et al. (2011) and their adoption of Lakoff and Johnson (1980)’s notion that metaphor is a way to move knowledge from a concrete domain to an ab- stract one. Hence, there should be a correlation between the “degree of abstractness in a word’s context [...] with the likelihood that the word is used metaphorically” (Tur- ney et al., 2011). Recent studies have suggested that there is a strong correlation between the concreteness scores of words, as annotated by humans, and the visibility of words, as calculated as a function of their occurrences in a visual …

Content truncated to 500 words