Research On Key Issues In Entity Linking

Ellipsis recovery is a key task in the field of Natural Language Processing.By recovering the ellipsis,we can get complete sentences from both syntactic and semantic perspective.In comparison with other expression forms in NLP,dialogue is more difficult to be understood for machines due to the free and flexible expressions.Moreover,ellipsis recovery plays an important role for other NLP applications,such as question answering,automatic dialogue systems.In recent years,the rapid development of the Internet has produced a lot of data.How to make full use of these data has become the research hotspot.However,due the limitation of annotated resources,there is little related work on dialogue.In the context of dialogue,this paper carried out a series of work.(1)For scene of single round dialogue,we propose an end-to-end neural model based on sequence to sequence architecture.In particular,we encode the sequence in both directions,use attention mechanism to keep soft alignment from semantic perspective,and employ the generative decoding unit to get the final sequence.Experiment results show that,our model has a good robustness for all kinds of ellipsis.(2)In order to address some problems in the sequence to sequence generative model,we propose an improved scheme from three aspects.Firstly,we use a fine-grained word representation method combining with character information of words at the word embedding layer.Secondly,compared with the bidirectional encoding,the self-attention information of words is added,and the feature extraction ability of the encoding layer is improved.Finally,aiming at the exposure bias and semantic deviation of traditional generative decoding algorithm,a new decoding algorithm with sequence constraints is presented.Experiment results show that the above three improvements can effectively improve the performance of ellipsis recovery in a single round of dialogue.(3)For scene of multi-turn dialogue,we propose an end-to-end neural model based on the combination of dynamic copy and mask mechanism.In this model,gating mechanism is used to adaptively fuse the probability distribution of words in copy mode and generation mode,and two kinds of sequence boundary information are used to improve the accuracy of ellipsis recovery in decoding stage.Experiment results show the effectiveness of our approach.