jeudi 1 novembre 2018

Paper Review 4 :SuperNMT: Neural Machine Translation with Semantic Supersenses and Syntactic Supertags

In this post, the paper "SuperNMT: Neural Machine Translation with Semantic Supersenses and Syntactic Supertags"is summarized.


Eva Vanmassenhove, Andy Way, SuperNMT: Neural Machine Translation with Semantic Supersenses and Syntactic Supertags, Proceedings of ACL 2018, Student Research Workshop, pages 67–73 Melbourne, Australia, July 15 - 20, 2018. Association for Computational Linguistics


Summary:

   Neural Machine Translation (NMT) learn by generalizing patterns based on the raw, sentences providing semantic supersensetags and syntactic supertag together make NMT system learn multi-word expressions.

This exclusive research enhances word-embeddings in NMT systems by providing a combination of semantic supersenses like (CCG supertags) and syntactic supertag like (SST).

CCG tags every word in a sentence with its correct role (verb/noun/auxiliary) and therefore resolve ambiguity in terms of prepositional attachment.

SST, in the other hand, classifies words independently into their part-of-speech (Modal/ Adverb/ noun)

These 2 new embedding vectors are then concatenated into the classical embedding vector and used in the model.

Adding explicitly this level of semantics provides the translation system with a higher level of abstraction beneficial to learn more complex constructions.

3 NMT systems are trained and tested on 1M sentences of the Europarl corpus for EN–FR and EN–DE:
One based on supersenses (SST), one on syntactic supertag (CCG), and one on both (SST–CCG).

Results for the EN–DE system shows that SST system converges faster, as hypothesized. The learning curve is also more consistent.

But if we focus on later stages of the learning process, CCG-SST model outperforms the best model: translations are 5% better comparing to the 2 other systems.

The results for the EN–DE system and EN–FR system are very similar and lead to the same conclusion: combining semantic and syntactic features is beneficial for generalization which lead to a better translation.

Aucun commentaire:

Enregistrer un commentaire

Presentation: Dorra EL MEKKI

Link :   presentation_dorraElMekki