Grammar Constrained Double-Layer Encoder Decoder For Neural Semantic Parsing

A great deal of information in today's world is stored in structured and semi-structured knowledge bases.An efficient and easy way to query it is essential and cannot be limited to people with expertise in query languages.The field of semantic parsing involves the conversion of natural language statements into computer-executable logical forms.With the help of semantic parsing,one can convert natural language statements into logical form without the help of experts through semantic parsers.The development of semantic parsing goes through three stages,rule-based model,statisticbased learning model,and neural network-based model,with each stage less dependent on experts than the previous stage.The Seq2 seq model based on encoder-decoder architecture is a neural network based model.Some models introduce the abstract syntax tree,and the syntax a priori of the meaning representation is incorporated into the action sequence that is converted by traversing the abstract syntax tree,compared to the model that treats the meaning representation as a sequence.To address the above issues,a grammar-based double layer encoder decoder system(GCDL enc-dec system)is proposed.The main work consists of the following two parts.In the first part,in order to introduce a syntactic a priori to ensure that the model generates legal meaning representations,the grammar model in the GCDL enc-dec system and the transition system are constructed,and the grammar model implements the translation of meaning representations and trees.The transition system implements the transformation of a tree into a sequence of operations,by which the syntax is incorporated a priori into the sequence of operations and the problem of generating a meaning representation is transformed into the problem of generating a sequence of actions.In the second part,the GCDL enc-dec model in the GCDL enc-dec system is constructed,which decomposes the meaning representation generation process into two processes,the first layer encoder-decoder encodes the text to generate the aciton sequence of abstract sketch tree,and the second layer encoder-decoder encodes the results of the previous layer and fuses the encoded information of the first layer to generate the actions sequence of abstract syntax tree.The GCDL encdec model uses a two-layer structure to first generate the less complex abstract sketch tree sequence and then generate the final abstract syntax tree.The grammar model,the transition system,and the GCDL enc-dec model make up the GCDLenc-dec system.The experimental results show that the GCDL enc-dec system has improved accuracy and BLEU scores on data sets such as semantic parsing and code generation,proving the effectiveness of the system.