| The information in the field of finance and economics contains great value.The analysis of the information in the field of finance and economics is significant to guide individual and social economic activities.However,the amount of information in the field of finance and economics is large,and the growth rate is fast.It is difficult to make full use of this information,if we process this information manually.The entity relationship in free text can be extracted automatically by using relation extraction technology,which can help analysts make use of information efficiently.So it is important to study the entity relationship extraction technology in the field of finance and economics.However,the relationship extraction methods or models in other field can't achieve higher precision and recall in the field of finance and economics,because these is something special about the information in the field of finance and economics.Specially:The length of sentence about this information is generally longer;These are some words with special meanings in this information which are different from the meanings in usual;There are some slangy expressions with some characteristics;Some sentence contains multiple domain entities and different entity pairs have different relationship types.In view of the characteristics of information in the field of finance and economics and the shortcomings of the existing method models,this thesis proposes a joint model based on BiGRU and attention mechanism named BGAJM model.The model has the following characteristics:(1)In view of the fact that the sentence length is generally long.The BGAJM model uses BiGRU as the hidden layer netxwork structure.Compared with the model using BiLSTM structure,the BiGRU has simpler structure and fewer learning parameters.When the sentence length is long,it can avoid the problem of gradient disappearance or explosion better.This is helpful for the model to learn better features,and then improve the precision and recall of the model.(2)In view of the fact that a sentence contains multiple entities and there are different relationships between different entities.BGAJM model uses attention mechanism to automatically assign different words different weights in order to extract the relationship between different entities.Compared with the traditional method of using the distance between words and entities as the feature when the sentence length is long then the distance feature is sparse and noisy,the BGAJM model uses the semantic information of the word with better discrimination as the feature.At the same time,the input dimension is reduced by twice the sentence length,then the calculation is reduced.(3)In view of the fact that some words in financial domain information have domain meaning and are quite different from common usage,as well as the characteristics of slang expression.The BGAJM model uses word vectors specially trained by using information in the financial field in the word embedding layer.This can help BGAJM model learn the domain semantic features of words better,and then improve the precision and recall of the model.(4)BGAJM model is a joint model,which is combined by entity recognition module and relation extraction module by sharing parameters in training,which achieves the effect that recognizes entity and extracts relationship in a model.The error propagation problem caused by entity recognition and relation extraction as two independent subtasks is avoided,and the performance of the model is improved.The experiment shows that the BGAJM model proposed by this thesis has a better effect than other models in the specific field of relationship extraction in the field of finance and economics.The area under the P-R curve of BGAJM model is 0.87,more than other models 0.19.Finally,based on BGAJM model,this thesis developed an information extraction system in the financial field.The system can automatically crawl information and extract relationships,which has a certain practical value. |
PDF Full Text Request