Show simple item record

dc.contributor.authorGupta, Rohit
dc.date.accessioned2017-01-17T16:22:03Z
dc.date.available2017-01-17T16:22:03Z
dc.date.issued2016
dc.identifier.urihttp://hdl.handle.net/2436/620338
dc.descriptionA thesis submitted in partial fulfilment of the requirements of the University of Wolverhampton for the degree of Doctor of Philosophy
dc.description.abstractCurrent Translation Memory (TM) tools lack semantic knowledge while matching. Most TM tools compute similarity at the string level, which does not take into account semantic aspects in matching. Therefore, semantically similar segments, which differ on the surface form, are often not retrieved. In this thesis, we present five novel and efficient approaches to incorporate advanced semantic knowledge in translation memory matching and retrieval. Two efficient approaches which use a paraphrase database to improve translation memory matching and retrieval are presented. Both automatic and human evaluations are conducted. The results on both evaluations show that paraphrasing improves matching and retrieval. An approach based on manually designed features extracted using NLP systems and resources is presented, where a Support Vector Machine (SVM) regression model is trained, which calculates the similarity between two segments. The approach based on manually designed features did not retrieve better matches than simple edit-distance. Two approaches for retrieving segments from a TM using deep learning are investigated. The first one is based on Long Short Term Memory (LSTM) networks, while the other one is based on Tree Structured Long Short Term Memory (Tree-LSTM) networks. Eight different models using different datasets and settings are trained. The results are comparable to a baseline which uses simple edit-distance.
dc.language.isoen
dc.titleUSE OF LANGUAGE TECHNOLOGY TO IMPROVE MATCHING AND RETRIEVAL IN TRANSLATION MEMORY
dc.typeThesis or dissertation
refterms.dateFOA2018-08-21T13:40:04Z
html.description.abstractCurrent Translation Memory (TM) tools lack semantic knowledge while matching. Most TM tools compute similarity at the string level, which does not take into account semantic aspects in matching. Therefore, semantically similar segments, which differ on the surface form, are often not retrieved. In this thesis, we present five novel and efficient approaches to incorporate advanced semantic knowledge in translation memory matching and retrieval. Two efficient approaches which use a paraphrase database to improve translation memory matching and retrieval are presented. Both automatic and human evaluations are conducted. The results on both evaluations show that paraphrasing improves matching and retrieval. An approach based on manually designed features extracted using NLP systems and resources is presented, where a Support Vector Machine (SVM) regression model is trained, which calculates the similarity between two segments. The approach based on manually designed features did not retrieve better matches than simple edit-distance. Two approaches for retrieving segments from a TM using deep learning are investigated. The first one is based on Long Short Term Memory (LSTM) networks, while the other one is based on Tree Structured Long Short Term Memory (Tree-LSTM) networks. Eight different models using different datasets and settings are trained. The results are comparable to a baseline which uses simple edit-distance.


Files in this item

Thumbnail
Name:
Gupta_PhD thesis.pdf
Size:
904.9Kb
Format:
PDF

This item appears in the following Collection(s)

Show simple item record