Event and Event Actor Alignment in Phrase Based Statistical Machine Translation
Anup Kumar Kolya, Santanu Pal, Asif Ekbal, Sivaji Bandyopadhyay · 2013
This paper proposes the impacts of event and event actor alignment in English and Bengali phrase based Statistical Machine Translation (PB-SMT) System. Initially, events and event actors are identified from English and Bengali parallel corpus. For events and event actor identification in English we proposed a hybrid technique and it was carried out within the TimeML framework. Events in Bengali are identified based on the concept of complex predicate structures. There can be one-to-one and one-to-many mappings between English and Bengali events and event actors. We preprocess the parallel corpus by single tokenizing the multiword events and event-actors which reflects some significant gain on the PB-SMT system. We represent a hybrid alignment approach of events and event-actors in both English-Bengali training corpus by defining a rule based aligner and a statistical hybrid aligner. The rule base aligner assumes a heuristic that the sequence of events and event actors on the source (English) side are also maintained in the target (Bengali) side. The performance of PB-SMT system could vary depending on the number of events and event-actors that are identified in the parallel training data. The proposed system achieves significant improvements (5.79 BLEU points absolute, 53.02% relative improvement) over the baseline system on an English-Bengali translation task.