Driving inversion transduction grammar induction with semantic evaluation

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 55–63, Berlin, Germany, August 11-12, 2016. Driving inversion transduction grammar induction with semantic evaluation Meriem Beloucif and Dekai Wu Human Language Technology Center Hong Kong University of Science and Technology Clear Water Bay, Hong Kong mbeloucif|dekai@cs.ust.hk Abstract We describe a new technique for improv- ing statistical machine translation training by adopting scores from a recent crosslin- gual semantic frame based evaluation met- ric, XMEANT, as outside probabilities in expectation-maximization based ITG (inversion transduction grammars) align- ment. Our new approach strongly biases early-stage SMT learning towards seman- tically valid alignments. Unlike previ- ous attempts that have proposed using se- mantic frame based evaluation metrics as the objective function for late-stage tun- ing of less than a dozen loglinear mix- ture weights, our approach instead applies the semantic metric at one of the earliest stages of SMT training, where it may im- pact millions of model parameters. The choice of XMEANT is motivated by em- pirical studies that have shown ITG con- straints to cover almost all crosslingual se- mantic frame alternations, which resem- ble the crosslingual semantic frame match- ing measured by XMEANT. Our exper- iments purposely restrict training data to small amounts to show the technique’s util- ity in the absence of a huge corpus, to study the effects of semantic generaliza- tions while avoiding overreliance on mem- orization. Results show that directly driv- ing ITG training with the crosslingual se- mantic frame based objective function not only helps to further sharpen the ITG con- straints, but still avoids excising relevant portions of the search space, and leads to better performance than either conven- tional ITG or GIZA++ based approaches. 1 Introduction We propose a new technique that biases early stage statistical machine translation (SMT) learn- ing towards semantics. Our algorithm adopts the crosslingual evaluation metric XMEANT (Lo et al., 2014) to initialize expectation-maximization (EM) outside probabilities during inversion trans- duction grammar or ITG (Wu, 1997) induction. We show that injecting a crosslingual semantic frame based objective function in the actual learn- ing of the translation model helps to bias the train- ing of the SMT model towards semantically more relevant structures. Our approach is highly mo- tivated by recent research which showed that in- cluding a semantic frame based objective function during the formal feature weights tuning stage in- creases the translation quality. More precisely, Lo et al. (2013a); Lo and Wu (2013); Lo et al. (2013b); Beloucif et al. (2014) showed that tuning against a semantic frame based evaluation metric like MEANT (Lo et al., 2012), improves the trans- lation adequacy. Our choice to improve ITG alignments is moti- vated by the fact that they have already previously been empirically shown to cover essentially 100% of crosslingual semantic frame alternations, even though they rule out the majority of incorrect align- ments (Addanki et al., 2012). Our technique uses XMEANT for rewarding good translations while learning bilingual correlations of the translation model. We also show …

Content truncated to 500 words