Research On The Representation Of Word Embedding Based On Knowledge Fusion

The representation of words is the core problem of natural language processing based on deep learning.Word embedding is widely used in many natural language processing tasks,which is an effective way to improve the performance of downstream tasks.At present,the mainstream word embedding learning model is based on the distributed hypothesis,but there are some problems in this method.According to Ziff's law,there are always a lot of low frequency words in the corpus.Because the language phenomenon of low frequency words is not rich,the semantic representation quality of low frequency word embedding is poor,and they can not represent their own semantics well.To solve this problem,this paper proposes a low frequency word enhancement method based on How Net Related Concept Field.This method uses it as pseudo context of low frequency words without the help of corpus.The experimental results show that this method achieves better experimental results,and can improve the semantic representation quality of low frequency words better than the baseline method.The method of knowledge fusion is an effective way to improve the performance of deep learning system.Generally,knowledge fusion methods only consider the elements of knowledge base,but not the structure of its network.In order to solve this problem,this paper proposes a network embedding driven word representation learning model based on retrofitting.Firstly,the elements and network structure in the knowledge base are encoded into vectors,and then the word embedding learning process is based on the vectors by retrofitting constraint.The experimental results show that the model proposed in this paper has achieved good experimental results,and can obtain higher quality word vectors than the two strong baseline models.Finally,we build a word representation learning system based on the two methods proposed in this paper and give a detailed description of the design and implementation of the system.