CN116306589B

CN116306589B - Method and device for medical text error correction and intelligent extraction of emergency scene

Info

Publication number: CN116306589B
Application number: CN202310521134.7A
Authority: CN
Inventors: 李劲松; 杨宗峰; 周逸飞; 田雨; 周天舒
Original assignee: Zhejiang Lab
Current assignee: Zhejiang Lab
Priority date: 2023-05-10
Filing date: 2023-05-10
Publication date: 2024-02-09
Anticipated expiration: 2043-05-10
Also published as: CN116306589A

Abstract

The specification discloses a method and a device for medical text error correction and intelligent extraction of an emergency scene, which can acquire an emergency medical text subjected to voice recognition, then determine error positions in the emergency medical text according to a statistical language model and/or an error word recognition model, further determine candidate replacement words corresponding to each error position, determine candidate error correction texts subjected to error correction on the emergency medical text according to the candidate replacement words corresponding to each error position, so as to select a target text from the candidate error correction texts, finally input a preset medical information type and the target text into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and input the prompt information vector and the target text into a second network layer of the information extraction model, so that medical information under the medical information type is extracted from the target text through the information extraction model, and the accuracy of information extraction is improved.

Description

Method and device for medical text error correction and intelligent extraction of emergency scene

Technical Field

The present disclosure relates to the field of neural networks, and in particular, to a method and apparatus for error correction and intelligent extraction of medical text in emergency situations.

Background

In medical systems, it is often necessary to enter some information of a patient into a system of a hospital, so that subsequent doctor treatment and re-diagnosis are facilitated, and the same is true in pre-hospital medical emergency, and medical staff can upload medical information related to the patient into the system of the hospital during pre-hospital medical emergency.

Currently, medical information related to a patient is entered into a hospital system, and a session of medical personnel input can be converted into structured medical information, which is then stored in the system. However, the biggest difference between pre-hospital medical first aid and medical treatment in hospital is that medical staff can not input a certain speech manually in emergency, but the speech input can be caused by different human accent differences, errors of on-site dictation, noise of on-site environment and other factors, more errors can exist in the speech recognized text, and the speech recognized text is converted into structured medical information, so that the accuracy of the result can be influenced.

Therefore, how to improve the accuracy of extracting the structured medical information of the patient is a urgent problem to be solved.

Disclosure of Invention

The present disclosure provides a method and apparatus for medical text correction and intelligent extraction in emergency situations, so as to partially solve the above-mentioned problems in the prior art.

The technical scheme adopted in the specification is as follows:

the specification provides a method for medical text error correction and intelligent extraction of emergency scenes, which comprises the following steps:

receiving an information extraction request, and acquiring a voice-recognized emergency medical text through the information extraction request;

determining the error position in the emergency medical text according to a preset statistical language model and/or an error word recognition model;

determining candidate replacement words corresponding to each error position, and determining each candidate error correction text after error correction of the emergency medical text according to the candidate replacement words corresponding to each error position so as to select a target text from each candidate error correction text;

inputting a preset medical information type and the target text into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the target text into a second network layer of the information extraction model, so that medical information under the medical information type is extracted from the target text through the information extraction model.

Optionally, determining, according to a preset statistical language model, an error location existing in the emergency medical text specifically includes:

for each word in the emergency medical text, determining a first conditional probability and a second conditional probability corresponding to the word according to the statistical language model, wherein the first conditional probability is used for representing the conditional probability that the word appears on the right of the left context under the condition that the left context of the word is consistent with the emergency medical text, and the second conditional probability is used for representing the conditional probability that the word appears on the left of the right context under the condition that the right context of the word is consistent with the emergency medical text;

determining the error rate determined for the word through the statistical language model according to the first conditional probability and the second conditional probability, and taking the error rate as a first error rate corresponding to the word;

and determining the error position in the emergency medical text according to the first error rate corresponding to each word.

Optionally, determining, according to a preset wrong word recognition model, a wrong position existing in the emergency medical text specifically includes:

inputting the emergency medical text into a pre-trained wrong word recognition model to output a second error rate corresponding to each word in the emergency medical text through the wrong word recognition model;

And determining the error position in the emergency medical text according to the second error rate corresponding to each word.

Optionally, determining the candidate replacement word corresponding to each error position specifically includes:

each error word in the error word sequence formed by each error position is used as a word which does not independently form a word, and candidate replacement words corresponding to the error word at each error position are determined from a preset Chinese word list according to the pinyin similarity;

determining each wrong word from the wrong word sequence, and determining candidate replacement words corresponding to each wrong word from a preset Chinese word list according to the pinyin similarity for each wrong word, wherein one wrong word comprises at least two adjacent words in the first-aid medical text in the wrong word sequence;

and determining candidate replacement words corresponding to each wrong word according to the candidate replacement words corresponding to each wrong word.

Optionally, selecting a target text from the candidate error correction texts, which specifically includes:

for each candidate error correction text, splicing the candidate error correction text and the emergency medical text, and inputting the spliced candidate error correction text and the emergency medical text into a pre-trained error correction recognition model so that the error correction recognition model outputs the accuracy of the candidate error correction text;

And taking the candidate error correction text with the highest accuracy in the candidate error correction texts as the target text.

Optionally, before inputting the preset medical information type and the target text into the first network layer of the information extraction model, the method further includes:

performing first pre-training on a second network layer in the information extraction model through a natural language training sample;

and performing second pre-training on the second network layer through the medical language training sample, wherein the pre-training is to randomly replace part of characters in the training sample with preset characters to obtain a replacement text, input the replacement text into the second network layer, and perform pre-training on the second network layer by taking the replaced part of text output by the second network layer as a training target.

Optionally, inputting the prompt information vector and the target text to a second network layer of the information extraction model, so as to extract the medical information under the medical information type from the target text through the information extraction model, which specifically includes:

and inputting the spliced text obtained by splicing the target text, the prompt information vector and the preset characters into a second network layer of the information extraction model so as to extract medical information under the medical information type from the target text through the information extraction model.

Optionally, training the information extraction model specifically includes:

obtaining a training sample;

inputting the medical information type and the sample text in the training sample into a first network layer of an information extraction model so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the sample text into a second network layer of the information extraction model so that the second network layer outputs predicted information of the medical information type in the sample text;

and training the first network layer and the second network layer in the information extraction model by taking the difference between the minimized predicted information and the marked information in the training sample as a training target, wherein the marked information is used for representing actual information under the medical information type in the sample text.

The present specification provides a device for medical text correction and intelligent extraction of emergency scenes, comprising:

the receiving module is used for receiving an information extraction request and acquiring a voice-recognized emergency medical text through the information extraction request;

the error positioning module is used for determining the error position in the emergency medical text according to a preset statistical language model and/or an error word recognition model;

The error correction module is used for determining candidate replacement words corresponding to each error position, and determining each candidate error correction text after error correction of the emergency medical text according to the candidate replacement words corresponding to each error position so as to select a target text from each candidate error correction text;

the extraction module is used for inputting the preset medical information type and the target text into a first network layer of a pre-trained information extraction model so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the target text into a second network layer of the information extraction model so as to extract medical information under the medical information type from the target text through the information extraction model.

Optionally, the error positioning module is specifically configured to determine, for each word in the emergency medical text, according to the statistical language model, a first conditional probability and a second conditional probability corresponding to the word, where the first conditional probability is used to represent a conditional probability that the word appears to the right of the left context if the left context of the word is consistent with the emergency medical text, and the second conditional probability is used to represent a conditional probability that the word appears to the left of the right context if the right context of the word is consistent with the emergency medical text; determining the error rate determined for the word through the statistical language model according to the first conditional probability and the second conditional probability, and taking the error rate as a first error rate corresponding to the word; and determining the error position in the emergency medical text according to the first error rate corresponding to each word.

Optionally, the error positioning module is specifically configured to input the emergency medical text into a pre-trained word-staggering recognition model, so as to output, through the word-staggering recognition model, a second error rate corresponding to each word in the emergency medical text; and determining the error position in the emergency medical text according to the second error rate corresponding to each word.

Optionally, the error correction module is specifically configured to use each error word in the error word sequence formed by each error position as a word that does not form a word independently, and determine, according to the pinyin similarity, a candidate replacement word corresponding to the error word at each error position from a preset chinese word list; determining each wrong word from the wrong word sequence, and determining candidate replacement words corresponding to each wrong word from a preset Chinese word list according to the pinyin similarity for each wrong word, wherein one wrong word comprises at least two adjacent words in the first-aid medical text in the wrong word sequence; and determining candidate replacement words corresponding to each wrong word according to the candidate replacement words corresponding to each wrong word.

Optionally, the error correction module is specifically configured to splice, for each candidate error correction text, the candidate error correction text and the emergency medical text, and input the spliced candidate error correction text and the emergency medical text into a pre-trained error correction recognition model, so that the error correction recognition model outputs the accuracy of the candidate error correction text; and taking the candidate error correction text with the highest accuracy in the candidate error correction texts as the target text.

The present specification provides a computer readable storage medium storing a computer program which when executed by a processor implements the above-described medical text correction and intelligent extraction method for emergency situations.

The present specification provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the above-mentioned method for medical text correction and intelligent extraction of emergency situations when executing the program.

The above-mentioned at least one technical scheme that this specification adopted can reach following beneficial effect:

according to the method for correcting the medical text and intelligently extracting the medical text of the emergency scene, an information extraction request can be received, the voice-recognized emergency medical text is obtained through the information extraction request, then the error positions in the emergency medical text are determined according to a preset statistical language model and/or an error word recognition model, further, candidate replacement words corresponding to each error position are determined, each candidate correction text after correction of the emergency medical text is determined according to the candidate replacement words corresponding to each error position, so that a target text is selected from each candidate correction text, finally, the preset medical information type and the target text can be input into a first network layer of a pre-trained information extraction model, so that a prompt information vector is output by the first network layer, the prompt information vector and the target text are input into a second network layer of the information extraction model, and medical information under the medical information type is extracted from the target text through the information extraction model.

According to the method for intelligently extracting the medical text oriented to the emergency scene, errors in the emergency medical text can be positioned, so that the errors in the emergency medical text can be corrected in two different modes, a target text can be obtained, further, a prompt information vector integrating the target text and a medical information type of information required to be extracted can be determined through a first network layer in an information extraction model, finally, the prompt information vector and the target text are input into a second network layer, and medical information under the medical information type in the target text is determined.

Drawings

The accompanying drawings, which are included to provide a further understanding of the specification, illustrate and explain the exemplary embodiments of the present specification and their description, are not intended to limit the specification unduly. In the drawings:

Fig. 1 is a schematic flow chart of a method for medical text correction and intelligent extraction of emergency scenes provided in the present specification;

FIG. 2 is a schematic diagram of a training process for an information extraction model composed of a first network layer and a second network layer according to the present disclosure;

fig. 3 is a schematic flow chart of information extraction by an information extraction model formed by a first network layer and a second network layer provided in the present specification;

fig. 4 is a schematic diagram of a device for medical text correction and intelligent extraction of an emergency scene provided in the present specification;

fig. 5 is a schematic view of the electronic device corresponding to fig. 1 provided in the present specification.

Detailed Description

For the purposes of making the objects, technical solutions and advantages of the present specification more apparent, the technical solutions of the present specification will be clearly and completely described below with reference to specific embodiments of the present specification and corresponding drawings. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the present specification. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are intended to be within the scope of the present disclosure.

The following describes in detail the technical solutions provided by the embodiments of the present specification with reference to the accompanying drawings.

Fig. 1 is a flow chart of a method for medical text correction and intelligent extraction of emergency scenes provided in the present specification, specifically including the following steps:

s100: and receiving an information extraction request, and acquiring the voice-recognized emergency medical text through the information extraction request.

In practical applications, pre-hospital emergency situations are usually urgent, and medical staff needs to perform emergency treatment on patients, and also needs to input information related to the patients and medical treatment into a system, such as symptom information of the patients, basic information of height and weight, body part conditions, and the like.

In this case, the healthcare worker can enter patient medical-related information into the system of the hospital by means of voice input.

Based on the information, the server can receive the information extraction request and acquire the voice-recognized emergency medical text through the information extraction request, wherein the emergency medical text can be obtained by voice recognition of voice input by medical staff in a pre-hospital emergency scene through a terminal.

The main body for performing voice recognition may be a server or a terminal of a medical care person, and is not limited herein. In the above description, the execution subject of the method for performing the correction and intelligent extraction of the medical text of the emergency scene is described as a server, and the execution subject for performing the method is not limited herein, and may be a server, a desktop computer, or a large-sized service platform.

S102: and determining the error position in the emergency medical text according to a preset statistical language model and/or an error word recognition model.

After the emergency medical text is determined, the error position in the emergency medical text can be determined according to a preset statistical language model and/or an error word recognition model.

Wherein determining the location of the error present in the emergency medical text by the statistical language model may be understood as determining the probability of each word in the emergency medical text being erroneous by statistics. Specifically, for each word in the emergency medical text, a first conditional probability and a second conditional probability corresponding to the word can be determined according to a statistical language model.

Wherein the first conditional probability is used to represent the conditional probability that the word appears to the right of the left context if the left context of the word is consistent with the first-aid medical text and the second conditional probability is used to represent the conditional probability that the word appears to the left of the right context if the right context of the word is consistent with the first-aid medical text.

That is, the first conditional probability and the second conditional probability mentioned above may be conditional probabilities calculated from a large number of samples of natural language, the first conditional probability referring to a conditional probability that a word appears at a current position given a left few words, the second conditional probability referring to a conditional probability that a word appears at a current position given a right few words, and the number of words in the left context and the right context may be set according to actual requirements.

Then, the error rate determined for the word by the statistical language model may be determined according to the first conditional probability and the second conditional probability, and as the first error rate corresponding to the word, the following formula may be specifically used.

Wherein,for the first error rate, c represents an error word at an error location of the emergency medical document,for the first conditional probability, ++>For the second conditional probability, ++>Represents the left-hand context of c, +.>Representing the right context of c.

According to the first error rate corresponding to each word, the error position existing in the emergency medical text can be determined. For example, words having a first error rate above a certain threshold may be used as error words to determine the location of errors present in the emergency medical document. Of course, the first error rate may also be used in combination with a second error rate that is subsequently determined by the miscord recognition model.

The wrong word recognition model is a deep learning model, and the emergency medical text can be input into the pre-trained wrong word recognition model so as to output a second error rate corresponding to each word in the emergency medical text through the wrong word recognition model; and determining the error position in the emergency medical text according to the second error rate corresponding to each word.

The principle of the wrong word recognition model (deep learning model) is that the whole first-aid medical text is input into a deep neural network, and the neural network can be used for each layerFor each word in the text, a hidden state (hidden state, i.e. a hidden state vector) is calculated, e.g. the hidden state of word c is calculated from its hidden state of the previous layer together with the hidden states of the other words. Obtaining hidden state of each word in last layer by multi-layer iterative operation, such as wordThe hidden state at the last layer is noted asInputting the hidden state of the last layer into a two-class neural network, and calculating to obtain the probability (second error rate) of error of the word c based on the deep learning model as +.>. The deep learning model used in this embodiment may be a 12-layer self-focusing neural network transducer.

Likewise, the word with the second error rate higher than the certain threshold may be used as an error word, so as to determine the error position existing in the emergency medical text, and of course, the first error rate and the second error rate may be weighted and summed to obtain a comprehensive error rate, and the word with the comprehensive error rate higher than the certain threshold may be used as an error word, so as to determine the error position existing in the emergency medical text, as shown in the following formula.

Wherein,is the integrated error rate. />For the first error rate, +.>For the second error rate, +.>Weight corresponding to the first error rate, < ->And the weight corresponding to the second error rate.

S104: and determining candidate replacement words corresponding to each error position, and determining each candidate error correction text after error correction of the emergency medical texts according to the candidate replacement words corresponding to each error position so as to select a target text from each candidate error correction text.

And determining error positions in the emergency medical texts, namely, after error positioning is carried out on the emergency medical texts, determining candidate replacement words corresponding to each error position, and determining each candidate error correction text after error correction is carried out on the emergency medical texts according to the candidate replacement words corresponding to each error position so as to select a target text from the candidate error correction texts.

That is, for a wrong word at one wrong position, a plurality of candidate replacement words that can replace the wrong word can be determined, and by combining the candidate replacement words corresponding to the wrong word at each wrong position, a plurality of candidate correction texts after correction of the emergency medical text can be obtained, and the text that corrects the emergency medical text most accurately can be selected from the candidate correction texts as the target text.

The candidate replacement words corresponding to the error word can be selected in a pinyin similarity mode, and it is to be noted that a plurality of error positions can exist in the emergency medical text, that is, a plurality of error words exist, and the error words can be continuous or discontinuous, that is, one error word can exist between correct words, and a plurality of continuous error words can also exist between correct words. In the latter case, the candidate replacement word may be determined differently for a succession of erroneous words, one of which is considered to be an independent word, or to be a word concatenated with the other erroneous word.

Thus, in this application, all possibilities that one miscord is independent or connected to another miscord are listed. Therefore, each wrong word in the wrong word sequence formed by each wrong position can be used as a word which independently does not form a word, and candidate replacement words corresponding to the wrong word at each wrong position are determined from a preset Chinese word list according to the pinyin similarity; determining each wrong word from the wrong word sequence, and determining candidate replacement words corresponding to each wrong word from a preset Chinese word list according to the pinyin similarity, wherein one wrong word comprises at least two adjacent words in the first-aid medical text in the wrong word sequence; and determining candidate replacement words corresponding to each wrong word according to the candidate replacement words corresponding to each wrong word.

For arbitrary Chinese characters/wordsCan use +.>Representing its pinyin, e.g.)>，。

Emergency medical text for speech recognitionIs +.>The sequence of miscords located in this sentence is expressed as: />Wherein->Representing the number of words in the sentence, +.>Indicate->Error word->Indicate->The position of the individual miscords throughout the sentence. For error word->And words in Chinese vocabulary ++>Pinyin similarity +.A Pinyin similarity between Pinyin can be calculated based on the relative string distance between them>：

Wherein the method comprises the steps ofRepresenting the string distance between the two Pinyin, the Levenshtein distance used in this example is used to calculate the string distance,/I>And->The character string lengths of the two Pinyin respectively. Error correction candidates (candidate replacement words and candidate replacement words) are then generated according to the following procedure: (1) First consider all candidate replacement words for a single miscord, i.e. for +.>Calculating it and arbitrary word ++in Chinese vocabulary>Is used for determining the pinyin similarity of the Chinese characters,when their pinyin similarity exceeds a predetermined threshold, i.e. satisfies +.>At the time, will be->Marked as->Is a candidate replacement word for a word; (2) Correction candidates for a misword consisting of two adjacent miswords, i.e. for +.in a sequence of miswords, are considered >And->If->Indicating that the two miswords are adjacent, they are joined to a word +.>Calculating the misword and all words including two words in the Chinese vocabulary ++>Pinyin similarity +.>When the pinyin similarity meets the condition +.>At the time, will be->Marked as error word->Is a candidate replacement word for (a); (3) Consider a misword consisting of three adjacent miswords, i.e. for a sequence of miswords、/>And->If->Representing that the three miswords are adjacent, they are connected to a word +.>Then, candidate replacement words containing three characters are found in the Chinese vocabulary by the same method; (4) The above steps are repeated until there is no longer adjacent staggered word sequence. In this embodiment, the threshold value of Pinyin similarity may be taken +.>。

The above mentioned Chinese word list may be a preset Chinese word library, and each Chinese word or Chinese word in the Chinese word list may correspond to the pinyin corresponding to the Chinese word or Chinese word.

That is, each of the error characters can be used as an independent character to determine the character with similar pinyin similarity as the candidate replacement character, and then the candidate replacement characters corresponding to each of the error characters are found out from the error characters according to the two error characters capable of forming a word and the three error characters capable of forming a word in sequence. Then, each candidate error correction text can be obtained in a combined mode, and candidate replacement words in different candidate error correction texts can be different in a certain way.

After each candidate error correction text is determined, a target text can be selected from each candidate error correction text and used as the most accurate text obtained by correcting the emergency medical text.

Specifically, for each candidate error correction text, the candidate error correction text and the emergency medical texts are spliced and input into a pre-trained error correction recognition model, so that the error correction recognition model outputs the accuracy of the candidate error correction text, and then the candidate error correction text with the highest accuracy in the candidate error correction texts is used as a target text.

The candidate error correction text and the emergency medical text can be spliced to obtain a spliced text, the spliced text is input into an error correction recognition model, and the error correction recognition model can be a multi-layer deep learning language model to obtain a hidden state (hidden state vector) of the last layer，/>Is the original emergency medical text, ++>For the candidate error correction text, further carrying out regression by using the hidden state to obtain the candidate error correction text +.>Scoring->：

The highest scoring candidate error correction text may be used as the target text, and the error correction recognition model may be a 12-layer transducer model. The above W and b are parameters that need to be obtained by training.

In the above description, the emergency medical texts are regarded as a whole, each candidate error correction text is determined, and the target text is selected from the candidate error correction texts. Of course, the emergency medical text can be split into multiple sentences, each sentence determines a candidate error correction sentence corresponding to the sentence, a target sentence is selected from the candidate error correction sentences, and the target text can be determined according to the target sentence corresponding to each sentence.

S106: inputting a preset medical information type and the target text into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the target text into a second network layer of the information extraction model, so that medical information under the medical information type is extracted from the target text through the information extraction model.

After the target text is obtained, the preset medical information type and the target text can be input into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and the prompt information vector and the target text are input into a second network layer of the information extraction model, so that medical information under the medical information type is extracted from the target text through the information extraction model. The prompt information vector may be a vector in which information of a preset medical information type and target text are fused.

The second network layer mentioned above may be a generative language model, which requires pre-training so that the second network layer has some semantic understanding capability. The training for the second network layer may be two-stage, specifically, the first pre-training may be performed on the second network layer in the information extraction model through a natural language training sample, and then the second pre-training may be performed on the second network layer through a medical language training sample, where the pre-training refers to randomly replacing part of the characters in the training sample with preset characters to obtain a replacement text, inputting the replacement text into the second network layer, outputting the replaced part of the text by the second network layer as a training target, and performing the pre-training on the second network layer.

That is, the two pre-training steps mentioned above make the tasks performed by the second network layer identical, and cover a portion of the text in the complete text corresponding to the training sample, so that the second network layer predicts the portion of the text.

However, the training sample used in the first pre-training may be a natural language training sample, that is, may be a large amount of natural text in an unlimited field, and is mainly aimed at letting the language model learn the expression mode and the included knowledge of the natural language, and the training sample used in the second pre-training may be a medical language training sample, which may be derived from medical records and medical documents of a hospital, so as to further enhance the understanding ability of the second network layer on the medical expertise.

Specifically, the end can be set in sentences contained in the training sampleThe individual words are replaced with the character "[ MASK ]]"replace>May be a randomly generated MASK length, in which case the sentence becomes suffixed with "[ MASK ]]"New sentence of character (i.e., [ MASK ]]For the preset characters). The replaced sentences are input into a second network layer (a generative language model), and the training task is to generate the replaced texts. For example, the original sentence of the sample text in the training text is "patient is very elderly, feel left thigh pain suddenly", and cover length is taken randomly for it +.>The patient is changed into 'super-aged, sudden feeling of patient' after covering treatment [ MASK ]]The sentence after conversion is used as the input of a pre-trained generated language model, the training target of the language model is corresponding left thigh pain, and finally the language model is input through a maximum likelihood function +.>Optimizing the model, wherein +.>Representing covered text, < >>Representing the original text in the training sample, +.>Representative participation optimizationIs described.

When the medical information under the medical information type is extracted from the target text through the information extraction model, the target text, the prompt information vector determined through the first network layer and the preset characters can be spliced to obtain a spliced text, and the spliced text is input into the second network layer of the information extraction model so as to extract the medical information under the medical information type from the target text through the information extraction model.

The determination method of the prompt message vector can be as follows:

first, a medical information type such as "[ medicine information-medicine name" may be added in front of a piece of target text]By "a piece of text is obtained, of course, a plurality of medical information types can be simultaneously contained in the piece of text, and the type of one or more medical information which we wish to extract for the current target text is represented. Then using the embellishing layer of the T5 model to encode the input, and converting each word in the text segment with the target text spliced with the medical information type into a word vectorWherein->Representing the position of a word in the piece of text, then adjusting the importance of the key information according to the context, ignoring the unimportant information by means of the word vector +.>And word vectors of other words in the text of this paragraph +.>The term vector for each word in the piece of text is modified. The position in this text is +.>The words and positions of (2) are->Is of the word of (a)Similarity of word vectors->The method comprises the following steps:

wherein the method comprises the steps ofAnd->Are matrix parameters. The position is +.>Correction word vector of the word->The method comprises the following steps:

wherein the method comprises the steps ofSimilarity for word vector->The normalization parameter of (2) is calculated by

Wherein, Representing traversing each word in the piece of text one pass.

The corrected word vector of each character of the text is spliced together according to the position sequence to form a matrix O, and then the character of deeper level is extracted through nonlinear change once to obtain the prompt information vector：

,

In the above formulaAnd->For matrix parameters +.>And->The vector parameters are all obtained through training.

Examples of the medical information types are given below.

The above table one shows a secondary structure of medical information types, wherein the medical information types of the first hierarchy in the first column may include the medical information types of the second hierarchy, for example, symptoms, body parts, azimuth words, frequency, causes, descriptions, positive symptoms, and negative symptoms may be included under the symptom information.

When inputting the medical information type into the information extraction model, the medical information type of the first hierarchy may be input, or both the medical information type of the first hierarchy and the medical information type of the second hierarchy may be input.

It should be noted that, after the second network layer training is completed, the complete training of the information extraction model formed by the first network layer and the second network layer may be started. The training process of the information extraction model is a supervised training process, a training sample can be obtained, the medical information type and the sample text in the training sample are input into a first network layer of the information extraction model, so that the first network layer outputs a prompt information vector, the prompt information vector and the sample text are input into a second network layer of the information extraction model, so that the second network layer outputs prediction information of the medical information type in the sample text, further, the first network layer and the second network layer in the information extraction model are trained by taking the difference between the minimum prediction information and the labeling information in the training sample as a training target, and the labeling information is used for representing actual information of the medical information type in the sample text. The prediction information is represented by the second network layer.

As an example of a training process, a medical information type to be extracted is added to the forefront part of a section of sample text (emergency medical text), a first network layer of an information extraction model is input to obtain a prompt information vector, then the original sample text and the prompt information vector are spliced, and a "[ MASK ]" character (the preset character) is added to the tail part of the text, and finally, the learning target of fine tuning training is the information under the medical information type truly contained in the section of sample text. For example, sample text= "patient is very elderly, feeling left thigh pain suddenly", containing medical information: the body part is the "left thigh" and the symptoms are "pain".

Original emergency medical text and medical information type "[ symptom information-body part ]]"and" [ symptom information-symptom ]]"the prompt information vector is obtained by the first network layer through the joint calculationThen the original emergency medical text and +.>And the preset characters are spliced together, so that the patient is very aged, and the pain of the left thigh of the sudden sensation is: -the patient is relieved>[MASK]Inputting the second network layer (the generated language model) to perform fine tuning training, wherein the training aims at generating a result of 'left thigh pain', and simultaneously performing fine tuning training on the first network layer Is optimized.

By maximum likelihood functionOptimizing the model, wherein +.>Representing the target text (labeling information) "left thigh pain", ">Representing the entered splice text, < >>Representing the parameter set involved in the optimization, comprising the parameters of the first network layer and the second network layer, by means of a counter-propagating gradient descent>The parameters in the model are updated.

That is, the training process of the information extraction model composed of the first network layer and the second network layer and the process of information extraction by the information extraction model may be as shown in fig. 2 and 3.

Fig. 2 is a schematic flow chart of training an information extraction model formed by a first network layer and a second network layer provided in the present specification.

Fig. 3 is a schematic flow chart of information extraction by an information extraction model formed by a first network layer and a second network layer provided in the present specification.

After the server extracts the medical information under the medical information types required in the emergency medical texts, the medical information under the medical information types can be input into a hospital system, and the intelligent tool on the information platform can be used for providing auxiliary decisions for medical staff through the input medical information.

From the above, it can be seen that the present invention provides a method for error correction and intelligent extraction of medical text in emergency medical scenes, in which errors in emergency medical text can be located, so that errors in emergency medical text can be corrected by combining two different modes, and a target text can be obtained, further, a prompt information vector integrating the target text and a medical information type of information to be extracted can be determined through a first network layer in an information extraction model, and finally, the prompt information vector and the target text are input into a second network layer, and medical information in the medical information type in the target text can be determined.

Fig. 4 is a schematic diagram of a device for medical text correction and intelligent extraction of emergency scenes provided in the present specification, including:

a receiving module 401, configured to receive an information extraction request, and obtain a voice-recognized emergency medical text through the information extraction request;

An error positioning module 402, configured to determine an error location in the emergency medical text according to a preset statistical language model and/or a misword recognition model;

the error correction module 403 is configured to determine candidate replacement words corresponding to each error location, and determine each candidate error correction text after error correction is performed on the emergency medical text according to the candidate replacement words corresponding to each error location, so as to select a target text from the candidate error correction texts;

the extracting module 404 is configured to input a preset medical information type and the target text into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and input the prompt information vector and the target text into a second network layer of the information extraction model, so as to extract medical information under the medical information type from the target text through the information extraction model.

Optionally, the error localization module 402 is specifically configured to determine, for each word in the emergency medical text, according to the statistical language model, a first conditional probability and a second conditional probability corresponding to the word, where the first conditional probability is used to represent a conditional probability that the word appears to the right of the left context if the left context of the word is consistent with the emergency medical text, and the second conditional probability is used to represent a conditional probability that the word appears to the left of the right context if the right context of the word is consistent with the emergency medical text; determining the error rate determined for the word through the statistical language model according to the first conditional probability and the second conditional probability, and taking the error rate as a first error rate corresponding to the word; and determining the error position in the emergency medical text according to the first error rate corresponding to each word.

Optionally, the error positioning module 402 is specifically configured to input the emergency medical text into a pre-trained word-staggering recognition model, so as to output, through the word-staggering recognition model, a second error rate corresponding to each word in the emergency medical text; and determining the error position in the emergency medical text according to the second error rate corresponding to each word.

Optionally, the error correction module 403 is specifically configured to use each error word in the error word sequence formed by each error location as a word that does not form a word independently, and determine, according to the pinyin similarity, a candidate replacement word corresponding to the error word at each error location from a preset chinese word list; determining each wrong word from the wrong word sequence, and determining candidate replacement words corresponding to each wrong word from a preset Chinese word list according to the pinyin similarity for each wrong word, wherein one wrong word comprises at least two adjacent words in the first-aid medical text in the wrong word sequence; and determining candidate replacement words corresponding to each wrong word according to the candidate replacement words corresponding to each wrong word.

Optionally, the error correction module 403 is specifically configured to, for each candidate error correction text, splice the candidate error correction text with the emergency medical text, and input the spliced candidate error correction text and the emergency medical text into a pre-trained error correction recognition model, so that the error correction recognition model outputs the accuracy of the candidate error correction text; and taking the candidate error correction text with the highest accuracy in the candidate error correction texts as the target text.

Optionally, before inputting the preset medical information type and the target text into the first network layer of the information extraction model, the apparatus further includes:

the training module 405 is configured to perform a first pre-training on a second network layer in the information extraction model through a natural language training sample; and performing second pre-training on the second network layer through the medical language training sample, wherein the pre-training is to randomly replace part of characters in the training sample with preset characters to obtain a replacement text, input the replacement text into the second network layer, and perform pre-training on the second network layer by taking the replaced part of text output by the second network layer as a training target.

Optionally, the extracting module 404 is specifically configured to input, to the second network layer of the information extraction model, a spliced text obtained by splicing the target text, the prompt information vector, and the preset character, so as to extract, by using the information extraction model, medical information under the medical information type from the target text.

Optionally, the training module 405 is specifically configured to obtain a training sample; inputting the medical information type and the sample text in the training sample into a first network layer of an information extraction model so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the sample text into a second network layer of the information extraction model so that the second network layer outputs predicted information of the medical information type in the sample text; and training the first network layer and the second network layer in the information extraction model by taking the difference between the minimized predicted information and the marked information in the training sample as a training target, wherein the marked information is used for representing actual information under the medical information type in the sample text.

The present specification also provides a computer readable storage medium storing a computer program operable to perform the above method of medical text correction and intelligent extraction of emergency situations.

The present specification also provides a schematic structural diagram of the electronic device shown in fig. 5. At the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile storage, as illustrated in fig. 5, although other hardware required by other services may be included. The processor reads the corresponding computer program from the nonvolatile memory to the memory and then runs the computer program to realize the method for medical text correction and intelligent extraction of the emergency scene.

Of course, other implementations, such as logic devices or combinations of hardware and software, are not excluded from the present description, that is, the execution subject of the following processing flows is not limited to each logic unit, but may be hardware or logic devices.

In the 90 s of the 20 th century, improvements to one technology could clearly be distinguished as improvements in hardware (e.g., improvements to circuit structures such as diodes, transistors, switches, etc.) or software (improvements to the process flow). However, with the development of technology, many improvements of the current method flows can be regarded as direct improvements of hardware circuit structures. Designers almost always obtain corresponding hardware circuit structures by programming improved method flows into hardware circuits. Therefore, an improvement of a method flow cannot be said to be realized by a hardware entity module. For example, a programmable logic device (Programmable Logic Device, PLD) (e.g., field programmable gate array (Field Programmable Gate Array, FPGA)) is an integrated circuit whose logic function is determined by the programming of the device by a user. A designer programs to "integrate" a digital system onto a PLD without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Moreover, nowadays, instead of manually manufacturing integrated circuit chips, such programming is mostly implemented by using "logic compiler" software, which is similar to the software compiler used in program development and writing, and the original code before the compiling is also written in a specific programming language, which is called hardware description language (Hardware Description Language, HDL), but not just one of the hdds, but a plurality of kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), lava, lola, myHDL, PALASM, RHDL (Ruby Hardware Description Language), etc., VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog are currently most commonly used. It will also be apparent to those skilled in the art that a hardware circuit implementing the logic method flow can be readily obtained by merely slightly programming the method flow into an integrated circuit using several of the hardware description languages described above.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer readable medium storing computer readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable logic controllers, and embedded microcontrollers, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, atmel AT91SAM, microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic of the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller in a pure computer readable program code, it is well possible to implement the same functionality by logically programming the method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. Such a controller may thus be regarded as a kind of hardware component, and means for performing various functions included therein may also be regarded as structures within the hardware component. Or even means for achieving the various functions may be regarded as either software modules implementing the methods or structures within hardware components.

The system, apparatus, module or unit set forth in the above embodiments may be implemented in particular by a computer chip or entity, or by a product having a certain function. One typical implementation is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being functionally divided into various units, respectively. Of course, the functions of each element may be implemented in one or more software and/or hardware elements when implemented in the present specification.

It will be appreciated by those skilled in the art that embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the present specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present description can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present description is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the specification. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In one typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include volatile memory in a computer-readable medium, random Access Memory (RAM) and/or nonvolatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

Computer readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media for a computer include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, does not include transitory computer-readable media (transmission media), such as modulated data signals and carrier waves.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises the element.

It will be appreciated by those skilled in the art that embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the present specification may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present description can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, as relevant to see a section of the description of method embodiments.

The foregoing is merely exemplary of the present disclosure and is not intended to limit the disclosure. Various modifications and alterations to this specification will become apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, or the like, which are within the spirit and principles of the present description, are intended to be included within the scope of the claims of the present description.

Claims

1. A method for medical text correction and intelligent extraction of emergency scenes is characterized by comprising the following steps:

for each word in the emergency medical text, determining a first conditional probability and a second conditional probability corresponding to the word according to a statistical language model, wherein the first conditional probability is used for representing the conditional probability that the word appears on the right of the left context under the condition that the left context of the word is consistent with the emergency medical text, and the second conditional probability is used for representing the conditional probability that the word appears on the left of the right context under the condition that the right context of the word is consistent with the emergency medical text;

determining the error position in the emergency medical text by referring to the second error rate on the basis of the first error rate corresponding to each word;

2. The method of claim 1, wherein determining candidate replacement words for each error location comprises:

3. The method of claim 1, wherein selecting the target text from the candidate error correction texts, specifically comprises:

4. The method of claim 1, wherein prior to entering a predetermined one of the medical information types and the target text into the first network layer of the information extraction model, the method further comprises:

5. The method of claim 4, wherein inputting the hint information vector and the target text to a second network layer of the information extraction model to extract medical information of the one medical information type from the target text by the information extraction model, comprises:

6. The method according to claim 1 or 5, wherein training the information extraction model specifically comprises:

obtaining a training sample;

inputting one medical information type and sample text in the training sample into a first network layer of an information extraction model, so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the sample text into a second network layer of the information extraction model, so that the second network layer outputs predicted information of the one medical information type in the sample text;

7. The utility model provides a device that medical text correction and intelligence of first aid scene draw which characterized in that includes:

an error locating module, for each word in the emergency medical text, determining a first conditional probability and a second conditional probability corresponding to the word according to a statistical language model, wherein the first conditional probability is used for representing a conditional probability that the word appears right to the left context under the condition that the left context of the word is consistent with the first conditional probability in the emergency medical text, the second conditional probability is used for representing a conditional probability that the word appears left to the right context under the condition that the right context of the word is consistent with the first conditional probability in the emergency medical text, determining an error rate determined for the word according to the first conditional probability and the second conditional probability, inputting the first conditional probability as a first error rate corresponding to the word into a pre-trained error word recognition model, outputting a second error rate corresponding to each word in the emergency medical text through the word recognition model, and determining the existence of the position in the medical text by referring to the second error rate based on the first error rate corresponding to each word;

the extraction module is used for inputting a preset medical information type and the target text into a first network layer of a pre-trained information extraction model, so that the first network layer outputs a prompt information vector, and inputting the prompt information vector and the target text into a second network layer of the information extraction model, so that medical information under the medical information type is extracted from the target text through the information extraction model.

8. The apparatus of claim 7, wherein the error correction module is specifically configured to use each of the error words in the error word sequence formed by each of the error positions as a word that does not form a word independently, and determine, according to the pinyin similarity, a candidate replacement word corresponding to the error word at each of the error positions from a preset chinese word list; determining each wrong word from the wrong word sequence, and determining candidate replacement words corresponding to each wrong word from a preset Chinese word list according to the pinyin similarity for each wrong word, wherein one wrong word comprises at least two adjacent words in the first-aid medical text in the wrong word sequence; and determining candidate replacement words corresponding to each wrong word according to the candidate replacement words corresponding to each wrong word.

9. The apparatus of claim 7, wherein the error correction module is specifically configured to, for each candidate error correction text, concatenate the candidate error correction text with the emergency medical text and input the concatenated candidate error correction text into a pre-trained error correction recognition model, such that the error correction recognition model outputs an accuracy of the candidate error correction text; and taking the candidate error correction text with the highest accuracy in the candidate error correction texts as the target text.

10. A computer readable storage medium, characterized in that the storage medium stores a computer program which, when executed by a processor, implements the method of any of the preceding claims 1-6.

11. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the method of any of the preceding claims 1-6 when executing the program.