A Multi-Task Architecture on Relevance-based Neural Query Translation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6339–6344 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 6339 A Multi-Task Architecture on Relevance-based Neural Query Translation Sheikh Muhammad Sarwar Hamed Bonab Center for Intelligent Information RetrievalJames Allan College of Information and Computer Sciences University of Massachusetts Amherst Amherst, MA 01003 {smsarwar, bonab, allan}@cs.umass.edu Abstract We describe a multi-task learning approach to train a Neural Machine Translation (NMT) model with a Relevance-based Auxiliary Task (RAT) for search query translation. The trans- lation process for Cross-lingual Information Retrieval (CLIR) task is usually treated as a black box and it is performed as an indepen- dent step. However, an NMT model trained on sentence-level parallel data is not aware of the vocabulary distribution of the retrieval cor- pus. We address this problem with our multi- task learning architecture that achieves 16% improvement over a strong NMT baseline on Italian-English query-document dataset. We show using both quantitative and qualitative analysis that our model generates balanced and precise translations with the regulariza- tion effect it achieves from multi-task learning paradigm. 1 Introduction CLIR systems retrieve documents written in a lan- guage that is different from search query language (Nie, 2010). The primary objective of CLIR is to translate or project a query into the language of the document repository (Sokokov et al., 2013), which we refer to as Retrieval Corpus (RC). To this end, common CLIR approaches translate search queries using a Machine Translation (MT) model and then use a monolingual IR system to retrieve from RC. In this process, a translation model is treated as a black box (Sokolov et al., 2014), and it is usually trained on a sentence level parallel corpus, which we refer to as Translation Corpus (TC). We address a pitfall of using existing MT models for query translation (Sokokov et al., 2013). An MT model trained on TC does not have any knowl- edge of RC. In an extreme setting, where there are no common terms between the target side of TC and RC, a well trained and tested translation model would fail because of vocabulary mismatch between the translated query and documents of RC. Assuming a relaxed scenario where some com- monality exists between two corpora, a translation model might still perform poorly, favoring terms that are more likely in TC but rare in RC. Our hy- pothesis is that a search query translation model would perform better if a translated query term is likely to appear in the both retrieval and translation corpora, a property we call balanced translation. To achieve balanced translations, it is desired to construct an MT model that is aware of RC vo- cabulary. Different types of MT approaches have been adopted for CLIR task, such as dictionary- based MT, rule-based MT, statistical MT etc. (Zhou et al., 2012). However, to the best of our knowl- edge, a neural search query translation approach has yet to be taken by the community. NMT models …

Content truncated to 500 words