Research Experiences and Projects

Currently, I’m focusing on the research of Statistical Machine Translation. My research interests also include topics like  Coreference Resolution and other NLP problems. I’m especially interested at using various machine learning techniques to solve complicated NLP problems.

  Statistical Machine Translation:                                                                                              Sep. 2008 - Now @NLP.NJU

We initiate a SMT project which will first build up a system with the help of Moses and then add extensions or modifications to involve more syntactic or semantic information. We are planning researches and experiments about various parts of the whole SMT procedure, for example word alignment, phrase extraction and scoring, phrase reordering, etc.

Our work starts with word alignment (Publication [2]). Observing that AER scores do not correlate well with BLEUs, we propose a variation, Error Sensitive AER (ESAER), for the evaluation of alignment results. ESAER takes the phrase extraction procedure into consideration and punishes different types of errors according to their affects to phrase extraction result. Experiments show a huge improvement of correlation to BLEU over original AER. We are also trying supervised and semi-supervised methods to improve the alignment result.

On a preprocessing step, we are trying to segmenting long sentence pairs into shorter ones, which may make followed steps more efficient. See Publication [6] for a preliminary result.

  Statistical Machine Translation for evaluation:                                                             Dec. 2008 - Aug. 2009 @NLP.NJU

We participated two machine translation evaluations, NIST Open MT 2009 Evaluation (MT09) (system report) and the evaluation of 5th China Workshop on Machine Translation (CWMT2009) (system report, in Chinese). Our systems were built mainly based on Moses. We added some phrase table feature and employed factored model to improve the result.

  Statistical Machine Translation:                                                                                    Apr. 2008 - Aug. 2008 @NLC.MSRA

Collaborating with Mu Li, we tried to improve the performance of word alignment by combining the outputs of different systems, for example Giza++, Berkeley aligner, etc. Experiments showed an increasing of BLEU score of about 0.5 points after combination.

I was also responsible for the rule extraction part of SMT system. Firstly I maintained the code for  Hiero rule extraction and did some rule filtering experiments. Later on, I also implemented some syntactic rule extraction procedures according to some papers from ISI, USC.

  Chinese Couplet project:                                                                                               Oct. 2007 - Mar. 2008 @NLC.MSRA

This project aims at building automatic second-line and banner generation systems for Chinese Couplet. I mainly worked with Long Jiang, on the automatic evaluation of second-line generation system. We thought about various evaluation metrics in IR and SMT, and finally employed BLEU, the evaluation metric of SMT, and built our evaluation framework based on some pseudo-reference second-lines.

  Coreference Resolution:                                                                                                  Jul. 2006 - Jul. 2007 @NLP.NJU

Coreference Resolution aims at identifying different mentions in a document that refer to the same entity. We firstly worked on unsupervised methods based on graph cut. Then some supervised learning techniques such as perceptron and Markov Logic Networks are employed. Also, a clustering technique called correlation clustering is used in the experiments. Later, we also did some research on automatic generation of first order logic rules and various clustering techniques. Some of these work are presented in Publication [1][3][5][7].

  Chinese word segmentation and POS tagging:                                                             Nov. 2005 - Apr. 2006 @NLP.NJU

In-house implementation of a Chinese word segmenter and a POS tagger based on HMM.