Key Techniques For Detection And Interpretation Of Chinese Sentence-level Metaphor

We make metaphors in our daily speech for their poetic power,aesthetic pleasure or ability to elucidate abstract concepts and experiences.Our cognitive mechanism and conceptual system are based on metaphor.Therefore,metaphor computing studies play a critical role in the area of natural language processing(NLP).Metaphor computing studies,which aim to explore and characterize the universal metaphorical mechanism in language,generally include the following two aspects:one is to detect the metaphorical expression in the corpus and distinguish it from the literal expression;the other is to describe the literal meaning of metaphor and achieve an effective interpretation of metaphorical meaning.Starting with the cognitive and linguistic features of metaphor,this paper proposes the computational models to detect and interpret the Chinese metaphor in the sentence level.In the task of metaphor detection,we aim to determine whether an utterance is metaphorical or not.A neural network based model has been active in detecting metaphor,but only few models introduce the attention mechanism in the detection task.Considering that the conceptual semantic features,such as a degree of abstractness,play important role when detecting metaphor,we argue that the abstractness degrees of words help machine to capture more critical information in the sentence for metaphor detection,so as to better determine the metaphorical tendency of an utterance.Therefore,we introduced a neural network model with an attention mechanism that considers the abstractness degree.Referring to the conceptual metaphor theory,we assume that if a word has a higher abstractness degree,it has a greater likelihood of being associated with the target domain.We selected the word with the highest degree of abstractness as the attention word of the corresponding sentence.This paper constructs a bidirectional long-short term memory network(BiLSTM)to capture the features of variable length sentences.After that,we calculated a position-weighted sentence features by considering the position information of each word.We also introduced a recurrent attention module to calculate the attention scores of each memory slices.Finally,our model performs binary(metaphorical/literal)classification on metaphors.In the task of metaphor interpretation,this paper presents a cooperative network model to interpret sentential metaphors based on the theory of interaction.We assume that the target and source domains obtain the process of seeking common ground while reserving differences through the mechanism of cooperation,dynamically construct the semantic association between them and output the literal meaning that adapted to the contextual information.In this paper,we extracted the properties from source domain and calculated their corresponding cooperative strengths with the contextual words as well as the properties of the target domain.The salient property with the highest cooperative strength is highlighted as the interpretation result of metaphor.To conclude,this paper proposes an abstractness attention-based network for metaphor detection,and presents a cooperative network model for metaphor interpretation.The experimental results demonstrate that these two models perform well.