Name List Only? Target Entity Disambiguation in Short Texts

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 654–664, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Name List Only? Target Entity Disambiguation in Short Texts Yixin Cao1, Juanzi Li1, Xiaofei Guo1, Shuanhu Bai2, Heng Ji3, Jie Tang1 1 Tsinghua National Laboratory for Information Science and Technology Dept. of Computer Science and Technology, Tsinghua University, China 100084 2 Sina Corporation, China 100084 3 Dept. of Computer Science, Rensselaer Polytechnic Institute, USA 12180 {caoyixin2011,lijuanzi2008,sophiaguo.thu,jery.tang}@gmail.com shuanhu@staff.sina.com.cn, jih@rpi.edu Abstract Target entity disambiguation (TED), the task of identifying target entities of the same domain, has been recognized as a critical step in various important applica- tions. In this paper, we propose a graph- based model called TremenRank to collec- tively identify target entities in short texts given a name list only. TremenRank prop- agates trust within the graph, allowing for an arbitrary number of target entities and texts using inverted index technology. Fur- thermore, we design a multi-layer direct- ed graph to assign different trust levels to short texts for better performance. The experimental results demonstrate that our model outperforms state-of-the-art meth- ods with an average gain of 24.8% in accu- racy and 15.2% in the F1-measure on three datasets in different domains. 1 Introduction Currently, a growing number of people prefer to express their views and comments online. These messages, which are updated at the rate of mil- lions per day, become a potentially rich source of information. From such a large number of texts, entity disambiguation is a critical step when ex- tracting text information from these messages, and for various applications, such as natural language processing and knowledge acquisition (Dredze et al., 2010). Take the application of customer feed- back analysis (CFA) as an example. An enterprise is typically interested in public reviews of its own products as well as those of its competitors’; the identification of these entities is thus critical for further analysis. The product names comprise a list of entities to be identified. We refer to these entities of the same domain as target entities, and the identifica- tion process is called target entity disambigua- tion (Wang et al., 2012). All of target entities (e.g., car brands) share a common domain, which we re- fer to as the target domain. The target domain is the only constraint to target entities, which implies that the entities are in a specific domain, rather than being general things. In the case of entity recognition from short texts, the disambiguation can be performed on the document level. Given a collection of short documents, our goal is to deter- mine which documents contain the target entities. Challenge and Related Work In contrast to traditional entity disambiguation tasks, TED in short texts can require as little in- formation as a name list. There are three types of information that are utilized in Named Entity Dis- ambiguation related tasks: knowledge resources, the context in which a target word occurs and sta- tistical information (Navigli, 2009). However, the lack …

Content truncated to 500 words