A Dialogue Summary Generation Method for Medical Consultation

Abstract

This invention discloses a method for generating dialogue summaries for medical consultations, relating to the field of natural language processing. Addressing the problem in existing technologies where invalid information in sentences leads to significant discrepancies between the generated doctor-patient dialogue summaries and the actual facts, resulting in low accuracy, this application establishes an extractive summarization component and a generative summarization component. First, several important sentences containing valid information are selected from the text. These are then input into a generative summarization model. The generative summarization model merges the sentences extracted in the extractive summarization component into shorter sentences containing more information. This preserves the valid factual information in the original text while generating fluent and concise sentences, thus improving the accuracy of the generated doctor-patient dialogue summaries. It solves the problems of significant discrepancies between the summarized results and the actual facts of the consultation dialogue, resulting in poor readability, and also helps doctors complete case summaries.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, specifically to a method for generating dialogue summaries for medical consultations. Background Technology

[0002] In recent years, with the widespread use of computers and smartphones, the number of internet users has increased dramatically, leading to an explosive growth in online data. Online medical services have become a popular online service in recent years, and the huge number of internet users has provided a vast market for their development. More and more people are now enjoying online medical services, able to communicate face-to-face with doctors and resolve health concerns without leaving home. Medical consultation dialogue summaries have become an urgent need and a problem of great research value. Automatic dialogue summaries offer an efficient solution. However, existing technologies suffer from low accuracy due to invalid information in sentences, resulting in significant discrepancies between the generated doctor-patient dialogue summaries and the facts. Summary of the Invention

[0003] The purpose of this invention is to address the problem in existing technologies where invalid information in sentences leads to significant discrepancies in the facts and low accuracy of the generated doctor-patient dialogue summaries. This invention proposes a dialogue summary generation method for medical consultations.

[0004] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0005] A method for generating dialogue summaries for medical consultations includes the following steps:

[0006] Step 1: Obtain the original medical consultation dialogue data, and simultaneously obtain the dialogue content summary already summarized by the doctor or doctor's assistant, in order to construct text data;

[0007] Step 2: Use the dialogue between doctors and patients in the text data as features and the summary of the dialogue content summarized by doctors or doctor assistants as labels to obtain labeled data;

[0008] The text data includes dialogues between doctors and patients, including medical history, description of the condition, symptoms, doctor's consultation and advice, and the doctor's diagnosis and reasonable treatment plan;

[0009] Step 3: Train a BiLSTM neural network using the labeled data. The BiLSTM neural network includes an extractive summarization network, a generative summarization network, and a pointer generator network. The generative summarization network includes an encoder and a decoder.

[0010] The neural network first segments the labeled data into words and encodes the segmentation results to obtain multiple word vectors. Then, the extractive digest filters the word vectors to retain word vectors that contain effective factual information.

[0011] The encoder is used to extract features from the sentences retained by the extractive summary, and then concatenate the extracted features to obtain the final feature representation.

[0012] The decoder is used to decode the final feature representation to obtain the probability distribution;

[0013] The pointer generator network is used to process the probability distribution by combining the copying mechanism and the covering mechanism to obtain the final distribution;

[0014] Step 4: Input the medical consultation dialogue data to be identified into the trained BiLSTM neural network, and retain the medical consultation dialogue summary with the highest probability according to the final distribution.

[0015] Furthermore, the original doctor-patient dialogue data in step one was obtained from the Internet by crawling the urllib request library.

[0016] Furthermore, the specific steps of the neural network first segmenting the labeled data into words and then encoding the segmentation results are as follows:

[0017] For the labeled data, each sentence is processed with [CLS] as the beginning and [SEP] as the end. The processed sentences are then converted into IDs using a word segmentation dictionary. The length of the input model is set to 512. For sentences longer than the input length, the first 512 IDs are retained. For sentences shorter than the input length, [PAD] is used to pad them. Finally, the converted IDs are input into the pre-trained model to obtain the encoded information, i.e., word vectors.

[0018] Furthermore, the extractive summarization performs information filtering on word vectors using the TextRank algorithm, which is expressed as:

[0019]

[0020] Among them, WS(v i ) indicates the statement v i The final score, where d represents the damping coefficient, with an initial value of 0.85, w ji w represents the similarity value between sentences j and i. jk v represents the similarity value between sentences j and k. k Indicates that except for sentence v i All sentences other than WS(v) j ) indicates the statement vj The final score.

[0021] Furthermore, the encoder is composed of a bidirectional long short-term memory network, and the hidden layer of the bidirectional long short-term memory network includes three gating structures and a hidden state a. t The hidden state includes the forget gate f t Input gate i t and output gate o t ;

[0022] The hidden layers of a bidirectional long short-term memory network are represented as follows:

[0023] f t =σ(W f ·[h t-1 ,x t ]+b f )

[0024] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0025] a t =tanh(W a ·[h t-1 ,x t ]+b a )

[0026] o t =σ(W o ·[h t-1 ,x t ]+b o )

[0027] Where, x t h represents the input at time t. t-1 W represents the hidden state value at time t-1; f W i W o W a These are the forget gate, input gate, output gate, and h during feature extraction. t-1 Weighting coefficients; b f b i b o b a These represent the bias values ​​for the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, tanh is expressed as:

[0028]

[0029] σ represents the Sigmoid activation function, and σ is expressed as:

[0030]

[0031] The forget gate and input gate are used to calculate the state c(t) at time t, which is expressed as:

[0032] c(t)=c(t-1)⊙f(t)+i(t)⊙a(t)

[0033] Where ⊙ represents the Handamard product;

[0034] The hidden state h(t) at time t is determined by the output gate o. t and the current state c t Find it, and represent it as:

[0035] h(t)=o(t)⊙tanh(c(t)).

[0036] Furthermore, the decoder employs an attention mechanism, by analyzing the input state s in each decoding step. t Hidden state h i The input contribution value a is obtained by weighting and summing. t The input contribution value a t Represented as:

[0037] a t =softmax(v T tanh(W h h i +W s s t ))

[0038] Among them, v and W h and W s It is possible to learn parameters;

[0039] Based on the input contribution value a t and hidden state h i Calculate the weighted output vector Represented as:

[0040]

[0041] in, The contribution value for the i-th sentence;

[0042] Using weighted output vectors and input state s t The probability distribution of the output words at time t is obtained by the following formula:

[0043]

[0044] Where V' and V are learnable parameters, y t This is the decoder output at time t.

[0045] Furthermore, the pointer generator network combines copying and covering mechanisms to process the probability distribution, and the specific steps to obtain the final distribution are as follows:

[0046] The pointer generator network calculates the generation probability p during decoding at time t. gen , represented as:

[0047]

[0048] Among them, y t-1 This represents the decoder output at time t-1. w s and w y These are learnable parameters;

[0049] The contribution distribution of copied words is represented as follows:

[0050]

[0051] Among them, P vocab (y t ) represents the probability distribution of generated words;

[0052] The sum of the attention distributions from previous time steps of the decoder yields the coverage vector c. t , represented as:

[0053]

[0054] Among them, c t This indicates the degree of coverage of these words by the attention mechanism up to time t;

[0055] The covering vector is used in the attention mechanism to represent the input contribution value a. t The formula is improved as follows:

[0056]

[0057] Among them, w c and b attn This allows us to learn parameters.

[0058] Furthermore, during the training process of the BiLSTM neural network, a beam search method is used for decoding, retaining only 5 candidate results each time to generate a medical consultation dialogue summary. This summary is then compared with the input summary, and the negative log-likelihood loss function is calculated. All parameters in the model are updated through backpropagation. Based on minimizing the loss as the training objective, the Adam optimizer is used to repeat the training process. The model obtained throughout the training process and its corresponding model parameters are saved, and hyperparameters are set simultaneously.

[0059] The hyperparameters are set as follows: training epochs are set to 30, tolerance is set to 50, training batch size is set to 4, learning rate is 1e-5, and BiLSTM hidden layer dimension is set to 150.

[0060] The beneficial effects of this invention are:

[0061] This application establishes an extractive summarization component and a generative summarization component. First, several important sentences containing valid information are selected from the text. These are then input into the generative summarization model. The generative summarization model merges the sentences extracted in the extractive summarization component into shorter sentences containing more information. This preserves the valid factual information in the original text while generating fluent and concise sentences, thus improving the accuracy of generating doctor-patient dialogue summaries.

[0062] It can solve the problems of existing summary generation methods, such as large discrepancies between the summary results and the actual consultation dialogue, and poor readability. At the same time, it helps doctors complete case summaries. Attached Figure Description

[0063] Figure 1 This is a flowchart of the application;

[0064] Figure 2 This is a schematic diagram of the TextRank algorithm extraction.

[0065] Figure 3 This is a schematic diagram of a generative summarization model. Detailed Implementation

[0066] It should be noted that, where there is no conflict, the various embodiments disclosed in this application can be combined with each other.

[0067] Specific implementation method one: Refer to Figure 1 This embodiment describes a method for generating dialogue summaries for medical consultations, comprising:

[0068] Step S1. Obtain raw training data via web crawling. Use the urllib request library to crawl medical consultation dialogue data from the Internet, and simultaneously obtain dialogue summaries already compiled by doctors or doctor assistants. Then construct a medical consultation dialogue corpus.

[0069] Step S2. Divide the text dataset into training dataset, test dataset and validation dataset, preprocess the training dataset, test dataset and validation dataset to generate training samples, test samples and validation samples.

[0070] Step S3. Construct a medical consultation dialogue summarization model. This model consists of an extraction part and a generative part. The extraction part uses the TextRank summarization algorithm to extract dialogue sentences containing useful information from the doctor-patient conversation. The generative part consists of an encoder and a decoder. The encoder uses a bidirectional long short-term memory network, while the decoder uses a long short-term memory network incorporating a pointer generation network and an attention mechanism, enabling it to generate not only words from the vocabulary but also words copied from the input sequence.

[0071] Step S4. Train the dialogue summarization generation model using training samples, employ the Adam optimizer as the optimization method, and update the dialogue summarization generation model using the negative log-likelihood loss function.

[0072] Step S5. Evaluate the updated dialogue summarization model using validation samples to obtain the medical consultation dialogue summarization generation model with the best evaluation results; test the optimized dialogue summarization generation model using test samples;

[0073] Step S6. Input a set of doctor-patient dialogue texts into the model to obtain a summary of the dialogue content.

[0074] In steps S1 and S2, the original doctor-patient dialogue text data obtained through web crawling and the dialogue content summary summarized by the doctor or doctor's assistant are used as data labels. A medical consultation dialogue corpus is constructed, and the dataset is divided into a training dataset, a test dataset, and a validation dataset.

[0075] In steps S3 to S5, the dataset obtained in step S1 is first partially extracted using the TextRank algorithm to extract sentences containing valid information, which are then fed into the generation model for training. The Adam optimizer is used as the optimization method, and the negative log-likelihood loss function is used to update the dialogue summarization generation model. Finally, the model with the best results is selected through testing and retained as the dialogue summarization generation model.

[0076] In step S6, model prediction is performed on the model representation from steps S3 to S5.

[0077] The following will explain each point:

[0078] (1) Based on the request library in urllib, medical consultation dialogue data was crawled from the Internet, and the dialogue content summaries already summarized by doctors or doctor assistants were obtained simultaneously to construct a medical consultation dialogue corpus. A total of about 1.8 million dialogue data were obtained. The content of each dialogue is roughly as follows: the patient first describes his / her medical history, condition, symptoms, etc., consults the doctor for advice, and then the doctor gives a diagnosis and reasonable treatment plan.

[0079] (2) After obtaining the data, it is preprocessed, using the dialogue between the doctor and patient as features and the summary of the dialogue content summarized by the doctor or doctor's assistant as labels. The labeled dataset is then obtained.

[0080] For each sentence, it is processed into a form like "[CLS]sentence[SEP]", and the maximum length of the input model is set to 512. For sentences shorter than the input length, [PAD] tags are used to complete them. The data is then divided into training dataset, test dataset, and validation dataset.

[0081] (3) First, the data to be used for training is segmented into words. After the sentences are segmented, word embedding is used to convert the words of natural language into word vectors that can be recognized by machines. The extraction part of the model uses the TextRank algorithm to extract invalid sentences in the dialogue. The TextRank algorithm is based on the PageRank algorithm and can be used to calculate keywords and key sentences. The TextRank algorithm treats each sentence in the text as a node and the similarity between sentences as the edges of the nodes. Because the similarity between sentences is mutual, the edges between sentences are undirected, thus forming a weighted undirected graph. The formula for calculating TextRank is:

[0082]

[0083] Among them, WS(v i ) indicates the statement v i The final score, where d is the damping coefficient, typically with an initial value of 0.85, and w ji This represents the similarity value between sentences j and i.

[0084] The generative part consists of an encoder and a decoder. The encoder is a bidirectional long short-term memory network, which is composed of two long short-term memory networks in different directions. Its main structure is similar to a traditional recurrent neural network. The main improvement is the addition of three gating structures in the hidden layers: a forget gate, an input gate, and an output gate, along with an additional hidden state. Its formal definition is:

[0085] f t =σ(W f ·[h t-1 ,x t ]+b f )#(2)

[0086] i t =σ(W i ·[h t-1 ,x t ]+b i )#(3)

[0087] a t =tanh(W a ·[h t-1 ,x t ]+b a )#(4)

[0088] o t =σ(W o ·[h t-1 ,x t ]+b o )#(5)

[0089] Where, x t h represents the input at time t. t-1 W represents the hidden state value at time t-1; f W i W o W a These are the forget gate, input gate, output gate, and h during feature extraction. t-1 Weighting coefficients; b f b i b o b a These represent the bias values ​​for the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, and its formula is as follows:

[0090]

[0091] σ represents the Sigmoid activation function, and its formula is as follows:

[0092]

[0093] The forget gate and input gate are used to calculate the state c(t) at time t, and their formulas are as follows:

[0094] c(t)=c(t-1)⊙f(t)+i(t)⊙a(t)#(8)

[0095] Where ⊙ represents the Handamard product. Finally, the hidden state h(y) at time t is determined by the output gate o. t and the current state c t The formula is:

[0096] h(t)=o(t)⊙tanh(c(t))#(9)

[0097] Bidirectional Long Short-Term Memory (LSTM) networks consist of two independent LSM networks. The input sequence is fed into both LSM networks in both forward and reverse order for feature extraction. The feature vectors extracted by the two networks are then concatenated to form the final feature representation. Due to its ability to extract features simultaneously in both forward and backward directions, the bidirectional LSM network possesses both forward and backward information. This network model has proven to be significantly more efficient and performant in extracting text features than single LSM network structures.

[0098] Since using only a Long Short-Term Memory (LSTM) network to decode and generate words may lead to inconsistencies between the generated summary and the facts, an attention mechanism is employed to improve the decoding and summary generation. This mechanism works by monitoring the input state s at each decoding step. t Hidden state h i The input contribution value a is obtained by weighting and summing. t The formula is as follows:

[0099] a t =softmax(v T tanh(W h h i +W s s t ))#(10)

[0100] Among them, v and W h and W s It is possible to learn parameters. The weighted output vector is calculated using the following formula:

[0101]

[0102] Finally, the probability distribution of the output words is obtained at time t, and its formula is:

[0103]

[0104] Where V' and V are learnable parameters.

[0105] The pointer generator network allows both copying words using pointers and generating words from a fixed vocabulary. The pointer generator network calculates the generation probability p during decoding at time t. genThis probability determines the probability of generating a word from the vocabulary or the probability of directly copying a word from the input sequence (1-p). gen The formula for calculating the generation probability is:

[0106]

[0107] Where y t-1 This represents the decoder output at time t-1. w s and w y For learnable parameters, the attention distribution in formula (10) is reused for the contribution distribution of copied words. The final output probability distribution is:

[0108]

[0109] Where P vocab (y t The same as formula (12) is used. A covering mechanism is also employed to address the shortcomings of traditional sequence-to-sequence models in generating repeated words, by adding a covering vector c. t Its value is the sum of the attention distributions of the decoder over past time steps, and its formula is:

[0110]

[0111] c t This represents the coverage of these words received by the attention mechanism up to time t. The coverage vector is used in the attention mechanism, and formula (10) is improved as follows:

[0112]

[0113] Where w c and b attn In order to learn parameters, this ensures that the attention mechanism can more easily avoid repeatedly focusing on the same position, thereby avoiding the generation of duplicate text.

[0114] (4) Based on the constructed dataset and neural network structure, the Adam optimizer and negative log-likelihood loss function are used to train the model. A large-scale medical consultation dialogue and its summary are input to train the constructed neural network. A beam search method is used for decoding, retaining only a specific number of candidate results each time to generate a medical consultation dialogue summary. This summary is then compared with the input summary, and the negative log-likelihood loss function is calculated. All parameters in the model are updated through backpropagation. Based on minimizing the loss as the training objective, the Adam optimizer is used to repeat the training process, saving the model and its corresponding parameters obtained throughout the training process. Hyperparameters are set as follows: training epochs are set to 30, tolerance is set to 50 (training stops when the metric on the validation set no longer increases after 50 iterations to avoid overfitting), training batch size is set to 4, learning rate is 1e-5, and the BiLSTM hidden layer dimension is set to 150.

[0115] (5) The evaluation is conducted using the ROUGE value, with higher scores indicating better dialogue summarization. In addition, based on the algorithm evaluation, a group of medical experts conducts manual review and evaluation of the automatic results based on the content of the medical consultation dialogue, and finally retains the medical consultation dialogue summary with the highest output accuracy.

[0116] (6) Based on a word embedding dictionary and a trained automatic medical consultation dialogue summarization model, intelligent medical consultation dialogue summarization is achieved, and the required summary is output for newly input online medical consultation dialogues. The specific process is as follows: input dialogue text, perform data preprocessing on the text, segment the text based on the dictionary, use the trained model to predict the summary of the segmented text, and output the medical consultation dialogue summary.

[0117] It should be noted that the specific embodiments are merely explanations and illustrations of the technical solution of the present invention and should not be used to limit the scope of protection. Any modifications made in accordance with the claims and specification of the present invention that are only partial should still fall within the protection scope of the present invention.

Claims

1. A method for generating dialogue summaries for medical consultations, characterized in that... Includes the following steps: Step 1: Obtain the original medical consultation dialogue data, and simultaneously obtain the dialogue content summary already summarized by the doctor or doctor's assistant, in order to construct text data; Step 2: Use the dialogue between doctors and patients in the text data as features and the summary of the dialogue content summarized by doctors or doctor assistants as labels to obtain labeled data; The text data includes dialogues between doctors and patients, including medical history, description of the condition, symptoms, doctor's consultation and advice, and the doctor's diagnosis and reasonable treatment plan; Step 3: Train a BiLSTM neural network using the labeled data. The BiLSTM neural network includes an extractive summarization network, a generative summarization network, and a pointer generator network. The generative summarization network includes an encoder and a decoder. The neural network first segments the labeled data into words and encodes the segmentation results to obtain multiple word vectors. Then, the extractive digest filters the word vectors to retain word vectors that contain effective factual information. The encoder is used to extract features from the sentences retained by the extractive summary, and then concatenate the extracted features to obtain the final feature representation. The decoder is used to decode the final feature representation to obtain the probability distribution; The pointer generator network is used to process the probability distribution by combining the copying mechanism and the covering mechanism to obtain the final distribution; Step 4: Input the medical consultation dialogue data to be identified into the trained BiLSTM neural network, and retain the medical consultation dialogue summary with the highest probability according to the final distribution. The pointer generator network combines replication and overlay mechanisms to process the probability distribution, and the specific steps to obtain the final distribution are as follows: The pointer generator network calculates the generation probability during decoding at time t. , is represented as: in, This represents the decoder output at time t-1. , and For learning parameters, The weighted output vector, For input status, This represents the Sigmoid activation function; The contribution distribution of copied words is represented as follows: in, Represents the probability distribution of generated words. Let i be the contribution value of the i-th sentence. This is the decoder output at time t; The sum of the attention distributions at previous time steps of the decoder yields the coverage vector. , is represented as: in, This indicates the degree of coverage of these words by the attention mechanism up to time t; Covering vectors are used in attention mechanisms to represent the contribution values ​​of the input. The formula is improved as follows: in, and For learning parameters, , and These are the learning parameters, where tanh represents the tangent hyperbolic function. It is in a hidden state.

2. The method for generating dialogue summaries for medical consultation according to claim 1, characterized in that... In step one, the original doctor-patient dialogue data was obtained from the Internet using the urllib request library.

3. The method for generating dialogue summaries for medical consultation according to claim 2, characterized in that... The neural network first segments the labeled data into words, and then encodes the segmentation results as follows: For the labeled data, each sentence is processed with [CLS] as the beginning and [SEP] as the end. The processed sentences are then converted into IDs using a word segmentation dictionary. The length of the input model is set to 512. For sentences longer than the input length, the first 512 IDs are retained. For sentences shorter than the input length, [PAD] is used to pad them. Finally, the converted IDs are input into the pre-trained model to obtain the encoded information, i.e., word vectors.

4. The method for generating dialogue summaries for medical consultation according to claim 3, characterized in that... The extractive summarization performs information filtering on word vectors using the TextRank algorithm, which is expressed as follows: in, Statement The final score, where d represents the damping coefficient, with an initial value of 0.

85. This represents the similarity value between sentences j and i. This represents the similarity value between sentences j and k. Indicates except sentence All sentences other than Statement The final score.

5. A method for generating dialogue summaries for medical consultation according to claim 4, characterized in that... The encoder is composed of a bidirectional long short-term memory network, and the hidden layer of the bidirectional long short-term memory network includes three gating structures and one hidden state. The hidden state includes a forget gate. Input gate and output gate ; The hidden layers of a bidirectional long short-term memory network are represented as follows: in, This represents the input at time t. This represents the hidden state value at time t-1; , , , These are the forget gate, input gate, output gate, and feature extraction process. Weighting coefficients; , , , These represent the bias values ​​for the forget gate, input gate, output gate, and feature extraction process, respectively; tanh represents the tangent hyperbolic function, tanh is expressed as: This represents the Sigmoid activation function. Represented as: The forget gate and input gate are used to calculate the state at time t. , Represented as: in For Handamard products; Hidden state at time t From the output gate and the current state Find it, and represent it as: 。 6. A method for generating dialogue summaries for medical consultation according to claim 5, characterized in that... The decoder employs an attention mechanism, which analyzes the input state at each decoding step. Hidden state The weighted sum is used to obtain the contribution value of the input. Input contribution value Represented as: in, , and These are the learning parameters; Based on the input contribution value and hidden state Calculate the weighted output vector , is represented as: in, The contribution value for the i-th sentence; Using weighted output vectors and input status The probability distribution of the output words at time t is obtained by the following formula: in, and For learning parameters, This is the decoder output at time t.

7. A method for generating dialogue summaries for medical consultation according to claim 1, characterized in that... The BiLSTM neural network is trained using a beam search approach for decoding, retaining only 5 candidate results at a time to generate a medical consultation dialogue summary. This summary is then compared with the input summary, and a negative log-likelihood loss function is calculated. All parameters in the model are updated through backpropagation. Based on minimizing the loss as the training objective, the Adam optimizer is used to repeat the training process. The model and its corresponding parameters obtained during the entire training process are saved, and hyperparameters are set simultaneously. The hyperparameters are set as follows: training epochs are set to 30, tolerance is set to 50, training batch size is set to 4, learning rate is 1e-5, and BiLSTM hidden layer dimension is set to 150.

Images