An Approach To Chinese Metaphor Identification Based On Word Abstractness

As a common phenomenon in human language, metaphor has become a tendencyin the area of natural language processing. Metaphor identification is the basis for themetaphor understanding, so how to use related theories about metaphor to recognizemetaphor is the prime task of the metaphor computation. At present, the most of thestudies on Chinese metaphor identification are concentrated in using the crossdomains of vocabulary as feature to recognize the metaphor like “A is B”. In thispaper, with the fundamental theory--the cognitive metaphor theory “the metaphor isoften used to represent abstract ideas with concrete things”, we proposed analgorithm to automatically calculate the word abstractness, then put it to recognize themetaphor sentences like “A is B”.Firstly, two algorithms for word abstractness computation are proposed. One ofis on the basis of the lexicon with abstractness information. We first get the fortyparadigm words by computing the Pearson correlation coefficient with the MRCPsycholinguistic Database. Second, we translate the Chinese words to English wordsor phrases. Third, we get the word abstractness by computing the similarity betweenEnglish words and the forty paradigm words. The other is an approach based onlogistic regression. At the first step, the words vectors are created through neuralnetwork language model. Then, logistic regression algorithm is introduced to computethe word abstractness degree. After introducing the algorithms, we evaluate of thesetwo methods by questionnaire and get the result that the latter algorithm is better thanthe former.On the premise of successfully calculating the word abstractness, we design ametaphor identification method based on it. At firstly, we get the words which may bethe source domain and target domain from the sentence by the condition randommodel. The second step is computing the word abstract on these words. Finally, withputting the word abstract to the features, we use non-linear support vector machinemodel to identify metaphor sentences. Meanwhile, to the contradistinction, we alsouse the semantic categories and the similarity to identify the metaphors. Through theexperiments, the method based on the word abstract computed by logistic regressionis better than others. To sum up, the Chinese words abstractness by automatically calculated iscorresponding to the abstract degree of the word. And the result about metaphoridentification method based on the word abstractness is better in our corpus. It canprovide a fundamental method for further research on Chinese metaphoridentification.