Research Of Natural Language Evaluation Model Based On Hyperbolic Space

Language evaluation is an issue of concern in the field of natural language processing.Researchers in the fields of machine translation,text summarization,text retelling,image caption,etc.all rely on language evaluation models to carry out their research work.The language evaluation model can use computers to automatically,low-costly,and rapidly evaluate the quality of the text generation model by reference materials.Researchers can use the language evaluation model to carry out the work of model evaluation and model selection,or to conduct ablation experiments to analyze indepth the details of the text generation model.However,the current commonly used language evaluation models have some problems.Either they cannot measure the relevance between the reference text and the generated text from a semantic view,resulting in that the evaluation accuracy rate is far less than that of humans,or it will consume a lot of computation resources.The problems of existing language evaluation models have brought obstacles to the research work in related fields.Aiming at the deficiencies of existing research work,this thesis proposes a new language evaluation method,Hy LEU,which can take advantage of the powerful representation ability of hyperbolic space and use hyperbolic geodesic distance to measure the se-mantic correlation between ngram phrases more precisely.Furthermore,it can measure the semantic relationship between the reference text and the generated text more precisely.The specific work of this thesis is as follows:(1)This thesis proposes the hyperbolic word-embedding-alignment regularization,which using the highly structured semantic tree established by the Word Net knowledge base,to adjust the hyperbolic word embedding model Poincare Glo Ve.Experiments have proved that after the optimization by the hyperbolic word-embedding-alignment regularization,hyperbolic word embedding vectors have the characteristics of high cohesion and low coupling,and the geodesic distances of synonym pairs and unrelated word pairs in the hyperbolic space are significantly distinguishable.(2)By analyzing the structure of the word embedding in hyperbolic space,this thesis finds that the cosine similarity is not an ideal metric to measure the semantic similarity for hyperbolic word vectors.For this reason,this thesis proposes a method to calculate word similarity in hyperbolic space: distance mapping function.The distance mapping function can convert the hyperbolic geodesic distance into the similarity measure of the vocabulary.Word similarity evaluation experiments conducted on multiple datasets have proved the superior performance of this method.(3)This thesis established the Hy LEU model,a natural language evaluation model based on hyperbolic space.Hy LEU combines the first two work to better judge the semantic relevance between the reference text and the machinegenerated text,and then better complete the language evaluation task.Experiments on several machine translation evaluation tasks show that Hy LEU has lower computation cost and higher evaluation quality,and meets the expected design requirements.