Common Space Embedding of Primal-Dual Relation Semantic Spaces

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1579–1590, Dublin, Ireland, August 23-29 2014. Common Space Embedding of Primal-Dual Relation Semantic Spaces Hidekazu Oiwa∗ The University of Tokyo Tokyo, Japan hidekazu.oiwa@gmail.com Jun’ichi Tsujii Microsoft Research Beijing, China jtsujii@microsoft.com Abstract Explicit continuous vector representation such as vector representation of words, phrases, etc. has been proven effective for various NLP tasks. This paper proposes a novel method of constructing such vector representation for both entity-pairs and relation expressions which link them in text. Based on the insight of the duality of relations, the representation is constructed by embedding of two separately constructed semantic spaces, one for entity-pairs and the other for relation expressions, into a common semantic space. By representing the two different types of objects (i.e. entity-pairs and relation expressions) in the same semantic space, we can treat the two tasks, relation mining and relation expression mining (a.k.a. pattern mining), systematically and in a unified manner. The approach is the first attempt to construct a continuous vector representation for expressions whose validity can be explicitly checked by their proximities to known sets of entity-pairs. We also experimentally validate the effectiveness of the common space for relation mining and relation expression mining. 1 Introduction Learning continuous vector representation for expressions which consist of more than one word has gained attention in recent years. Various representations have been constructed and used to measure semantic similarities between expressions in various tasks, such as analogical reasoning (Turney et al., 2003; Mikolov et al., 2013) and sentiment analysis (Turney and Littman, 2003; Socher et al., 2012). Many algorithms have been proposed to construct such continuous representations, depending on specific tasks in mind. In this paper, we propose a method for constructing a vector representation for binary relations, i.e., relations with two arguments. We demonstrate the effectiveness of the representation for relation mining and relation expression mining. The method exploits the duality of a relation (Bollegala et al., 2010). While Bollegala et al. (2010) uses the duality in their co-clustering algorithm, we construct an explicit semantic space which reflects the two aspects of a given relation. We first construct two separate semantic spaces, one for pairs of named entities and another for relation expressions in text which link an entity-pair. A relation is supposed to correspond to a subset in each of these two spaces. The subset of entity-pairs is a set of pairs between which the relation holds. The subset is called the extension set of the relation. The subset of relation expressions consists a set of expressions which are used to link entity-pairs in the extension set. The two semantic spaces are then embedded into a single common space. Figure 1 illustrates a brief summary of constructing a common semantic space. While the subsets which correspond to a specific relation are supposed to constitute natural clusters in the two original spaces, objects in the two spaces exchange useful information to each other and form a tighter cluster in the …

Content truncated to 500 words