A method, system and medium for correcting errors in an Alzheimer's disease scale based on reading comprehension

Abstract

The present invention discloses a method, system and medium for correcting an Alzheimer's scale based on reading comprehension. The method steps include: 1) obtaining the audio of the tester on the Alzheimer's scale, and transcribing the tester's audio into a text to be corrected; 2) retrieving the neuropsychological scale text that matches the text to be corrected, and recording it as candidate characters; 3) combining the text to be corrected and the matching candidate characters into a text pair, and inputting them into a BERT pre-training model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale; 4) selecting the character with the highest probability from the candidate character list as the final output of the character. The system includes: an audio acquisition and conversion module, a candidate character matching module, a BERT processing module, and an error correction module; the present invention combines the language characteristics of the neuropsychological scale and proposes a proprietary error correction data generation method to make up for the problem of lack of training data.

Description

Technical Field

[0001] The invention relates to the field of Alzheimer's scale testing, and in particular to an error correction method, system and medium for an Alzheimer's scale based on reading comprehension. Background Art

[0002] Since neuropsychological tests are subjective, it is difficult to screen and diagnose dementia. Most general hospitals, community health centers, and nursing homes do not have the testing capabilities, which limits the correct diagnosis rate of Alzheimer's disease. Technological breakthroughs are urgently needed to solve the problems of long neuropsychological tests and subjective influences. Automatic speech recognition in artificial intelligence technology can solve the above problems in terms of test efficiency and accuracy, and is expected to promote neuropsychological scale tests in more communities.

[0003] Therefore, automatic speech recognition technology that can accurately identify Chongqing dialect will help promote early screening for dementia, especially Alzheimer's disease. Clinical studies have shown that quantifiable signs of cognitive decline associated with Alzheimer's disease and mild cognitive impairment are detectable in the patient's language. Machine learning models have been shown to successfully detect dementia using speech and language variables. For example, analyzing the content of sentences from speech-recognized text can infer the patient's health status. Since the transcript should be accurate enough to correctly represent syntactic and linguistic features, current methods generally rely on transcription by trained transcribers. But this will take a lot of time and effort, and at present any single speech recognition algorithm cannot guarantee the correct recognition of key information of patients in the process of answering neuropsychological scales. Chinese is a language based on syllables as the basic unit. We have also found differences in syllables between Mandarin and Chongqing dialect, mainly comparing the differences between initial consonants and finals. At the same time, some elderly people have slurred pronunciation, which puts higher requirements on the speech recognition model serving such users. A speech recognition model with a single model structure will find it difficult to capture the different pronunciations of characters. In the application of speech recognition in Chongqing dialect, its accuracy is very limited.

[0004] The accuracy of input data is the basic premise of common natural language processing tasks and the key to applying artificial intelligence technology in neuropsychological scale testing. In the mainstream solutions in the industry, the error detection technology based on LSTM+CRF is difficult to be widely implemented due to its reliance on a large number of labeled samples, and the error detection technology based on N-gram also has poor algorithm performance due to its "hard" judgment rules. Chinese error correction technology is an important technology for realizing automatic inspection and automatic error correction of Chinese sentences. Its purpose is to improve language correctness while reducing manual verification costs. As the most basic module of natural language processing, the importance of the error correction module is self-evident.

[0005] In the general field, the problem of Chinese text error correction is a problem that has been sought to be solved since the beginning of the Internet. In search engines, a good error correction system can provide error correction prompts for the query terms entered by users, or directly display the correct answer. Currently, due to the randomness of user input and the error-proneness of handwriting input methods, error strings account for as much as 10%-15% of the input text, which also highlights the necessity of the Chinese text error correction module. In vertical fields, such as the field of neuropsychological scales, due to the patient's education level, speaking habits and other reasons, there are a large number of near-implicit or homophonic errors in user questions and answers.

[0006] The mainstream error correction solutions in the industry generally include three parts: error detection → candidate recall → candidate ranking, or use the idea of ​​error correction methods based on machine translation models, analogize error correction tasks to machine translation tasks, and use models to translate erroneous sentences into correct sentences, and use Seq2seq models to complete the end-to-end correction process. However, these technologies are all based on error correction in open fields.

[0007] In summary, the speech recognition technology for neuropsychological scales needs to be optimized in terms of language model, as summarized as follows:

[0008] Statistics-based language models and language models in speech recognition are affected by the size of the corpus, have limited effectiveness, and statistical information has limited expressive power at the semantic level.

[0009] It is difficult to collect specific corpus for error correction in open fields. The text field of neural scales is relatively special, so it is relatively easy to build proprietary error correction training data.

[0010] 3) In neuropsychological scales, prior questions are very important, and the existing inventions have not yet utilized this information in error correction. Summary of the invention

[0011] The purpose of the present invention is to provide a method for correcting errors in an Alzheimer's scale based on reading comprehension, comprising the following steps:

[0012] 1) Obtain the tester's audio about the Alzheimer's scale and transcribe the tester's audio into text to be corrected;

[0013] 2) In the neuropsychological scale, the neuropsychological scale text matching the text to be corrected is retrieved and recorded as a candidate character;

[0014] 3) Establish a BERT pre-training model;

[0015] 4) Combine the text to be corrected and the matching candidate characters into a text pair and input it into the BERT pre-trained model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale;

[0016] 5) Select the character with the highest probability from the candidate character list as the final output of the character, and repeat steps 2) to 4) until the decoding encounters the end character.

[0017] Furthermore, the character x in the text to be corrected i The conditional probability Pc(yi=j|text) of being replaced by the jth character in the candidate character list is as follows:

[0018] P c (y i = j|Text) = softmax(W*h i +b)[j] (1)

[0019] Where W and b are weight and bias parameters, is the embedding result of input xi; softmat is the activation function; It is the last hidden state in the BERT pre-trained model.

[0020] Furthermore, the BERT pre-training model includes a multi-head attention layer, a forward propagation layer, and an encoder output layer.

[0021] Furthermore, the output of the multi-head attention layer MultiHead(Q, K, V) is as follows:

[0022] MultiHead(Q,K,V)=Concat(head1,...,head h )W O (2)

[0023] Where W 0 is the weight;

[0024] Among them, the parameter head i As shown below:

[0025] head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h (3)

[0026] Where W i Q , W i K , W i V Represents weight; Q = XW i Q , K = XWi K 、V=XW i V represents the query vector, key vector and value vector; h is the number of characters;

[0027] Among them, attention (Q, K, V) is as follows:

[0028]

[0029] Where, d k Represents the vector dimension.

[0030] Furthermore, the output FFN(x) of the forward propagation layer is as follows:

[0031] FFN(x)=max(0,xW1+b1)W2+b2 (5)

[0032] Where W1 and W2 represent weights; b1 and b2 represent biases;

[0033] The parameter x is as follows:

[0034] x=norm(X+MultiHead(Q,K,V)) (6)

[0035] Where X is the input.

[0036] Furthermore, the output Y of the encoder output layer is as follows:

[0037] Y=FFN(x) (7)

[0038] Where FFN(x) is the output of the forward propagation layer.

[0039] Furthermore, the input of the BERT pre-trained model is:

[0040] X = CLS + Question + SEP + Text + End; where End represents the end character; Text is the text to be corrected; SEP and CLS are separators and sentence vectors;

[0041] The text Question looks like this:

[0042] Question = {w1, w2, ..., w m}=encoder(E(w1),E(w2),...,E(w m )) (8)

[0043] In the formula, m represents the number of words in the question, encoder represents the encoder in BERT, and E represents the processing function for each character in the encoder. m For characters.

[0044] A system for applying an error correction method for the Alzheimer's scale based on reading comprehension, including an audio acquisition and conversion module, a candidate character matching module, a BERT processing module, and an error correction module;

[0045] The audio acquisition and conversion module acquires the tester's audio about the Alzheimer's scale, transcribes the tester's audio into text to be corrected, and transmits it to the candidate character matching module and the BERT processing module;

[0046] The candidate character matching module stores a neuropsychological scale;

[0047] The candidate character matching module retrieves the neuropsychological scale text that matches the text to be corrected from the neuropsychological scale, records it as a candidate character, and transmits it to the BERT processing module;

[0048] The BERT processing module stores a BERT pre-training model;

[0049] The BERT processing module combines the text to be corrected and the matching candidate characters into a text pair, and inputs it into the BERT pre-training model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale, and transmits it to the error correction module;

[0050] The error correction module selects the character with the highest probability from the candidate character list as the final output of the character until the decoding encounters the end character.

[0051] A computer-readable storage medium having a computer program stored thereon;

[0052] When the computer program is called, the steps of the above method are executed.

[0053] The technical effect of the present invention is unquestionable. The present invention proposes a segmented processing method for the speech recognition problem in the neuropsychological scale, that is, a separate error correction model is proposed to serve as a language model, thereby improving the accuracy rate.

[0054] The present invention proposes for the first time to use the question text in the neuropsychological scale as a priori condition and complete text error correction in combination with a reading comprehension model.

[0055] The present invention combines the language characteristics of the neuropsychological scale and proposes a proprietary error correction data generation method to make up for the problem of lack of training data. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flowchart for error correction;

[0057] Figure 2 It is the voice data processing flow;

[0058] Figure 3 For the model structure. DETAILED DESCRIPTION

[0059] The present invention is further described below in conjunction with the embodiments, but it should not be understood that the above subject matter of the present invention is limited to the following embodiments. Without departing from the above technical ideas of the present invention, various substitutions and changes are made according to the common technical knowledge and customary means in the art, which should all be included in the protection scope of the present invention.

[0060] Embodiment 1:

[0061] See also Figures 1 to 3 , a method for correcting errors in an Alzheimer's scale based on reading comprehension, comprising the following steps:

[0062] 1) Obtain the tester's audio about the Alzheimer's scale and transcribe the tester's audio into text to be corrected;

[0063] 2) In the neuropsychological scale, the neuropsychological scale text matching the text to be corrected is retrieved and recorded as a candidate character;

[0064] 3) Establish a BERT pre-training model;

[0065] 4) Combine the text to be corrected and the matching candidate characters into a text pair and input it into the BERT pre-trained model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale;

[0066] 5) Select the character with the highest probability from the candidate character list as the final output of the character, and repeat steps 2) to 4) until the decoding encounters the end character.

[0067] Character x in the text to be corrected i The conditional probability Pc(yi=j|text) of being replaced by the jth character in the candidate character list is as follows:

[0068] P c (y i = j|Text) = softmax(W*h i +b)[j] (1)

[0069] Where W and b are weight and bias parameters, is the embedding result of input xi; softmat is the activation function; It is the last hidden state in the BERT pre-trained model.

[0070] The BERT pre-training model includes a multi-head attention layer, a forward propagation layer, and an encoder output layer.

[0071] The output of the multi-head attention layer MultiHead(Q, K, V) is as follows:

[0072] MultiHead(Q,K,V)=Concat(head1,...,head h )W O (2)

[0073] Where W 0 is the weight;

[0074] Among them, the parameter head i As shown below:

[0075] head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h (3)

[0076] Where W i Q , W i K , W i V Represents weight; Q = XW i Q , K = XW i K 、V=XW i V represents the query vector, key vector and value vector; h is the number of characters;

[0077] Among them, attention (Q, K, V) is as follows:

[0078]

[0079] Where, d k Represents the vector dimension.

[0080] The output FFN(x) of the forward propagation layer is as follows:

[0081] FFN(x)=max(0,xW1+b1)W2+b2 (5)

[0082] Where W1 and W2 represent weights; b1 and b2 represent biases;

[0083] The parameter x is as follows:

[0084] x=norm(X+MultiHead(Q,K,V)) (6)

[0085] Where X is the input.

[0086] The output Y of the encoder output layer is as follows:

[0087] Y=FFN(x) (7)

[0088] Where FFN(x) is the output of the forward propagation layer.

[0089] The input of the BERT pre-trained model is:

[0090] X = CLS + Question + SEP + Text + End; where End represents the end character; Text is the text to be corrected; SEP and CLS are separators and sentence vectors;

[0091] The text Question looks like this:

[0092] Question = {w1, w2, ..., w m}=encoder(E(w1),E(w2),...,E(w m )) (8)

[0093] In the formula, m represents the number of words in the question, encoder represents the encoder in BERT, and E represents the processing function for each character in the encoder. m For characters.

[0094] Terminology explanation:

[0095] MRC: Machine Reading Comprehension;

[0096] ASR: Automatic Speech Recognition speech recognition technology;

[0097] NLP: Natural Language Processing;

[0098] BERT: pre-trained language model;

[0099] AI:Artificial Intelligence

[0100] LSTM+CRF: Long Short-Term Memory Network + Conditional Random Field;

[0101] Seq2seq: sequence to sequence;

[0102] NLG: Natural Language Generation.

[0103] Embodiment 2:

[0104] A system for applying an error correction method for the Alzheimer's scale based on reading comprehension, including an audio acquisition and conversion module, a candidate character matching module, a BERT processing module, and an error correction module;

[0105] The audio acquisition and conversion module acquires the tester's audio about the Alzheimer's scale, transcribes the tester's audio into text to be corrected, and transmits it to the candidate character matching module and the BERT processing module;

[0106] The candidate character matching module stores a neuropsychological scale;

[0107] The candidate character matching module retrieves the neuropsychological scale text that matches the text to be corrected from the neuropsychological scale, records it as a candidate character, and transmits it to the BERT processing module;

[0108] The BERT processing module stores a BERT pre-training model;

[0109] The BERT processing module combines the text to be corrected and the matching candidate characters into a text pair, and inputs it into the BERT pre-training model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale, and transmits it to the error correction module;

[0110] The error correction module selects the character with the highest probability from the candidate character list as the final output of the character until the decoding encounters the end character.

[0111] Embodiment 3:

[0112] A computer-readable storage medium having a computer program stored thereon;

[0113] When the computer program is called, the steps of the method described in Example 1 are executed.

[0114] Embodiment 4:

[0115] A method for correcting an Alzheimer's scale error based on reading comprehension comprises the following steps:

[0116] 1) Based on the audio, use the speech recognition model to output the text to be corrected:

[0117] 2) Based on the text of speech recognition, retrieve the matching text in the neuropsychological scale, combine the two into a text pair, and put it into the BERT pre-trained model.

[0118] The problem is dealt with as follows:

[0119] Question = {w1, w2, ..., w m}=encoder(E(w1),E(w2),...,E(w m ))

[0120] Where m represents the number of words contained in the question, encoder represents the encoder in BERT, and E represents the processing function for each character in the encoder.

[0121] The text to be corrected is processed in the same way as Question, and is finally input into BERT's X = CLS + Question + SEP + Text + End. The embedding is completed by introducing three matrices. The specific process is as follows: the input is converted into Q, K, V:

[0122] Q=XW Q ,K=XW K ,V=XW V

[0123] Attention calculation formula:

[0124]

[0125] Multi-head attention layer:

[0126] MultiHead(Q,K,V)=Concat(head1,...,head h )W O

[0127] in:

[0128] head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h

[0129] Forward propagation layer:

[0130] FFN(x)=max(0,xW1+b1)W2+b2

[0131] in:

[0132] x = norm(X + MultiHead(Q,K,V))

[0133] The output of the encoder is:

[0134] Y=FFN(x)

[0135] After the BERT pre-training model is used, for the character x in the text to be corrected, i The conditional probability of being replaced by the jth character in the candidate character list is:

[0136] P c (y i = j|Text) = softmax(W*h i +b)[j]

[0137] Among them, W and b are parameters, h i It is the embedding result of input X and the last hidden state in BERT After the residual link, we get: The last layer of the entire network uses a softmax function to select the character with the highest probability from the candidate character list as the final output of the character until the decoding encounters the end character. yi represents the character x in the text to be corrected i is replaced by the jth character in the candidate character list.

[0138] Embodiment 5:

[0139] A method for correcting an Alzheimer's scale error based on reading comprehension comprises the following steps:

[0140] 1) Obtain the tester's audio about the Alzheimer's scale, and transcribe the tester's audio into text to be corrected.

[0141] 2) In the neuropsychological scale, the neuropsychological scale text matching the text to be corrected is retrieved and recorded as a candidate character;

[0142] 3) Establish a BERT pre-training model;

[0143] 4) Combine the text to be corrected and the matching candidate characters into a text pair and input it into the BERT pre-trained model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale;

[0144] 5) Select the character with the highest probability from the candidate character list as the final output of the character, and repeat steps 2) to 4) until the decoding encounters the end character.

[0145] Embodiment 6:

[0146] A method for correcting errors in an Alzheimer's scale based on reading comprehension, the main contents of which are shown in Example 5, wherein the character x in the text to be corrected is i The conditional probability P of being replaced by the jth character in the candidate character list c (yi=j|text) is as follows:

[0147] P c (y i = j|Text) = softmax(W*h i +b)[j] (1)

[0148] Where W and b are weight and bias parameters, is the embedding result of input xi; softmat is the activation function; It is the last hidden state in the BERT pre-trained model.

[0149] Embodiment 7:

[0150] A method for correcting errors in an Alzheimer's scale based on reading comprehension, the main contents of which are shown in Example 5, wherein the BERT pre-training model includes a multi-head attention layer, a forward propagation layer, and an encoder output layer.

[0151] Embodiment 8:

[0152] A method for correcting an Alzheimer's scale based on reading comprehension, the main contents of which are shown in Example 5, wherein the output MultiHead(Q, K, V) of the multi-head attention layer is as follows:

[0153] MultiHead(Q,K,V)=Concat(head1,...,head h )W O (2)

[0154] Where W 0 is the weight;

[0155] Among them, the parameter head i As shown below:

[0156] head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h (3)

[0157] Where W iQ , W i K , W i V Represents weight; Q = XW i Q , K = XW i K 、V=XW i V represents the query vector, key vector and value vector; h is the number of characters;

[0158] Among them, attention (Q, K, V) is as follows:

[0159]

[0160] Where, d k Represents the vector dimension.

[0161] Embodiment 9:

[0162] A method for correcting an Alzheimer's scale error based on reading comprehension, the main contents of which are shown in Example 5, wherein the output FFN(x) of the forward propagation layer is as follows:

[0163] FFN(x)=max(0,xW1+b1)W2+b2 (5)

[0164] Where W1 and W2 represent weights; b1 and b2 represent biases;

[0165] The parameter x is as follows:

[0166] x=norm(X+MultiHead(Q,K,V)) (6)

[0167] Where X is the input.

[0168] Embodiment 10:

[0169] A method for correcting an Alzheimer's scale error based on reading comprehension, the main contents of which are shown in Example 5, wherein the output Y of the encoder output layer is as follows:

[0170] Y=FFN(x)(7)

[0171] Where FFN(x) is the output of the forward propagation layer.

[0172] Embodiment 11:

[0173] A method for correcting an Alzheimer's scale based on reading comprehension, the main contents of which are shown in Example 5, wherein the input X of the BERT pre-training model is CLS+Question+SEP+Text+End; wherein End represents the end character; Text is the text to be corrected; SEP and CLS are separators and sentence vectors;

[0174] The text Question looks like this:

[0175] Question = {w1, w2, ..., w m}=encoder(E(w1),E(w2),...,E(w m )) (8)

[0176] In the formula, m represents the number of words in the question, encoder represents the encoder in BERT, E represents the processing function for each character in the encoder; w m For characters.

Claims

1. A method for correcting errors in an Alzheimer's disease scale based on reading comprehension, characterized in that: The following steps are involved: 1) Obtain the tester's audio about the Alzheimer's scale and transcribe the tester's audio into text to be corrected; 2) In the neuropsychological scale, the neuropsychological scale text matching the text to be corrected is retrieved and recorded as a candidate character; 3) Establish a BERT pre-training model; 4) The text to be corrected and the matching candidate characters are combined into text pairs and input into the BERT pre-trained model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale; 5) Select the character with the highest probability from the neuropsychological scale as the final output of the character, and repeat steps 2) to 4) until the decoding encounters the End character; Character x in the text to be corrected i The conditional probability P of being replaced by the jth character in the neuropsychological scale c (y i =j|Text) is as follows: P c (y i <j|Text)=softmax(W*h i +b)[j](1) Where W and b are weight and bias parameters, is the input x i The embedding result; softmat is the activation function; It is the last hidden state in the BERT pre-training model; The BERT pre-training model includes a multi-head attention layer, a forward propagation layer, and an encoder output layer; The output of the multi-head attention layer MultiHead(Q, K, V) is as follows: MultiHead(Q,K,V)=Concat(head1,...,head h )W O (2) Where W 0 is the weight; Among them, the parameter head i As shown below: head i =Attention(QW i Q ,KW i K ,VW i V ),i=1,2,...,h (3) Where W i Q , W i K , W i V Represents weight; Q = XW i Q , K = XW i K 、V=XW i V represents the query vector, key vector and value vector; h is the number of characters; Among them, attention (Q, K, V) is as follows: Where, d k Represents the vector dimension; The input of the BERT pre-trained model is X=CLS+Question+SEP+Text+End; where End represents the end character; Text is the text to be corrected; SEP and CLS are separators and sentence vectors; The text Question looks like this: Question={w1,w2,...,w m }=encoder(E(w1),E(w2),...,E(w m )) (5) In the formula, m represents the number of words contained in the question, encoder represents the encoder in BERT, E represents the processing function for each character in the encoder; w m For characters.

2. The method for correcting an Alzheimer's scale based on reading comprehension according to claim 1, characterized in that: The output FFN(x) of the forward propagation layer is as follows: FFN(x)=max(0,xW1+b1)W2+b2(6) Where W1 and W2 represent weights; b1 and b2 represent biases; The parameter x is as follows: x=norm(X+MultiHead(Q,K,V))(7) Where X is the input.

3. The method for correcting the Alzheimer's scale based on reading comprehension according to claim 1, characterized in that: The output Y of the encoder output layer is as follows: Y=FFN(x)(8) Where FFN(x) is the output of the forward propagation layer.

4. A system using the method for correcting the error of the Alzheimer's scale based on reading comprehension as claimed in any one of claims 1 to 3, characterized in that: It includes audio acquisition and conversion module, candidate character matching module, BERT processing module and error correction module; The audio acquisition and conversion module acquires the tester's audio about the Alzheimer's scale, transcribes the tester's audio into text to be corrected, and transmits it to the candidate character matching module and the BERT processing module; The candidate character matching module stores a neuropsychological scale; The candidate character matching module retrieves the neuropsychological scale text that matches the text to be corrected from the neuropsychological scale, records it as a candidate character, and transmits it to the BERT processing module; The BERT processing module stores a BERT pre-training model; The BERT processing module combines the text to be corrected and the matching candidate characters into a text pair, and inputs the text into the BERT pre-training model to obtain the conditional probability that each character in the text to be corrected is replaced by the jth character in the neuropsychological scale, and transmits the conditional probability to the error correction module; The error correction module selects the character with the highest probability from the neuropsychological scale as the final output of the character until the decoding encounters the End character.

5. A computer-readable storage medium, characterized in that: A computer program is stored thereon; When the computer program is called, the method according to any one of claims 1 to 3 is executed.

Images