Research On Machine Reading Comprehension Model Based On Cross-lingual Transfer Technology

The research on machine reading comprehension has attracted extensive attention from many researchers in the field of natural language processing.At present,most machine reading comprehension tasks all rely on word vectors to achieve,and the unity of the word vector space makes the same model not universal in different languages.At the same time,most models tend to be trained for large languages like English,and there are too few experiments in other languages.If a separate machine reading comprehension model is established in the research field of each language,it will be takes a lot of time and cost.In response to these problems,this thesis studies the method of cross-lingual machine reading comprehension.First,for the monolingual extractive machine reading comprehension task,a Bi DAF model that integrates the self-attention mechanism is proposed to enhance the information interaction and effectiveness between the contexts and questions and the ability of extract features,also use the evaluation indicators of machine reading comprehension tasks as evaluation criteria to evaluate the effect of crosslingual transfer.The second is the study of cross-lingual transfer technology,which is mainly achieved in two ways:(1)Aiming at the word embedding method to implement a cross-lingual machine reading comprehension model.This thesis proposes a method of using pre-trained word vectors in different languages to build a cross-lingual word embedding model,using adversarial learning or self-learning methods to train shared word vectors under the situation of relying on parallel corpus,an unsupervised cross-lingual word embedding model is realized.In the adversarial training process,the Procrustes analysis method is introduced to fine-tune the transfer matrix and the word vector pairs,so that the word vector representations in the shared word vector space with the same meaning and different languages are similar.In self-learning training,cross-domain similarity local scaling is used to locate the mapped word vector pairs and iteratively amplify the word vector pairs.The final experimental results show that both the Procrustes analysis method and the cross-domain similarity local scaling have significantly improved the training effect of the cross-lingual word embedding model;(2)A cross-lingual machine reading comprehension model is implemented based on the translation alignment method.First,the target language is translated into the form of the source language,and if the target language has no training corpus,the model training is carried out with the help of the source language training corpus,and finally the answer is translated back into the target language.For the problem that the translated answer does not match the target language,consider using character-level matching to align the answer.At the same time,when there is a certain training corpus in the target language,the real answer of the target language can be merged and matched with the obtained answer.And verification,the experiment proves that this method is very effective to achieve the crosslingual machine reading comprehension task.