EXTENDED CONVOLUTIONAL NEURAL NETWORKS POST-TRAINED WITH FACTORED STATISTICAL MACHINE

This paper investigate the potential of coupling two machine translation research approaches while taking full advantage of each method, namely, the deterministic (neuronal) and probabilistic (statistical) approaches, in order to address three main problems occurring in MT, that is, language pairs having grammatical structure and word order that differs drastically, data sparseness and the number of out of vocabulary (OOV) words generated. Additionally, we integrated word-level linguistic features (Part-of- Speech with compounds, lemmatization and/or word class) so as to decrease the number of unknown words while significantly increasing the vocabulary coverage. We combined a fully-factored ConvS2S and a factored PB-SMT where, the pre-translated training data generated using the NMT system is employed to build the SMT system, parallely, tuning parameters using the pre-translated development set. Finally, the desired results where produced by operating in the post-translation step the SMT system to re-decode the pre-translated test set. Experiments are performed on comparisons between stand-alone vs. our hybrid MT models where our model outperforms the strong MTs state-of-the-art by over 3.88 %BLEU, 3.08 %BLEU and 3.52 %BLEU, on the WMT’16 English-Romanian, WMT’14 English-German and WMT’14 English-French translation tasks, respectively, and hybrid recurrent vs. hybrid ConvS2S MT models where our model outperforms the strong HMTs state-of-the-art by over 3.95 %BLEU and 5.68 %BLEU on the Japanese-English and Chinese-English test sets, respectively.Although the adequacy of statistical machine translations (SMT) has been surpassed or rivalled by the fluency of the neural machine translations models [1],[2],[3],[4],[5], which have achieved the state of the art in machine translation due to their fundamental design, NMT models are constraint by the need for the observation of many examples of a word till the reliability of its embedding (input representation) and the computational necessity that they use a fixed modest-size vocabulary to few tens of thousands of words which most often leads to coverage issues and the loss of rare words identity, thus performs poorly in translating infrequent or unseen words, by so doing NMT models still face major problems when dealing with morphologically-rich and/or low resource languages, which are characterized by their high lexical sparseness, due to the occurrence of unseen word forms (commonly known as uncommon and rare words) in source texts during training, caused by their lexical productivity and morphological variation