Research And Implementation Of Intention Recognition Based On Multi-dimensional Feature Fusion

Intention Detection and Slot Filling are two key tasks of natural language understanding.The mainstream research method is to establish a joint recognition model of intention and slot filling,and reflect the correlation between them through parameter sharing.However,the existing methods lack the attention to the word level features of text,and it is difficult to obtain fine-grained classification features.In terms of joint modeling,the existing methods also ignore the two-way information flow between the two tasks,and can not well model the correlation between them.Therefore,thesis studies the joint detection of intention and slot of multi-dimensional feature fusion.The main work is as follows:1.Feature extraction module based on multi-dimensional semantic representationAiming at the problem that the traditional word embedding method lacks the mapping of part of the vocabulary,which leads to the lack of text semantics,a method of integrating multiple word embeddings is designed to alleviate the influence of the lack of vocabulary mapping vectors.And use the local self-attention mechanism and Bi LSTM to extract semantic features of different scales,and finally prove the effectiveness of this module through experiments.2.Cross attention module based on two-way information flowAiming at the problem that the model lacks the correlation between the task of intent detection and slot filling,a cross-attention module based on bidirectional information flow is designed,so that the model can learn the complete joint distribution explicitly.Finally,it is experimentally verified that the model outperforms the baseline methods on multiple evaluation metrics on both datasets.3.Joint recognition system of intention and slot for flight query sceneIn order to reduce the labor cost in the flight query scenario,an intent and slot joint detection system for the flight query scenario is designed and implemented,so that the system can identify the user's query intent and return the flight information.An opensource application framework is used to build a system that predicts the intent and slot labels of the input text by invoking the joint model constructed in this study,and queries and visualizes the corresponding flight information based on the slot information.