Fast Domain Adaptation of SMT models without in-Domain Parallel Data

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1114–1123, Dublin, Ireland, August 23-29 2014. Fast Domain Adaptation of SMT models without in-Domain Parallel Data Prashant Mathur∗ Fondazione Bruno Keseler Povo - 38100 (IT) first@fbk.eu Sriram Venkatapathy Xerox Research Center Europe Meylan (FR) first.last@xrce.xerox.com Nicola Cancedda∗ Microsoft, London (UK) first.last@gmail.com Abstract We address a challenging problem frequently faced by MT service providers: creating a domain- specific system based on a purely source-monolingual sample of text from the domain. We solve this problem by introducing methods for domain adaptation requiring no in-domain parallel data. Our approach yields results comparable to state-of-the-art systems optimized on an in-domain parallel set with a drop of as little as 0.5 BLEU points across 4 domains. 1 Introduction We consider the problem of creating the best possible statistical machine translation (SMT) system for a specific domain when no parallel sample or training data from such domain is available. We assume that we have access to a collection of phrase tables (PT) and other models independently created from now unavailable corpora, and we receive a monolingual source language sample from a text source we would like to optimize for. For a MT provider to deliver a SMT system tailored to a customer’s domain, a sample dataset is requested. In most cases, the customer is able to provide an in-domain mono-lingual sample from his operations. However, it is generally not feasible for the customer to provide the translations as well because the customer has to hire professional translators to do that. In such a scenario, the translations has to be generated by MT service provider itself by hiring human translators thus requiring an investment upfront. The methods proposed in this paper aim to avoid that by building a good quality pilot SMT system leveraging only sample mono-lingual source corpus, and previously trained library of models. This in turn postpones the task of generating in-domain parallel data to a later date when there is a commitment by the customer. Unavailability of the raw parallel data could derive from a trading model where data owners share intermediate-level resources like PTs, Reordering Models (RM) and Language Models (LM), but can not, or do not want to, share the textual data such resources were derived from. This particular scenario has been explained in (Cancedda, 2012). This scenario is similar to the multi-model framework studied in (Sennrich et al., 2013), with the additional challenge that no parallel development set is available. We build on the linear mixture model combination of the cited work, extending it to our more challenging environment: 1. We propose a new measure derived from the popular BLEU score (Papineni et al., 2002) to assess the fitness of a PT to cope with a given monolingual sample S. This measure is computed from n-gram statistics that can be easily extracted from a PT. 2. We propose a new method for tuning the parameters of a log linear model that does not require an in-domain parallel …

Content truncated to 500 words