Robust Document Representations for Cross-Lingual Information Retrieval in Low-Resource Settings

Robust Document Representations for Cross-Lingual Information Retrieval in Low-Resource Settings Mahsa Yarmohammadi1, Xutai Ma2, Sorami Hisamoto3, Muhammad Rahman4, Yiming Wang5, Hainan Xu6, Daniel Povey7, Philipp Koehn8 and Kevin Duh9 Center for Language and Speech Processing, Johns Hopkins University, Baltimore, MD, USA {mahsa1,xutai ma2,sorami3,mahbubur4,yiming.wang5,phi8}@jhu.edu {hainan.xv6,dpovey7}@gmail.com kevinduh@cs.jhu.edu9 Abstract The goal of cross-lingual information re- trieval (CLIR) is to find relevant docu- ments written in languages different from that of the query. Robustness to transla- tion errors is one of the main challenges for CLIR, especially in low-resource settings where there is limited training data for building machine translation (MT) systems or bilingual dictionaries. If the test col- lection contains speech documents, addi- tional errors from automatic speech recog- nition (ASR) makes translation even more difficult. We propose a robust document representation that combines N-best trans- lations and a novel bag-of-phrases out- put from various ASR/MT systems. We perform a comprehensive empirical analy- sis on three challenging collections; they consist of Somali, Swahili, and Tagalog speech/text documents to be retrieved by English queries. By comparing various ASR/MT systems with different error pro- files, our results demonstrate that a richer document representation can consistently overcome issues in low translation accu- racy for CLIR in low-resource settings. 1 Introduction Cross-lingual Information Retrieval (CLIR) is a search task where the user’s query is written in a language different from that of the documents in the collection. There are some important niche applications, for example, a local news reporter c⃝ 2019 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. searching foreign-language news-feeds to obtain different perspectives for her story, or a patent writer exploring the patents in another country to understand prior art before submitting her applica- tion, or an aid worker monitoring the social media of a disaster-affected area, looking for unmet needs and new emergencies. In all these scenarios, CLIR increases the user base by enabling users who are not proficient in the foreign language to produc- tively participate as knowledge workers. Even if the user requires manual translations of the re- trieved documents to complete her task, CLIR can at least provide a triage/filtering step. CLIR performance depends critically on the ac- curacy of its underlying machine translation or bilingual dictionary component. Recent advances in MT suggest that it is now ever more possi- ble to build CLIR for practical use. In particu- lar, the availability of large amounts of parallel text in some language-pairs (e.g. English sen- tences and their aligned German translations from European Parliamentary proceedings) had led to dramatic improvements in MT quality. How- ever, there are many language-pairs–what we term ”low-resource” settings–where parallel text is lim- ited and the challenge is to make CLIR robust to translation errors. Missing words in translation may lead to degradation in recall, while extrane- ous words may lead to degradation in precision. In this work, we focus on the document trans- lation approach to CLIR, where all foreign doc- uments in the collection are …

Content truncated to 500 words