Feature-Enriched Character-Level Convolutions for Text Regression

Proceedings of the Conference on Machine Translation (WMT), Volume 2: Shared Task Papers, pages 575–581 Copenhagen, Denmark, September 711, 2017. c⃝2017 Association for Computational Linguistics Feature-Enriched Character-Level Convolutions for Text Regression Gustavo Henrique Paetzold and Lucia Specia Department of Computer Science University of Sheffield, UK {g.h.paetzold,l.specia}@sheffield.ac.uk Abstract We present a new model for text re- gression that seamlessly combine engi- neered features and character-level infor- mation through deep parallel convolution stacks, multi-layer perceptrons and multi- task learning. We use these models to create the SHEF/CNN systems for the sentence-level Quality Estimation task of WMT 2017 and Emotion Intensity Analy- sis task of WASSA 2017. Our experiments reveal that combining character-level clues and engineered features offers notice- able performance improvements over us- ing only one of these sources of informa- tion in isolation. 1 Introduction Text regression consists in estimating a numeric label based on information available from the text. The label can represent any abstract property of said text: its appropriateness, sentiment, fluency, simplicity, quality, etc. Due to their wide ap- plicability in both research and industry, some of these tasks have been gaining a lot of atten- tion. These include Quality Estimation and Emo- tion Intensity Analysis, which are the subjects of shared tasks held at the WMT 2017 conference1 and WASSA 2017 workshop2 (Mohammad and Bravo-Marquez, 2017), respectively. In Quality Estimation (QE), one attempts to es- timate the quality of a machine translated text based on the information that can be extracted from the original sentence and its translation. The task has many variants, given that the quality of a translation can be estimated at word, phrase, 1http://www.statmt.org/wmt17 2http://optima.jrc.it/wassa2017 sentence or even document level. Quality esti- mates can be incorporated in Machine Transla- tion (MT) decoding or used for re-ranking of top candidates, for example, allowing for a more in- telligently guided translation process (Avramidis, 2012), or they can be used to help human transla- tors decide which automatic translations are worth post-editing, and which should be re-translated from scratch (Turchi et al., 2015). Sentence-level QE is the most popular variant, mostly due the fact that most modern statistical and neural MT systems translate one sentence at a time. In this task, the input is the original-translated sentence pair and the output is some numeric label that rep- resents quality. The most commonly used label is HTER, which measures the human post-editing effort required to fix the translation in question (Snover et al., 2006). As shown in (Bojar et al., 2016), the perfor- mance of QE approaches submitted to the WMT shared tasks have steadily improved in recent years. However, the nature of these approaches have not changed much: most of the top rank- ing systems employ well-known regression meth- ods and extensive feature engineering. Some of the most notable examples are the RTM systems of WMT 2014 and 15, which managed to reach the top of the ranks by employing Referential Translation Machines trained with SVMs for re- gression (Bicici, 2016). The LORIA (Langlois, 2015) and YSDA …

Content truncated to 500 words