Multi-intent Recognition And Slot Filling For Interrogated Text

Intent recognition and slot filling tasks are key tasks in natural language semantic parsing and have mature applications in many downstream tasks.Especially in assisted consultation,it not only can interact with patients,answer their questions,guide them to tell their conditions and thus get more information about their symptoms,but also saves a lot of time for doctors and patients and greatly improves efficiency.Currently,there are still some shortcomings in the task of intention recognition and slot filling in Chinese medical field,such as the coarse granularity of intention recognition,which cannot correspond to specific slots in the text,and the frequent misspellings and colloquialisms in the question corpus described by users.In order to parse the user’s interrogative text more precisely and exhaustively,this paper investigates the following three subjects:(1)In order to build an intention model that can recognize corresponding slots and enhance the accuracy of natural language parsing in Chinese,this paper proposes a joint model of intention recognition and slot filling that incorporates word sound and word shape features.The model encoding side is based on MC-BERT model,incorporating word sound and word shape features,which not only enriches the semantic representation of text but also improves the tolerance of misspelled words;in the intention recognition decoding side,the intention of each Token is recognized and the intention type of the corresponding slot is voted,which improves the accuracy of intention recognition while mitigating the misleading degree of slot filling when the intention recognition is wrong;Incorporating Token-level intent information at the slot-filling decoding end to achieve the guiding effect of intent information on slot identification.(2)In order to be able to capture information about the yin and yang of disease and symptom slots of the interrogative corpus species.In this paper,we develop an aspect-level negativity discrimination model incorporating word sounds and word forms.The model embeds text sequences and aspect words in the MC-BERT input,namely,the word encoding,position encoding and sentence encoding of both are passed into the MC-BERT model,so that the model can identify the specific aspect word meaning and the position of the aspect word in the text sequence,thus solving the problem of multiple identical aspect words in the question text described by the user and inconsistent labeling of the aspect words.(3)Based on the medical questioning text,the dataset of intent types labeled on slots is constructed,which makes the joint model of intent recognition and slot filling have an F1 value of 93.13% in slot filling recognition and an overall F1 value of 89.38%in recognizing slots and the intent categories corresponding to slots,both of which are better than the comparison algorithm.Based on the feminine and masculine discrimination dataset of the interrogated text,the proposed aspect-level feminine and masculine discrimination model achieves a Macro-F1 value of 80.70% for the model on the test set.With the F1 values around 90% for label prediction of negative and positive,the model can accurately discriminate the femininity and masculinity of the patient’s interrogation text.