Data augmentation method based on abstract sentence and PICOS extraction classification method
By using automatic data augmentation and sentence concatenation methods, combined with sentence position information, a PICOS extraction and classification model was constructed. This solved the problem of insufficient data in sentence classification of medical literature abstracts and improved the accuracy of sentence recognition in the model.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HANGZHOU HUIYIDAO TECH CO LTD
- Filing Date
- 2022-09-14
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies for classifying sentences in medical literature abstracts suffer from insufficient data and difficulties in manual annotation, resulting in inadequate training data for the model and affecting classification accuracy.
By using an automatic data augmentation method, sentences from the same medical article abstract are randomly concatenated to generate new data samples. Combined with sentence position information, a PICOS extraction and classification model is constructed. The BERT pre-trained model and the concatenation module are used to improve the model's sentence recognition ability.
It significantly improved the model's accuracy in recognizing long sentences, solved the problems of limited data and difficulty in manual annotation, and improved the performance of PICOS extraction and classification.
Smart Images
Figure CN115758230B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of PICOS extraction and classification technology in evidence-based medicine, and specifically relates to a data augmentation method based on article summary sentences and a PICOS extraction and classification method. Background Technology
[0002] Currently, most medical literature is compiled according to the PICOS framework and principles. For the PICOS classification task in medical article abstracts, deep learning from artificial intelligence can be considered to solve large-scale text classification problems. This can be achieved through a pre-training-fine-tuning-downstream task approach, eliminating cumbersome manual feature engineering and providing an end-to-end solution.
[0003] However, when classifying article abstract sentences, there is a problem of insufficient data. It is necessary to annotate the article abstract sentences, but the current manual annotation method is difficult and complicated, and professional medical workers need to read through the abstract before they can annotate. Summary of the Invention
[0004] One of the objectives of this invention is to propose a data augmentation method based on article summary sentences, which solves the current technical problems of difficult data annotation and small data volume through automatic data augmentation.
[0005] To achieve the above objectives, the present invention adopts the following technical solution:
[0006] A data augmentation method based on article summary sentences includes the following steps:
[0007] Step 1. Data Sample Construction;
[0008] First, the abstracts of multiple medical articles are segmented into sentences. Then, each sentence after segmentation is categorized and its position within the abstract of the corresponding medical article is marked, i.e., its order information.
[0009] Each sentence is combined with its corresponding medical article abstract number, sentence text, sentence position, and sentence category to form a data sample.
[0010] Step 2. Data sample expansion;
[0011] Sentences belonging to the same medical article abstract and of the same category are randomly concatenated to generate new sentences;
[0012] The sentence position of the new sentence after splicing adopts the sentence position information of the first sentence before splicing. The sentence category of the new sentence after splicing is the same as the sentence category of each sentence before splicing. The medical article abstract number to which the new sentence belongs is the same as the medical article abstract number to which each sentence belongs before splicing.
[0013] The medical article abstract number to which each new sentence belongs, along with the sentence text, sentence position, and sentence category of that new sentence, are combined to form a new data sample.
[0014] The second objective of this invention is to propose a PICOS extraction and classification method based on article summary sentences. This method is based on the data augmentation method described above to improve the accuracy of the model in recognizing long sentences.
[0015] To achieve the above objectives, the present invention adopts the following technical solution:
[0016] A PICOS-based extraction and classification method for article summary sentences includes the following steps:
[0017] Step 1. Build the PICOS extraction and classification model;
[0018] The PICOS extraction and classification model consists of a pre-trained model, a splicing module, and a classification module.
[0019] The data sample processing in the PICOS extraction and classification model is as follows:
[0020] The sentence text in the data sample is embedded into a text using a pre-trained model to obtain sentence vectors; the sentence positions in the data sample are also embedded into text using a pre-trained model to obtain sentence order information vectors.
[0021] The sentence vector and sentence order information vector of the data sample are fed into the concatenation module to obtain new sentence vector information. Then, the new sentence vector information is fed into the classification module to output the sentence category.
[0022] Step 2. Train the PICOS extraction and classification model;
[0023] The PICOS extraction classification model is trained using the training data in the training dataset; wherein, the training data in the training dataset comes from the data samples obtained by the data augmentation method described above.
[0024] By minimizing the loss value of the classification module, the model parameters of the classification model are extracted by PICOS, and the model training is completed.
[0025] Step 3. Perform PICOS extraction and classification on the medical article abstracts;
[0026] The trained PICOS extraction and classification model is used to extract and classify the medical article abstract information to be classified.
[0027] Furthermore, the present invention also proposes a computer device corresponding to the above-described PICOS extraction and classification method based on article summary sentences, the computer device comprising a memory and one or more processors.
[0028] The memory stores executable code, and when the processor executes the executable code, it implements the steps of the PICOS extraction and classification method based on article summary sentences described above.
[0029] Furthermore, this invention also proposes a computer-readable storage medium corresponding to the PICOS extraction and classification method based on the above-mentioned article summary sentences, on which the program is stored.
[0030] When executed by the processor, this program implements the steps of the PICOS extraction and classification method based on article summary sentences described above.
[0031] The present invention has the following advantages:
[0032] As described above, this invention discloses a data augmentation method based on article summary sentences. This method automatically augments sentences of the same category from the same medical article summary using random concatenation, effectively solving the technical problems of difficult manual data annotation and limited data volume. Furthermore, this invention proposes a PICOS extraction and classification method based on article summary sentences. This method builds a PICOS extraction and classification model, where the model's training data comes from data samples obtained through the aforementioned data augmentation method. By augmenting the data and incorporating sentence position information during training, the accuracy of the model in recognizing long sentences can be significantly improved, while the model also gains additional input-related knowledge. Attached Figure Description
[0033] Figure 1 This is a flowchart of a data augmentation method based on article summary sentences in an embodiment of the present invention.
[0034] Figure 2 This is a flowchart of the PICOS extraction and classification method based on article summary sentences in an embodiment of the present invention. Detailed Implementation
[0035] Definitions:
[0036] PICOS: In evidence-based medicine, any study consists of patients, intervention, comparison, outcome, and trial design—the PICOS principle. The PICOS principle comprises five elements:
[0037] P (Population) Research Subjects: The population to be studied or issues related to the research subjects.
[0038] I (Intervention): Treatment interventions or observational indicators used in the study population.
[0039] C (Comparison) group: Represents the control group and the indicators that will be given treatment or observed.
[0040] O (Outcome): Represents the outcome indicators and related questions.
[0041] S (Study design): This refers to the type of research design, such as cohort study, case-control study, or cross-sectional study.
[0042] PMID: PubMed Unique Identifier, is the document number in the life sciences and medicine fields included in the PubMed search engine. In this invention, PMID is used to determine which sentences belong to the same medical article abstract.
[0043] Pandas is a tool based on NumPy, created to solve data analysis tasks.
[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:
[0045] like Figure 1 As shown in the figure, this embodiment describes a data augmentation method based on article summary sentences, which includes the following steps:
[0046] Step 1. Data sample construction.
[0047] Before constructing a data sample, data acquisition and preprocessing are required.
[0048] In this embodiment, multiple medical article abstracts are collected, such as 500-1000 medical article abstracts. These medical article abstract data come from open-source literature resource websites, such as CNKI, PubMed, etc.
[0049] After data acquisition, the medical article abstracts need to be cleaned. The main steps of data cleaning include removing medical article abstracts with coding problems or those with only one sentence.
[0050] Then, the obtained medical article abstracts are segmented into sentences according to punctuation marks (such as periods, question marks, and exclamation marks).
[0051] Each sentence after clause segmentation is categorized into six types: P, I, C, O, S, and six others. The sentence's position within the medical article abstract, i.e., its sequence information, is also indicated.
[0052] Each sentence is combined with its medical article abstract number (e.g., PMID), sentence text, sentence position, and sentence category to form a data sample.
[0053] All data samples were stored in an Excel file, with the storage format shown in Table 1 below.
[0054] Table 1
[0055]
[0056]
[0057] In Table 1, the sentence position indicates the order of the sentence in the abstract information of the medical article to which it belongs.
[0058] Step 2. Data sample expansion.
[0059] New sentences are generated by randomly concatenating sentences that belong to the same medical article abstract and are of the same category.
[0060] The sentence position of the new sentence after splicing is based on the sentence position information of the first sentence before splicing, and the sentence category of the new sentence after splicing is the same as the sentence category of each sentence before splicing.
[0061] The first sentence refers to the sentence that appears first among two or more sentences to be joined together.
[0062] A new data sample is formed by combining the medical article abstract number (e.g., PMID) to which each new sentence belongs, along with the sentence text, sentence position, and sentence category of that new sentence.
[0063] In this embodiment, the data sample expansion adopts an automatic data expansion method.
[0064] Specifically, this embodiment uses the pandas tool to process Excel files to achieve automated data augmentation.
[0065] The process of automatic data augmentation is as follows: First, pandas.groupby is used to filter out sentences from the data sample that belong to the same medical article abstract and are of the same category. Then, all the filtered sentences are put into a list.
[0066] The list is traversed, and sentences belonging to the same medical article abstract and the same category are randomly concatenated to generate new sentences.
[0067] For example, a medical article abstract contains sentences {s1, s2, s3, s4, s5}, as shown in Table 2 below.
[0068] Sentences s1, s2, and s4 belong to the same category.
[0069] New sentences are generated by splicing: s6 = s1 + s2, s7 = s1 + s4, s8 = s2 + s4.
[0070] The order information of the newly concatenated sentences s6, s7, and s8 is based on the position information of the first sentence before concatenation. Taking the new sentence s6 as an example, before concatenation, s1 was the first sentence, so the sentence position information of the new sentence s6 is consistent with that of s1.
[0071] The category of the new sentence after splicing is the same as the category of the sentence before splicing.
[0072] Table 2
[0073]
[0074]
[0075] This invention solves the technical problems of difficulty in manual data annotation and insufficient data volume through the above-mentioned data augmentation method.
[0076] Furthermore, this invention also proposes a PICOS extraction and classification method based on article summary sentences, such as... Figure 2 As shown. The PICOS extraction and classification method includes the following steps:
[0077] Step 1. Build the PICOS extraction and classification model.
[0078] The constructed PICOS extraction and classification model includes a pre-trained model, a concatenation module, and a classification module. These components are sequentially connected.
[0079] The input to the pre-trained model is the sentence text and sentence position in the data sample.
[0080] The classification module predicts the sentence category.
[0081] In this embodiment, the pre-trained model uses the BERT pre-trained model; the splicing module uses the torch.cat module; and the classification module includes a LayerNorm layer, two neural network layers, and a softmax classification layer.
[0082] The data sample processing procedure in the PICOS extraction and classification model is as follows:
[0083] The sentence text in the data sample is embedded using a BERT pre-trained model. The state of the last hidden layer of the BERT pre-trained model is used as its embedding output to obtain the sentence vector.
[0084] Similarly, the sentence positions in the data samples are embedded into the text using a pre-trained model to obtain sentence order information vectors.
[0085] The sentence vector and sentence order information vector of the data sample are fed into the torch.cat concatenation module. The torch.cat concatenation module concatenates the two vectors to obtain the new sentence vector information.
[0086] Due to the unique nature of the PICOS principle, this invention incorporates sentence sequence information into the PICOS sentence extraction and classification process of medical article abstracts to utilize the PICOS principle for model training, which can significantly improve the model's accuracy.
[0087] The new sentence vector information is fed into the classification module, which outputs the sentence category. The specific process is as follows:
[0088] The new sentence vector information first enters the LayerNorm layer, then enters a two-layer neural network structure (using ReLU activation function and dropout value of 0.3), and finally passes through softmax to obtain the sentence category. The classification module uses the cross-entropy loss function.
[0089] Step 2. Train the PICOS extraction and classification model.
[0090] The PICOS extraction classification model is trained using the training data in the training dataset; the training data in this dataset comes from the data samples obtained by the data augmentation method based on article summary sentences mentioned above.
[0091] Unlike the data samples mentioned above, the training data does not include medical article abstract numbers (e.g., PMIDs).
[0092] Sentences of the same category in an article describe the most similar information. This invention expands the data in this way, which will significantly improve the accuracy of the model in recognizing long sentences. At the same time, the model will gain additional knowledge related to the input.
[0093] By minimizing the loss value of the classification module, PICOS extracts the model parameters of the classification model, thus completing the model training.
[0094] Step 3. Perform PICOS extraction and classification on the medical article abstract information.
[0095] The trained PICOS extraction and classification model is used to extract and classify the medical article abstract information to be classified.
[0096] This invention significantly improves the accuracy of PICOS extraction and classification by incorporating sentence positional information and using data samples obtained through the aforementioned data augmentation method during the training process of the PICOS extraction and classification model.
[0097] Furthermore, this embodiment of the invention also proposes a computer device for implementing the above-described PICOS extraction and classification method based on article summary sentences. The computer device includes a memory and one or more processors.
[0098] The system contains executable code stored in memory. When the processor executes the executable code, it implements the steps of the PICOS extraction and classification method based on article summary sentences.
[0099] In this embodiment, the computer device can be any device or apparatus with data processing capabilities, and will not be described in detail here.
[0100] Furthermore, embodiments of the present invention also provide a computer-readable storage medium having a program stored thereon, which, when executed by a processor, is used to implement the steps of the above-described PICOS extraction and classification method based on article summary sentences.
[0101] The computer-readable storage medium can be an internal storage unit of any device or apparatus with data processing capabilities, such as a hard disk or memory, or an external storage device of any device with data processing capabilities, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc.
[0102] Of course, the above description is only a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. It should be noted that any equivalent substitutions or obvious modifications made by those skilled in the art under the guidance of this specification fall within the scope of this specification and should be protected by the present invention.
Claims
1. A PICOS extraction and classification method based on article summary sentences, characterized in that, Includes the following steps: Step 1. Build the PICOS extraction and classification model; The PICOS extraction and classification model consists of a pre-trained model, a splicing module, and a classification module. The data sample processing procedure in the PICOS extraction and classification model is as follows: The sentence text in the data sample is embedded into a text using a pre-trained model to obtain sentence vectors; the sentence positions in the data sample are also embedded into text using a pre-trained model to obtain sentence order information vectors. The sentence vector and sentence order information vector of the data sample are fed into the concatenation module to obtain new sentence vector information. Then, the new sentence vector information is fed into the classification module to output the sentence category. Step 2. Train the PICOS extraction and classification model; The PICOS extraction and classification model is trained using the training data in the training dataset, and the order information of sentences is added to train the model using the principles of PICOS. The training data in the training dataset comes from the following data samples obtained by the data augmentation method based on article summary sentences; By minimizing the loss value of the classification module, the model parameters of the classification model are extracted by PICOS, and the model training is completed. Step 3. Perform PICOS extraction and classification on the medical article abstracts; The trained PICOS extraction and classification model is used to extract and classify the medical article abstract information to be classified. The data augmentation method based on article summary sentences includes the following steps: Step I. Data Sample Construction; First, the abstracts of multiple medical articles are segmented into sentences. Then, each sentence after segmentation is categorized and its position within the abstract of the corresponding medical article is marked, i.e., its order information. Each sentence is combined with its corresponding medical article abstract number, sentence text, sentence position, and sentence category to form a data sample. Step II. Data Sample Enlargement; Sentences belonging to the same medical article abstract and of the same category are randomly concatenated to generate new sentences; The sentence position of the new sentence after splicing adopts the sentence position information of the first sentence before splicing. The sentence category of the new sentence after splicing is the same as the sentence category of each sentence before splicing. The medical article abstract number to which the new sentence belongs is the same as the medical article abstract number to which each sentence belongs before splicing. The medical article abstract number to which each new sentence belongs, along with the sentence text, sentence position, and sentence category of that new sentence, are combined to form a new data sample. There are six sentence categories: P, I, C, O, S, and six others.
2. The PICOS extraction and classification method based on article summary sentences according to claim 1, characterized in that, In step I, all data samples are stored in an Excel file; In step II, the pandas tool is used to process the Excel file to achieve automated data augmentation: Specifically, the process involves first using pandas.groupby to filter out sentences from the data sample that belong to the same medical article abstract and are of the same category, and then putting all the filtered sentences into a list. The list is traversed, and sentences belonging to the same medical article abstract and the same category are randomly concatenated to generate new sentences.
3. The PICOS extraction and classification method based on article summary sentences according to claim 1, characterized in that, In step I, the medical article abstract is divided into sentences according to punctuation marks.
4. The PICOS extraction and classification method based on article summary sentences according to claim 1, characterized in that, In step I, before segmenting the medical article abstract into sentences, there are also steps of data acquisition and preprocessing, namely: Collect medical article abstracts and perform data cleaning on the medical article abstracts.
5. The PICOS extraction and classification method based on article summary sentences according to claim 4, characterized in that, In step I, the medical article abstracts are sourced from open-source literature resource websites; In step I, data cleaning includes removing medical article abstracts that have coding problems or have only one sentence in their summary.
6. The PICOS extraction and classification method based on article summary sentences according to claim 1, characterized in that, The pre-trained model uses the BERT pre-trained model; the splicing module uses the torch.cat module; the classification module includes a LayerNorm layer, two neural network layers, and a softmax classification layer.
7. A computer device comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code Implement the PICOS extraction and classification method based on article summary sentences as described in claim 1.
8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the PICOS extraction and classification method based on article summary sentences as described in claim 1.