Lost and Found in Translation

Using cross-lingual question answering (CLQA), users can find information in languages that they do not know. In this thesis, we consider the broader problem of CLQA with result translation, where answers retrieved by a CLQA system must be translated back to the user's language by a machine translation (MT) system. This task is challenging because answers must be both relevant to the question and adequately translated in order to be correct. In this work, we show that integrating the MT closely with cross-lingual retrieval can improve result relevance and we further demonstrate that automatically correcting errors in the MT output can improve the adequacy of translated results. To understand the task better, we undertake detailed error analyses examining the impact of MT errors on CLQA with result translation. We identify which MT errors are most detrimental to the task and how different cross-lingual information retrieval (CLIR) systems respond to different kinds of MT errors. We describe two main types of CLQA errors caused by MT errors: lost in retrieval errors, where relevant results are not returned, and lost in translation errors, where relevant results are perceived irrelevant due to inadequate MT. To address the lost in retrieval errors, we introduce two novel models for cross-lingual information retrieval that combine complementary source-language and target-language information from MT. We show empirically that these hybrid, bilingual models outperform both monolingual models and a prior hybrid model. Even once relevant results are retrieved, if they are not translated adequately, users will not understand that they are relevant. Rather than improving a specific MT system, we take a more general approach that can be applied to the output of any MT system. Our adequacy-oriented automatic post-editors (APEs) use resources from the CLQA context and information from the MT system to automatically detect and correct phrase-level errors in MT at query time, focusing on the errors that are most likely to impact CLQA: deleted or missing content words and mistranslated named entities. Human evaluations show that these adequacy-oriented APEs can successfully adapt task-agnostic MT systems to the needs of the CLQA task. Since there is no existing test data for translingual QA or IR tasks, we create a translingual information retrieval (TLIR) evaluation corpus. Furthermore, we develop an analysis framework for isolating the impact of MT errors on CLIR and on result understanding, as well as evaluating the whole TLIR task. We use the TLIR corpus to carry out a task-embedded MT evaluation, which shows that our CLIR models address lost in retrieval errors, resulting in higher TLIR recall; and that the APEs successfully correct many lost in translation errors, leading to more adequately translated results.