A Two-Step Named Entity Recognizer for Open-Domain Search Queries

International Joint Conference on Natural Language Processing, pages 829–833, Nagoya, Japan, 14-18 October 2013. A Two-Step Named Entity Recognizer for Open-Domain Search Queries Andreas Eiselt Yahoo! Research Latin America Av. Blanco Encalada 2120, Santiago, Chile eiselt@yahoo-inc.com Alejandro Figueroa Yahoo! Research Latin America Av. Blanco Encalada 2120, Santiago, Chile afiguero@yahoo-inc.com Abstract Named entity recognition in queries is the task of identifying sequences of terms in search queries that refer to a unique con- cept. This problem is catching increas- ing attention, since the lack of context in short queries makes this task difficult for full-text off-the-shelf named entity recog- nizers. In this paper, we propose to deal with this problem in a two-step fashion. The first step classifies each query term as token or part of a named entity. The sec- ond step takes advantage of these binary labels for categorizing query terms into a pre-defined set of 28 named entity classes. Our results show that our two-step strategy is promising by outperforming a one-step traditional baseline by more than 10%. 1 Introduction Search engines are key players in serving as in- terface between users and web resources. Hence, they started to take on the challenge of modelling user interests and enhance their search experience. This is one of the main drivers of replacing the classical document-keyword matching, a.k.a. bag- of-word approach, with user-oriented strategies. Specifically, these changes are geared towards im- proving the precision, contextualization, and per- sonalization of the search results. To achieve this, it is vital to identify fundamental structures such as named entities (e.g., persons, locations and organi- zations) (Hu et al., 2009). Indeed, previous studies indicate that over 70% of all queries contain enti- ties (Guo et al., 2009; Yin and Shah, 2010). Search queries are on average composed of 2- 3 words, yielding few context and breaking the grammatical rules of natural language (Guo et al., 2009; Du et al., 2010). Thus, named entity recog- nizers for relatively lengthy grammatically well- formed documents perform poorly on the task of Named Entity Recognition in Queries (NERQ). At heart, the contribution of this work is a novel supervised approach to NERQ, trained with a large set of manually tagged queries and consisting of two steps: 1) performs a binary classification, where each query term is tagged as token/entity depending on whether or not it is part of a named entity; and 2) takes advantage of these binary to- ken/entity labels for categorizing each term within the query into one of a pre-defined set of classes. 2 Related Work To the best of our knowledge, there have been a few previous research efforts attempting to recog- nize named entities in search queries. This prob- lem is relatively new and it was first introduced by (Pas¸ca, 2007). Their weakly supervised method starts with an input class represented by a set of seeds, which are used to induce typical query- contexts for the respective input category. Con- texts are then used to acquire and select new can- didate instances for …

Content truncated to 500 words