Search

Metaphors are compact packages of information with rich cultural background information. As one of the most powerful linguistic forms with non-literal meaning, metaphor detection in natural language processing can be both challenging and rewarding. We propose an innovative method for metaphor detection and classification leveraging culturally grounded eventive information. This culturally grounded information is organized based on ontological structure, which in turn facilitates further semantic processing of the result of our classification. As a culturally bound ontological system, the Chinese writing system has basic concepts organized according to semantic radicals, which are symbols containing rich eventive information that represent categorical concepts. This paper illustrates the basic design principles of applying ontological structures in metaphor detection by taking into account radicals representing body parts, instruments, materials, and movements. Our approach to leverage the eventive information of the Chinese writing system in metaphor detection is based on the fact that such information is available as an integral part of the writing system of any text. We hypothesize that eventive information can be accessed through the “embodied” source domain information represented by the radicals without syntactic processing or annotation. In terms of the theory of metaphor, we further hypothesize that eventive types in the embodied source domain maps to, and hence can help to predict, eventive meaning in the target domain of metaphor. Our studies show that the event information encoded in lexical items can facilitate classification of metaphoric events and identification of metaphors in Chinese texts effectively. We achieved improvements in Chinese metaphor detection over state-of-the-art approaches in our first classification experiment, and our proposed approach is shown to be generalizable in a second experiment involving new sets of characters with the same radicals.

Distributional Semantic Models (DSMs) established themselves as a standard for the representation of word and sentence meaning. However, DSMs provide quantitative measurement of how strongly two linguistic expressions are related, without being able to automatically classify different semantic relations. Hence the notion of semantic similarity is underspecified in DSMs. We introduce Evalution-MAN in this paper as an effort to address this underspecification problem. Following the EVALution 1.0 dataset for English, we present a dataset for evaluating DSMs on the task of the identification of semantic relations in Mandarin Chinese. Moreover, we test different types of word vectors on the automatic learning of these semantic relations, and we evaluate them both in a unsupervised and in a supervised setting, finding that distributional models tend, in general, to assign higher similarity scores to synonyms and that deep learning classifiers are the best performing ones in the identification of semantic relations.

Finding methods to represent multiple types of nodes in heterogeneous networks is both challenging and rewarding, as there is much less work in this area compared with that of homogeneous networks. In this paper, we propose a novel approach to learn node embedding for heterogeneous networks through a joint learning framework of both network links and text associated with nodes. A novel attention mechanism is also used to make good use of text extended through links to obtain much larger network context. Link embedding is first learned through a random-walk-based method to process multiple types of links. Text embedding is separately learned at both sentence level and document level to capture salient semantic information more comprehensively. Then, both types of embeddings are jointly fed into a hierarchical neural network model to learn node representation through mutual enhancement. The attention mechanism follows linked edges to obtain context of adjacent nodes to extend context for node representation. The evaluation on a link prediction task in a heterogeneous network data set shows that our method outperforms the current state-of-the-art method by 2.5%-5.0% in AUC values with p-value less than 10-9, indicating very significant improvement.