Automatic Grammatical Error Correction for ESL-Learners by SMT - Getting it right

Speaker:
Marcin Junczys-Dowmunt
Abstract:
There has been an increasing interest in using statistical machine translation (SMT) for the task of Grammatical Error Correction (GEC) for English-as-a-Second-Language (ESL) learners. Two of the three highest-scoring systems of the CoNLL-2014 Shared Task were SMT-based. The currently highest-scoring result published for the CoNLL-2014 test set has been achieved by a system combination of the five best CoNLL-2014 submissions built with MEMT (a tool for MT system combination).  In this talk, we demonstrate how a single SMT-based system can match and outperform the result of the mentioned combined system. Furthermore, this system outperforms any other published results (including our own CoNLL-2014 submission) for a single system by a margin of several percent F-Score when the same training data is being used.   These results are achieved by adapting current state-of-the art methods for phrase-based SMT specifically to the problem of GEC. We report on the effects of:  - Parameter tuning for GEC - Introducing GEC-specific dense and sparse features - Using large-scale data
Length:
01:18:51
Date:
05/01/2015
views:

Images:
Preview of 1.jpg
Image 1.jpg
Preview of 2.jpg
Image 2.jpg
Preview of 3.jpg
Image 3.jpg
Attachments: (video, slides, etc.)
109 MB
167 downloads
1.6 GB
191 downloads
572 MB
179 downloads
298 MB
166 downloads
212 MB
165 downloads
246 downloads