Resolving Surface Forms to Wikipedia Topics

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1335–1343, Beijing, August 2010 Resolving Surface Forms to Wikipedia Topics Yiping Zhou Lan Nie Omid Rouhani-Kalleh Flavian Vasile Scott Gaffney Yahoo! Labs at Sunnyvale {zhouy,lannie,omid,flavian,gaffney}@yahoo-inc.com Abstract Ambiguity of entity mentions and con- cept references is a challenge to mining text beyond surface-level keywords. We describe an effective method of disambi- guating surface forms and resolving them to Wikipedia entities and concepts. Our method employs an extensive set of fea- tures mined from Wikipedia and other large data sources, and combines the fea- tures using a machine learning approach with automatically generated training da- ta. Based on a manually labeled evalua- tion set containing over 1000 news ar- ticles, our resolution model has 85% pre- cision and 87.8% recall. The performance is significantly better than three baselines based on traditional context similarities or sense commonness measurements. Our method can be applied to other languages and scales well to new entities and con- cepts. 1 Introduction Ambiguity in natural language is prevalent and, as such, it can be a difficult challenge for infor- mation retrieval systems and other text mining applications. For example, a search for “Ford” in Yahoo! News retrieves about 40 thousand ar- ticles containing Ford referring to a company (Ford Motors), an athlete (Tommy Ford), a place (Ford City), etc. Due to reference ambiguity, even if we knew the user was only interested in the company, they would still have to contend with articles referring to the other concepts as well. In this paper we focus on the problem of re- solving references of named-entities and con- cepts in natural language through their textual surface forms. Specifically, we present a method of resolving surface forms in general text docu- ments to Wikipedia entries. The tasks of resolu- tion and disambiguation are nearly identical; we make the distinction that resolution specifically applies when a known set of referent concepts are given a priori. Our approach differs from oth- ers in multiple aspects including the following. 1) We employ a rich set of disambiguation features leveraging mining results from large- scale data sources. We calculate context- sensitive features by extensively mining the cat- egories, links and contents of the entire Wikipe- dia corpus. Additionally we make use of context- independent data mined from various data sources including Web user-behavioral data and Wikipedia. Our features also capture the one-to- one relationship between a surface form and its referent. 2) We use machine learning methods to train resolution models with a large automatically la- beled training set. Both ranking-based and classi- fication-based resolution approaches are ex- plored. 3) Our method disambiguates both entities and word senses. It scales well to new entities and concepts, and it can be easily applied to other languages. We propose an extensive set of metrics to eva- luate not only overall resolution performance but also out-of-Wikipedia prediction. Our systems for English language are evaluated using real- world test sets and compared with a number of baselines. …

Content truncated to 500 words