Automatic Threshold Detection for Data Selection in Machine Translation

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 483–488 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics Automatic Threshold Detection for Data Selection in Machine Translation Mirela-Stefania Duma and Wolfgang Menzel University of Hamburg Natural Language Systems Division {mduma, menzel}@informatik.uni-hamburg.de Abstract We present in this paper the participation of the University of Hamburg in the Biomedical Translation Task of the Sec- ond Conference on Machine Translation (WMT 2017). Our contribution lies in adopting a new direction for perform- ing data selection for Machine Transla- tion via Paragraph Vector and a Feed For- ward Neural Network Classifier. Contin- uous distributed vector representations of the sentences are used as features for the binary classifier. Most approaches in data selection rely on scoring and ranking gen- eral domain sentences with respect to their similarity to the in-domain and setting a range of thresholds for selecting a percent- age of them for training various MT sys- tems. The novelty of our method consists in developing an automatic threshold de- tection paradigm for data selection which provides an efficient and simple way for selecting the most similar sentences to the in-domain. Encouraging results are ob- tained using this approach for seven lan- guage pairs and four data sets. 1 Introduction Data selection for Machine Translation (MT) rep- resents a standard domain adaptation technique with the aim of tackling the problem of select- ing from various general domain data the sen- tences that are most similar to sentences from the in-domain. Irrespective of having available vast amounts or small amounts of in-domain data, one of the advantages of data selection consists in pro- viding more in-domain data selected from large amounts of general domain data. Two difficult tasks arise when performing data selection: what method to use for scoring the sentences from the general domain according to their similarity to the in-domain and how many of the scored sentences to keep for later use in training MT systems. Standard state-of-the-art methods resolve the first difficulty by means of information retrieval, perplexity or edit distance methods. However, the second difficulty remains a challenge. There are no standard start-threshold and increment- threshold defined in the community. Axelrod et al. (2011), for example, uses the top N = {35k, 70k, 150k} sentence pairs from the scored general do- main data, while Bic¸ici and Yuret (2011) increas- ingly select N ∈ {100, 200, 500, 1000, 2000, 3000, 5000, 10000} instances for each test sentence for training and Kirchhoff and Bilmes (2014) select subsets of 10%, 20%, 30% and 40% of the data. We present a time and resource efficient method of performing data selection using Paragraph Vec- tor (Le and Mikolov, 2014) for representing the sentences and a Feed Forward Neural Network Classifier for determining which general domain sentences should be considered similar to the in- domain. The paragraph vectors and the binary classifiers are trained using standard parame- ters and have a great advantage of dropping the need to experiment with …

Content truncated to 500 words