Open Entity Extraction from Web Search Query Logs

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 510–518, Beijing, August 2010 Open Entity Extraction from Web Search Query Logs Alpa Jain Yahoo! Labs alpa@yahoo-inc.com Marco Pennacchiotti Yahoo! Labs pennac@yahoo-inc.com Abstract In this paper we propose a completely un- supervised method for open-domain en- tity extraction and clustering over query logs. The underlying hypothesis is that classes defined by mining search user activ- ity may significantly differ from those typ- ically considered over web documents, in that they better model the user space, i.e. users’ perception and interests. We show that our method outperforms state of the art (semi-)supervised systems based either on web documents or on query logs (16% gain on the clustering task). We also report evi- dence that our method successfully supports a real world application, namely keyword generation for sponsored search. 1 Introduction Search engines are increasingly moving beyond the traditional keyword-in document-out paradigm, and are improving user experience by focusing on user- oriented tasks such as query suggestions and search personalization. A fundamental building block of these applications is recognizing structured infor- mation, such as, entities (e.g., mentions of people, organizations, or locations) or relations among en- tities (Cao et al., 2008; Hu et al., 2009). For this, search engines typically rely on large collections of entities and relations built using information extrac- tion (IE) techniques (Chaudhuri et al., 2009). Commonly used IE techniques follow two main assumptions: (1) IE focuses on extracting infor- mation from syntactically and semantically “well- formed” pieces of texts, such as, news corpora and web documents (Pennacchiotti and Pantel, 2009); (2) extraction processes are bootstrapped with some pre-existing knowledge of the target domain (e.g entities are typically extracted for pre-defined cat- egories, such as Actors, Manufacturers, Persons, Locations (Grishman and Sundheim, 1996)). Prior work (Banko et al., 2007), has looked into relax- ing the second assumption and proposed open in- formation extraction (OIE), a domain-independent and scalable extraction paradigm, which however focuses mostly on web corpora. In this paper, we argue that for user-oriented ap- plications discussed earlier, IE techniques should go beyond the traditional approach of using “well- formed” text documents. With this in mind, we ex- plore the utility of search query logs, a rich source of user behaviors and perception, and build tech- niques for open entity extraction and clustering over query logs. We hypothesize that web docu- ments and query logs model two different spaces: web documents model the web space, i.e. general knowledge about entities and concepts in an objec- tive and generic way; search query logs model the user space, i.e. the users’ view and perception of the world in a more specific fashion, where avail- able information directly expresses users’ needs and intents. For example, in a web space, ‘brit- ney spears’ will tend to be similar and be clus- tered with other singers, such as ‘celine dion’ and ‘bruce springsteen’. On the contrary, in the users’ space, she is highly similar and clustered with other gossiped celebrities like …

Content truncated to 500 words