Literature efficacy information extraction model training method, literature efficacy information extraction method

Abstract

The present invention discloses a document efficacy information extraction model training method and a document efficacy information extraction method, which belong to the field of natural language processing technology and solve the problem of low accuracy and low efficiency of extracting efficacy information by the existing manual reading method. The method comprises: extracting efficacy sentences and non-efficacy sentences from document samples and adding corresponding labels to form an efficacy sentence extraction sample set; using the efficacy sentence extraction sample set to train and verify the efficacy sentence extraction function module in the document efficacy information extraction model; using the efficacy sentences in the document samples and the entity information of the efficacy sentences obtained by decomposing the efficacy sentences to train and verify the entity information extraction function module in the document efficacy information extraction model; using the entity information obtained by decomposing the efficacy sentences in the document samples and the efficacy phrases obtained by combining the entity information according to predetermined sentence rules to train and verify the efficacy phrase generation function module in the document efficacy information extraction model.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a document efficacy information extraction model training method and a document efficacy information extraction method. Background Art

[0002] With the advancement of science and technology and the rapid development of information technology, the importance of patent documents and other various scientific and technological documents in scientific and technological development is increasing day by day. It is particularly important to effectively analyze and use patent documents and other various scientific and technological documents.

[0003] In the specific practice, users are very concerned about the functions and effects that can be achieved by patent documents and other scientific and technological documents. Therefore, professional methods are needed to effectively extract and classify the function and effect information (i.e. "efficacy information") in the documents. In the existing methods, efficacy information is roughly extracted through simple manual reading, which has the problems of low accuracy and low efficiency.

[0004] Therefore, how to design and train a model for extracting literature efficacy information and achieve accurate and rapid extraction of literature efficacy information is a technical problem that needs to be solved urgently. Summary of the invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a document efficacy information extraction model training method and a document efficacy information extraction method to solve the problem of low accuracy and low efficiency in extracting efficacy information by existing manual reading methods.

[0006] In one aspect, the present invention discloses a method for training a model for extracting document efficacy information, the method comprising:

[0007] Obtain multiple literature samples in the target research field, extract effective sentences and non-effective sentences from the literature samples, and add corresponding labels to form an effective sentence extraction sample set;

[0008] Using the efficacy sentence extraction sample set to train and verify the efficacy sentence extraction function module in the literature efficacy information extraction model;

[0009] Using the efficacy sentences in the literature samples and the entity information of the efficacy sentences obtained by decomposing the efficacy sentences to train and verify the entity information extraction function module in the literature efficacy information extraction model;

[0010] The entity information obtained by decomposing the efficacy sentences in the document samples and the efficacy phrases obtained by combining the entity information according to predetermined sentence rules are used to train and verify the efficacy phrase generation function module in the document efficacy information extraction model.

[0011] On the basis of the above scheme, the present invention also makes the following improvements:

[0012] Furthermore, the efficacy sentence extraction function module is implemented based on the efficacy sentence extraction model; the efficacy sentence extraction model is trained and verified using the efficacy sentence extraction sample set, and the following is performed:

[0013] The efficacy sentence extraction sample set is divided into an efficacy sentence extraction sample training set and an efficacy sentence extraction sample verification set;

[0014] The efficacy sentence extraction model is trained using the efficacy sentence extraction sample training set;

[0015] Afterwards, the efficacy sentence extraction model is verified using the efficacy sentence extraction sample verification set, thereby obtaining a verified efficacy sentence extraction model.

[0016] Further, the efficacy sentence extraction model includes a chinese_roberta pre-training sub-model and an efficacy sentence classification sub-model; the efficacy sentence extraction model is trained and verified using the efficacy sentence extraction sample set, and further performs:

[0017] Using the open source framework Bert4keras, based on the chinese_roberta pre-trained model, we built a model framework for extracting efficacy sentences and set the training parameters of the model.

[0018] The efficacy sentences and non-efficacy sentences in the literature samples are respectively input into the chinese_roberta pre-trained sub-model to obtain the corresponding sentence vectors, and the sentence vectors are input into the efficacy sentence classification sub-model. Based on the corresponding labels, the training and verification of the efficacy sentence classification sub-model are realized.

[0019] Furthermore, the entity information extraction function module is implemented based on the entity annotation model; the entity annotation model is trained and verified by using the entity information of the function sentences obtained by decomposing the function sentences in the literature samples, and executing:

[0020] Build an entity annotation model based on GlobalPointer and set the parameters for entity annotation model training;

[0021] The entity annotation model is trained and verified by taking the functional sentences in the document samples as input and the entity information of the functional sentences obtained by decomposing the corresponding functional sentences as output, so as to obtain a verified entity annotation model.

[0022] Furthermore, the entity information includes the following categories: efficacy words, subject words, parameters, efficacy adjectives; wherein,

[0023] Function words refer to the verbs in function sentences;

[0024] The subject word refers to the object of the verb in the efficacy sentence;

[0025] Parameter refers to the object of action of the verb in the efficacy sentence;

[0026] Functional adjectives refer to changes in parameters.

[0027] Furthermore, the functional phrase generation module is implemented based on the functional phrase generation model; the functional phrase training and verification model is obtained by using the entity information obtained by decomposing the functional sentences in the literature samples and combining the entity information according to the predetermined sentence rules, and executing:

[0028] Using the open source framework Bert4keras, the seq2seq automatic title generation algorithm based on the chinese_roberta model was adopted as the built-in algorithm of the efficacy phrase generation model. The entity information obtained by decomposing the efficacy sentences in the literature samples was used as input, and the efficacy phrases obtained by combining according to predetermined sentence rules were used as output to train and verify the efficacy phrase generation model, and a verified efficacy phrase generation model was obtained.

[0029] Furthermore, the efficacy phrase generation model includes a plurality of efficacy phrase generation sub-models, and each sentence rule corresponds to one efficacy phrase generation sub-model;

[0030] The entity information obtained by decomposing the functional sentences in the literature samples is used as input, and the functional phrases obtained by combining according to the predetermined sentence rules are used as output. The functional phrase generation sub-model corresponding to the sentence rules is trained and verified to obtain a verified functional phrase generation sub-model.

[0031] Furthermore, the sentence rules include: efficacy word + parameter; efficacy word + subject word + parameter; subject word + efficacy adjective.

[0032] On the other hand, the present invention also provides a method for extracting document efficacy information, which is implemented based on the document efficacy information extraction model obtained by the above document efficacy information extraction model training method;

[0033] The method for extracting literature efficacy information comprises:

[0034] The validated efficacy sentence extraction function module is used to extract efficacy sentences from the document to be processed, and semantically removes duplicate efficacy sentences from the extracted efficacy sentences to obtain valid efficacy sentences of the document to be processed;

[0035] Input the effective efficacy sentences of the document to be processed into the verified entity information extraction function module, and the entity annotation model outputs the entity information in the effective efficacy sentences;

[0036] Input the entity information in the effective efficacy sentence into the verified efficacy phrase generation function module, and the efficacy phrase generation model outputs the efficacy phrase corresponding to the effective efficacy sentence;

[0037] The effective efficacy sentences of the document to be processed, the efficacy words in the effective efficacy sentences, and the efficacy phrases corresponding to the effective efficacy sentences are combined as the efficacy information extraction result of the document to be processed.

[0038] On the basis of the above scheme, the present invention also makes the following improvements:

[0039] Furthermore, the extracted effect sentences are semantically deduplicated by executing:

[0040] The efficacy sentences in the extracted documents to be processed are traversed, the efficacy sentence vectors are extracted based on the Chinese_Roberta model, and the similarity between the efficacy sentence vectors is calculated using the Roformer-Sim framework algorithm. After removing the duplicate efficacy sentences whose similarity exceeds the set first threshold, the effective efficacy sentences of the documents to be processed are obtained.

[0041] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0042] The document efficacy information extraction model training method provided by the present invention extracts efficacy sentences and non-efficacy sentences from document samples, and on this basis, performs the training and verification of the efficacy sentence extraction function module in the document efficacy information extraction model. On this basis, the efficacy sentences are decomposed into corresponding entity information, and the training and verification of the entity information extraction function module in the document efficacy information extraction model are realized. Finally, based on the entity information and the corresponding efficacy phrases, the training and verification of the efficacy phrase generation function module in the document efficacy information extraction model are realized. The conception of this method is novel, and it can realize the effective training of the document efficacy information extraction model, and provide a basis for the extraction of document efficacy information. At the same time, this method ensures the accuracy of the model training and verification results by selecting an adaptive model at different stages.

[0043] The document efficacy information extraction method provided by the present invention, based on the above-mentioned document efficacy information extraction model, first uses the efficacy sentence extraction function module to extract the efficacy sentences in the document to be processed, and obtains the effective efficacy sentences of the document to be processed after semantic deduplication; then inputs the effective efficacy sentences of the document to be processed into the entity information extraction function module, outputs the entity information; inputs the entity information in the effective efficacy sentences into the efficacy phrase generation function module, outputs the efficacy phrase; and combines the effective efficacy sentences of the document to be processed, the efficacy words in the effective efficacy sentences, and the efficacy phrases corresponding to the effective efficacy sentences as the efficacy information extraction result of the document to be processed. The method can effectively extract the efficacy information in the document to be processed with high accuracy. Since the steps are interconnected, there is less redundant information and the extraction speed of efficacy information is fast.

[0044] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0046] Figure 1 A flowchart of a method for training a document efficacy information extraction model provided by an embodiment of the present invention;

[0047] Figure 2 A flowchart of a method for extracting document efficacy information provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0049] Example 1

[0050] Specific embodiment 1 of the present invention discloses a method for training a model for extracting document efficacy information. The flowchart of the method is as follows: Figure 1 The specific steps are as follows:

[0051] Step S1: Obtain multiple literature samples in the target research field, extract effective sentences and non-effective sentences from the literature samples, and add corresponding labels to form an effective sentence extraction sample set.

[0052] In this embodiment, multiple document samples in the target research field are obtained as full-text corpus data to be extracted. In the process of extracting functional sentences and non-functional sentences from the full-text corpus data, the data is required to be diverse and universal. In this embodiment, in order to ensure that the above-mentioned training method has a wider scope of application, when selecting document samples, it is possible to consider selecting multiple categories and multiple document samples under each category according to the coverage of the target research field, and summarize all document samples. The training method in this embodiment is applicable to patent documents and scientific and technological documents. Taking patent documents as an example, at least 30 document samples can be randomly selected from the data of the past 10 years in 648 subcategories, and a total of 22,545 document samples can be extracted.

[0053] In the specific implementation process, the efficacy sentences and non-efficacy sentences in each document sample can be extracted by manual annotation, and corresponding labels can be added. The specific example is described as follows: Considering the writing of the patent document, the extraction scope of the efficacy sentence is preferably the abstract and invention content of the patent document, and is preferably extracted from the beneficial effect paragraph of the invention content. Through "the present invention has the following advantages", "the present invention has the following beneficial effects", "compared with the relevant technology", "the positive progressive effect of the invention is" ..., it can be determined that the corresponding paragraph belongs to the beneficial effect paragraph; then use characteristic word collocation, such as "through ..., to achieve the purpose of ...", "improve ...", "reduce ...", "improve ...", "optimize ...", "... enhance", "energy saving and environmental protection", "time saving and labor saving" ..., extract the efficacy sentences of the patent document, and lay the foundation for further extracting efficacy short words. In this embodiment, a total of 92,970 efficacy sentence indexing corpora are manually extracted, and after manual semantic deduplication, a total of 71,577 efficacy sentence indexing corpora are obtained, and the label of each efficacy sentence is "1". Then, more than 100,000 non-effective sentences were randomly extracted from the full text corpus as negative samples, and the label of each non-effective sentence was "0", and finally a total of 172,558 sample data were obtained. The order of the 172,558 sample data was randomly shuffled to form a sample set of effective sentence extraction.

[0054] Step S2: using the efficacy sentence extraction sample set to train and verify the efficacy sentence extraction function module in the literature efficacy information extraction model.

[0055] Preferably, the efficacy sentence extraction function module is implemented based on the efficacy sentence extraction model. In the specific implementation process, the efficacy sentence extraction sample set is divided into an efficacy sentence extraction sample training set and an efficacy sentence extraction sample verification set, and the efficacy sentence extraction model is trained using the efficacy sentence extraction sample training set; then, the efficacy sentence extraction model is verified using the efficacy sentence extraction sample verification set, thereby obtaining a verified efficacy sentence extraction model.

[0056] Preferably, the efficacy sentence extraction model includes a chinese_roberta pre-trained sub-model and an efficacy sentence classification sub-model. Using the open source framework Bert4keras, based on the chinese_roberta pre-trained model, the efficacy sentence extraction model framework is built, and the training parameters of the efficacy sentence extraction model are set. Among them, the efficacy sentences and non-efficacy sentences in the literature sample are respectively input into the chinese_roberta pre-trained sub-model to obtain corresponding sentence vectors, and the sentence vectors are input into the efficacy sentence classification sub-model, and the training and verification of the efficacy sentence classification sub-model are realized based on the corresponding labels.

[0057] Exemplarily, the training parameters of the efficacy sentence extraction model are set as follows:

[0058] num_classes=2, number of categories (0-non-effective sentence, 1-effective sentence);

[0059] maxlen=128, input sentence length;

[0060] batch_size=16, the number of training sentences submitted in each batch;

[0061] epochs=200, training iteration rounds.

[0062] In the specific implementation process, the trained efficacy sentence extraction model can be used to run a batch of new data, manually verify and continue to index, add the missing efficacy sentences to the training set, and continue training. Repeat the above three times, and the accuracy rate reaches 90%+, which basically achieves the effect of manual indexing.

[0063] Step S3: using the functional sentences in the document samples and the entity information of the functional sentences obtained by decomposing the functional sentences to train and verify the entity information extraction function module in the document functional information extraction model.

[0064] The entity information includes the following categories: efficacy words, subject words, parameters, efficacy adjectives. Based on the efficacy sentences of the literature samples, the words in the efficacy sentences are decomposed into efficacy words, subject words, parameters, and efficacy adjectives. Efficacy words refer to verbs in efficacy sentences, such as prolong, enhance, improve, reduce, and so on; subject words refer to the object subject of the verb in the efficacy sentence, such as "audio output" in "improve the quality of audio output"; parameters refer to the object of the verb in the efficacy sentence, such as "quality" in "improve the quality of audio output"; efficacy adjectives refer to words that change parameters, mostly adjectives, such as "low" in "low failure rate". The purpose of splitting is to facilitate the classification of different words in the later use process, and more than 14,000 efficacy words have been manually annotated.

[0065] Preferably, the entity information extraction function module is implemented based on the entity annotation model. In the specific implementation process, an entity annotation model based on GlobalPointer is built, and the parameters for entity annotation model training are set. The entity annotation model is trained and verified with the functional sentences in the literature sample as input and the entity information of the functional sentences obtained by decomposing the corresponding functional sentences as output, and a verified entity annotation model is obtained.

[0066] as follows:

[0067] maxlen=128, input sentence length;

[0068] epochs=100, training iteration rounds;

[0069] batch_size=16, the number of training sentences submitted in each batch,

[0070] learning_rate=2e-5, learning rate.

[0071] Using the trained entity labeling model, we ran a batch of new data, manually verified and continued the functional word indexing, added the missing functional words to the training set, and continued training. After three rounds of repetition, the accuracy rate reached 76.4%, which has met the basic requirements of downstream applications.

[0072] Step S4: using entity information obtained by decomposing functional sentences in document samples and functional phrases obtained by combining entity information according to predetermined sentence rules to train and verify the functional phrase generation function module in the document functional information extraction model.

[0073] Specifically, in this embodiment, on the premise of accurate extraction of the efficacy words, the efficacy phrases are manually annotated according to the predetermined sentence rules. The predetermined sentence rules can be: (1) efficacy word + parameter (such as: improve cutting efficiency); (2) efficacy word + subject word + parameter (such as: extend the service life of the supercharger); (3) subject word + efficacy adjective (such as: the method is simple to use), etc. Then, manual auxiliary annotation is performed as the efficacy phrase after the efficacy sentence is compressed, and more than 14,000 efficacy phrases are annotated.

[0074] Preferably, the functional module for generating efficacy phrases is implemented based on the efficacy phrase generation model. Using the open source framework Bert4keras, the seq2seq automatic title generation algorithm based on the chinese_roberta model is adopted as the built-in algorithm of the efficacy phrase generation model. The entity information obtained by decomposing the efficacy sentences in the literature samples is used as input, and the efficacy phrases obtained by combining according to the predetermined sentence rules are used as output to train and verify the efficacy phrase generation model, and obtain a verified efficacy phrase generation model. In the specific implementation process, the parameters are set to input sentence token length maxlen=128, training iteration rounds epochs=30, and finally the efficacy phrase generation model is saved.

[0075] The functional phrase generation model includes multiple functional phrase generation sub-models, and each sentence rule corresponds to a functional phrase generation sub-model. The entity information obtained by decomposing the functional sentences in the document sample is used as input, and the functional phrases obtained by combining according to the predetermined sentence rules are used as output to train and verify the functional phrase generation sub-model corresponding to the sentence rule, so as to obtain a verified functional phrase generation sub-model.

[0076] In the specific implementation process, we can also use the trained efficacy phrase generation model, run a batch of new data, manually verify, continue to index the efficacy phrases, add the missed efficacy phrases to the training set, and continue training. After three rounds of repetition, the accuracy rate reaches 91.2%, which is close to the quality of manual indexing.

[0077] The document efficacy information extraction model training method provided in the present embodiment extracts efficacy sentences and non-efficacy sentences from document samples, and performs the training and verification of the efficacy sentence extraction function module on this basis. On this basis, the efficacy sentences are decomposed into corresponding entity information to realize the training and verification of the entity information extraction function module. Finally, based on the entity information and the corresponding efficacy phrases, the training and verification of the efficacy phrase generation function module are realized. The conception of this method is novel, and it can realize the effective training of the document efficacy information extraction model, and provide a basis for the extraction of document efficacy information. At the same time, this method ensures the accuracy of the model training and verification results by selecting the adapted model at different stages.

[0078] Example 2

[0079] The embodiment of the present invention discloses a method for extracting literature efficacy information, the flow chart of which is as follows: Figure 2 As shown, the method is implemented based on the literature efficacy information extraction model training method provided in Example 1, and the method includes the following steps:

[0080] Step S1: extracting the efficacy sentences in the document to be processed by using the verified efficacy sentence extraction function module, and performing semantic deduplication on the extracted efficacy sentences to obtain the effective efficacy sentences of the document to be processed.

[0081] Specifically, after extracting the efficacy sentences of a certain patent, there will be a large number of repeated sentences, which need to be semantically deduplicated. To perform semantic deduplication on the extracted efficacy sentences, perform: traverse the efficacy sentences extracted from the document to be processed, extract the efficacy sentence vectors based on the Chinese_Roberta model, and use the Roformer-Sim framework algorithm to calculate the similarity between the efficacy sentence vectors, and after deduplication of the efficacy sentences whose similarity exceeds the set first threshold, obtain the valid efficacy sentences of the document to be processed.

[0082] Exemplarily, when the similarity between two function sentence vectors exceeds 90% (for example, the first threshold), they are considered to be semantically duplicated, and one of them can be deleted.

[0083] Step S2: Input the valid efficacy sentences of the document to be processed into the verified entity information extraction function module, and the entity annotation model outputs the entity information in the valid efficacy sentences.

[0084] Step S3: Input the entity information in the effective efficacy sentence into the verified efficacy phrase generation function module, and the efficacy phrase generation model outputs the efficacy phrase corresponding to the effective efficacy sentence.

[0085] Step S4: combining the effective efficacy sentences of the document to be processed, the efficacy words in the effective efficacy sentences, and the efficacy phrases corresponding to the effective efficacy sentences as the efficacy information extraction result of the document to be processed.

[0086] The document efficacy information extraction method provided in this embodiment, based on the above-mentioned document efficacy information extraction model, first uses the efficacy sentence extraction function module to extract the efficacy sentences in the document to be processed, and obtains the effective efficacy sentences of the document to be processed after semantic deduplication; then inputs the effective efficacy sentences of the document to be processed into the entity information extraction function module, outputs the entity information; inputs the entity information in the effective efficacy sentences into the efficacy phrase generation function module, outputs the efficacy phrase; combines the effective efficacy sentences of the document to be processed, the efficacy words in the effective efficacy sentences, and the efficacy phrases corresponding to the effective efficacy sentences as the efficacy information extraction result of the document to be processed. This method can effectively extract the efficacy information in the document to be processed with high accuracy. Since each step is interconnected, there is less redundant information and the extraction speed of efficacy information is fast.

[0087] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0088] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for training a model for extracting literature efficacy information, characterized in that: The method comprises: Obtain multiple categories of literature samples in the target research field, and multiple literature samples in each category, extract effective sentences and non-effective sentences from the literature samples, and add corresponding labels to form an effective sentence extraction sample set; Using the efficacy sentence extraction sample set to train and verify the efficacy sentence extraction function module in the literature efficacy information extraction model; The entity information extraction function module in the document efficacy information extraction model is trained and verified by using the efficacy sentences in the document samples and the entity information of the efficacy sentences obtained by decomposing the efficacy sentences; the entity information includes the following categories: efficacy words, subject words, parameters, and efficacy adjectives; wherein the efficacy words refer to the verbs in the efficacy sentences; the subject words refer to the objects acted on by the verbs in the efficacy sentences; the parameters refer to the objects acted on by the verbs in the efficacy sentences; and the efficacy adjectives refer to the words that change the parameters; The function phrase generation function module in the document efficacy information extraction model is trained and verified by using the entity information obtained by decomposing the efficacy sentences in the document samples and the efficacy phrases obtained by combining the entity information according to the predetermined sentence rules; the sentence rules include: efficacy word + parameter; efficacy word + subject word + parameter; subject word + efficacy adjective; The efficacy sentence extraction function module is implemented based on the efficacy sentence extraction model; the efficacy sentence extraction model includes the Chinese_roberta pre-training sub-model and the efficacy sentence classification sub-model; the efficacy sentence extraction model is trained and verified using the efficacy sentence extraction sample set, and further performs: Using the open source framework Bert4keras, based on the chinese_roberta pre-trained model, we built a model framework for extracting efficacy sentences and set the training parameters of the model. The efficacy sentences and non-efficacy sentences in the literature samples are respectively input into the chinese_roberta pre-trained sub-model to obtain the corresponding sentence vectors, and the sentence vectors are input into the efficacy sentence classification sub-model. Based on the corresponding labels, the training and verification of the efficacy sentence classification sub-model are realized.

2. The method for extracting document efficacy information from a model according to claim 1, characterized in that: Use the efficacy sentence extraction sample set to train and verify the efficacy sentence extraction model and execute: The efficacy sentence extraction sample set is divided into an efficacy sentence extraction sample training set and an efficacy sentence extraction sample verification set; The efficacy sentence extraction model is trained using the efficacy sentence extraction sample training set; Afterwards, the efficacy sentence extraction model is verified using the efficacy sentence extraction sample verification set, thereby obtaining a verified efficacy sentence extraction model.

3. The method for extracting document efficacy information from a model according to claim 2, characterized in that: The entity information extraction function module is implemented based on the entity annotation model; the entity annotation model is trained and verified by using the entity information of the efficacy sentences obtained by decomposing the efficacy sentences in the literature samples, and executing: Build an entity annotation model based on GlobalPointer and set the parameters for entity annotation model training; The entity annotation model is trained and verified by taking the functional sentences in the document samples as input and the entity information of the functional sentences obtained by decomposing the corresponding functional sentences as output, so as to obtain a verified entity annotation model.

4. The method for extracting document efficacy information from a model according to claim 1, characterized in that: The functional phrase generation module is implemented based on the functional phrase generation model; the functional phrase training and verification model is obtained by using the entity information obtained by decomposing the functional sentences in the literature samples and combining the entity information according to the predetermined sentence rules to perform: Using the open source framework Bert4keras, the seq2seq automatic title generation algorithm based on the chinese_roberta model was adopted as the built-in algorithm of the efficacy phrase generation model. The entity information obtained by decomposing the efficacy sentences in the literature samples was used as input, and the efficacy phrases obtained by combining according to predetermined sentence rules were used as output to train and verify the efficacy phrase generation model, and a verified efficacy phrase generation model was obtained.

5. The method for extracting document efficacy information from a model according to claim 4, characterized in that: The functional phrase generation model includes multiple functional phrase generation sub-models, and each sentence rule corresponds to a functional phrase generation sub-model; The entity information obtained by decomposing the functional sentences in the literature samples is used as input, and the functional phrases obtained by combining according to the predetermined sentence rules are used as output. The functional phrase generation sub-model corresponding to the sentence rules is trained and verified to obtain a verified functional phrase generation sub-model.

6. A method for extracting literature efficacy information, characterized in that: The method is implemented based on a document efficacy information extraction model obtained by a document efficacy information extraction model training method according to any one of claims 1 to 5; The method for extracting literature efficacy information comprises: The validated efficacy sentence extraction function module is used to extract efficacy sentences from the document to be processed, and semantically removes duplicate efficacy sentences from the extracted efficacy sentences to obtain valid efficacy sentences of the document to be processed; Input the effective efficacy sentences of the document to be processed into the verified entity information extraction function module, and the entity annotation model outputs the entity information in the effective efficacy sentences; Input the entity information in the effective efficacy sentence into the verified efficacy phrase generation function module, and the efficacy phrase generation model outputs the efficacy phrase corresponding to the effective efficacy sentence; The effective efficacy sentences of the document to be processed, the efficacy words in the effective efficacy sentences, and the efficacy phrases corresponding to the effective efficacy sentences are combined as the efficacy information extraction result of the document to be processed.

7. The method for extracting document efficacy information according to claim 6, characterized in that: To perform semantic deduplication on the extracted efficacy sentences, execute: The efficacy sentences in the extracted documents to be processed are traversed, the efficacy sentence vectors are extracted based on the Chinese_Roberta model, and the similarity between the efficacy sentence vectors is calculated using the Roformer-Sim framework algorithm. After removing the duplicate efficacy sentences whose similarity exceeds the set first threshold, the effective efficacy sentences of the documents to be processed are obtained.

Images