Bag-of-Words Forced Decoding for Cross-Lingual Information Retrieval

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1172–1182, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Bag-of-Words Forced Decoding for Cross-Lingual Information Retrieval Felix Hieber Computational Linguistics Heidelberg University 69120 Heidelberg, Germany hieber@cl.uni-heidelberg.de Stefan Riezler Computational Linguistics Heidelberg University 69120 Heidelberg, Germany riezler@cl.uni-heidelberg.de Abstract Current approaches to cross-lingual informa- tion retrieval (CLIR) rely on standard retrieval models into which query translations by sta- tistical machine translation (SMT) are inte- grated at varying degree. In this paper, we present an attempt to turn this situation on its head: Instead of the retrieval aspect, we em- phasize the translation component in CLIR. We perform search by using an SMT decoder in forced decoding mode to produce a bag-of- words representation of the target documents to be ranked. The SMT model is extended by retrieval-specific features that are optimized jointly with standard translation features for a ranking objective. We find significant gains over the state-of-the-art in a large-scale eval- uation on cross-lingual search in the domains patents and Wikipedia. 1 Introduction Approaches to CLIR have been plentiful and di- verse. While simple word translation probabilities are easily integrated into term-based retrieval mod- els (Berger and Lafferty, 1999; Xu et al., 2001), state-of-the-art SMT systems (Koehn, 2010; Chi- ang, 2007) are complex statistical models on their own. The use of established translation models for context-aware translation of query strings, effec- tively reducing the problem of CLIR to a pipeline of translation and monolingual retrieval, has been shown to work well in the past (Chin et al., 2008). Only recently, approaches have been presented to include (weighted) translation alternatives into the query structure to allow a more generalized term matching (Ture et al., 2012a; Ture et al., 2012b). However, this integration of SMT remains agnostic about its use for CLIR and is instead optimized to match fluent, human reference translations. In con- trast, retrieval systems often use bag-of-word repre- sentations, stopword filtering, and stemming tech- niques during document scoring, and queries are rarely fluent, grammatical natural language queries (Downey et al., 2008). Thus, most of a translation’s structural information is lost during retrieval, and lexical choices may not be optimal for the retrieval task. Furthermore, the nature of modeling transla- tion and retrieval separately requires that a single query translation is selected, which is usually done by choosing the most probable SMT output. Attempts to inform the SMT system about its use for retrieval by optimizing its parameters towards a retrieval objective have been presented in the form or re-ranking (Nikoulina et al., 2012) or ranking (Sokolov et al., 2014). In this paper, we take this idea a step further and directly integrate the task of scoring documents with respect to the query into the process of translation decoding. We make the full expressiveness of the translation search space available to the retrieval model, without enumerat- ing all possible translation alternatives. This is done by augmenting the linear model of the …

Content truncated to 500 words