Learning to translate from graded and negative relevance information

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3156–3166, Osaka, Japan, December 11-17 2016. Learning to translate from graded and negative relevance information Laura Jehl Computational Linguistics Heidelberg University 69120 Heidelberg, Germany jehl@cl.uni-heidelberg.de Stefan Riezler Computational Linguistics & IWR Heidelberg University 69120 Heidelberg, Germany riezler@cl.uni-heidelberg.de Abstract We present an approach for learning to translate by exploiting cross-lingual link structure in mul- tilingual document collections. We propose a new learning objective based on structured ramp loss, which learns from graded relevance, explicitly including negative relevance information. Our results on English-German translation of Wikipedia entries show small, but significant, im- provements of our method over an unadapted baseline, even when only a weak relevance signal is used. We also compare our method to monolingual language model adaptation and automatic pseudo-parallel data extraction and find small improvements even over these strong baselines. 1 Introduction Typically, parameters of an SMT system are learned on a small parallel data set from the domain or genre of interest. However, while many multilingual data sets, especially in the realm of user-generated data, contain document-level links, sentence-parallel training data are not always available. A small number of sentences can be manually translated for in-domain parameter tuning, but this ignores most of the available multilingual resource. Monolingual language model adaptation via concatenation or interpo- lation is one viable solution which makes use of the target side part of a collection (see e.g. Koehn and Schroeder (2007) or Foster and Kuhn (2007)). Additionally, there are several approaches to automatic parallel data extraction from cross-lingual document-level links, such as Munteanu and Marcu (2005)’s work on news data, or more recent work on Wikipedia by Wołk and Marasek (2015), and on websites by Smith et al. (2013). We argue that these approaches work well if the cross-lingual links are a strong signal for parallelism, but fail if the signal linking documents across languages is weaker. We propose a method for tuning sparse lexicalized features on large amounts of multilingual data which contain some cross-lingual document-level relevance annotation. We do so by re-formulating the structured ramp loss objective proposed by Chiang (2012) and Gimpel and Smith (2012) to incorporate graded and negative cross-lingual relevance signals. Using translation of Wikipedia entries as a running example, we evaluate the efficacy of our method along with the traditional approaches on a manually created in-domain test set. We show that our method is able to produce small, but significant, gains, even if only a weak relevance signal is used. Section 2 explains our learning objective and cost function. In Section 3 we describe the construction of our training and evaluation data, including pseudo-parallel data extraction. Section 4 contains details of our experimental setup and presents our experimental results. Section 5 concludes the paper. 2 Learning from graded relevance feedback 2.1 Learning objectives We work within a scenario where we want to learn the parameters of an SMT system, but have no in- domain reference translations available. What we have, is a …

Content truncated to 500 words