A method, device, electronic device and storage medium for generating a question-answer library based on a knowledge graph

Abstract

The present invention discloses a method, device, electronic device and storage medium for generating a question-answering database based on a knowledge graph, the method comprising: obtaining entity relationship combinations from a knowledge graph; randomly masking entities in the entity relationship combinations, extracting some of the masked entity relationship combinations, and generating fine-tuning data; fine-tuning a pre-trained model obtained by pre-training using the fine-tuning data to obtain a question generation model; and generating a question-answering dataset based on the full amount of masked entity relationship combinations and the question generation model. The scheme of the present invention reduces the cost of constructing a question-answering dataset, ensures the accuracy and quality of the question-answering dataset, and can achieve KBQA tasks through database retrieval and matching.

Description

Technical Field

[0001] The present invention relates to the field of question-answering systems, and in particular to a method, device, electronic device and storage medium for generating a question-answering library based on a knowledge graph. Background Art

[0002] The knowledge graph is the infrastructure for intelligent applications such as the next generation of question-answering systems. If the question-answering system is regarded as a brain, then the knowledge graph is the knowledge base in the brain, which enables the machine to analyze and think about problems from the perspective of entities and relationships. As the basis for training models for dialogue question-answering tasks, the construction of large-scale high-quality question-answering datasets is of great significance.

[0003] The existing question-answering datasets are mostly constructed by crawling and integrating Internet data or manually constructing them based on knowledge graphs. However, the traditional methods have shortcomings: first, the method of crawling data from the Internet cannot guarantee the quality of the dataset, which increases the burden of data cleaning. Internet data has problems such as data privacy, data rights confirmation and data bias, and there are certain legal risks and data quality problems that cannot be solved by data cleaning; secondly, the manual construction method based on knowledge graphs is costly and has a long construction period, so it is difficult for the dataset to meet the requirements of the question-answering system in terms of scale. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention provides the following technical solutions.

[0005] The first aspect of the present invention provides a method for generating a question-answer library based on a knowledge graph, comprising:

[0006] Obtain entity relationship combinations from the knowledge graph;

[0007] Randomly masking the entities in the entity relationship combination, extracting some of the masked entity relationship combinations, and generating fine-tuning data;

[0008] Using the fine-tuning data, fine-tuning the pre-trained model obtained by pre-training to obtain a question generation model;

[0009] Based on the masked combination of all entity relationships and the question generation model, a question-answering dataset is generated.

[0010] Preferably, the obtaining entity relationship combination from the knowledge graph further includes:

[0011] Processing triples in the knowledge graph based on heuristic rules as entity relationship combinations;

[0012] The heuristic rules further include:

[0013] Templates matching first-degree relationships of entities;

[0014] Templates that match second-degree relationships of entities; and / or

[0015] A template that joins two entities to a first degree relationship.

[0016] Preferably, the pre-trained model is pre-trained in the following manner:

[0017] Get reading comprehension corpus and encyclopedia corpus;

[0018] Generate questions using the paragraphs and answers in the reading comprehension corpus to obtain a first training data set;

[0019] Performing mask prediction on entities marked with hyperlinks in the encyclopedia corpus to obtain a second training data set;

[0020] The pre-trained model is obtained by training a specific model using the first training data set and the second training data set.

[0021] Preferably, the generating fine-tuning data further comprises:

[0022] For the extracted masked partial entity relationship combinations, questions are obtained by manually asking questions, and then each masked entity relationship combination and the question asked about it are used as a question-and-answer data to generate the fine-tuning data.

[0023] Preferably, the fine-tuning the pre-trained model using the fine-tuning data to obtain a question generation model further includes:

[0024] Based on the generated fine-tuning data, the masked entity relationship combination is used as the input of the encoder end of the pre-trained model, and the questions in the fine-tuning data are used as the output of the decoder end of the pre-training model. The pre-trained model is end-to-end fine-tuned to obtain a question generation model that generates questions from the masked entity relationship combination.

[0025] Preferably, the step of generating a question-answering dataset based on the masked full entity relationship combination and the question generation model further comprises:

[0026] The masked full entity relationship combination is input into the encoder end of the question generation model. For each input masked entity relationship combination, the decoder end of the question generation model outputs the corresponding question, using the masked entity in each entity relationship combination as the answer, and associating the question and answer to form the question and answer dataset.

[0027] Another aspect of the present invention provides a question-answering method based on a knowledge graph, comprising:

[0028] Receive user input questions;

[0029] Searching a question-answer corpus according to the question to obtain a question-answer pair, wherein the question-answer corpus is generated based on a knowledge graph according to the method described in the first aspect above;

[0030] The answer in the question-answer pair is returned to the user.

[0031] The third aspect of the present invention provides a device for generating a question-answer library based on a knowledge graph, characterized in that it includes:

[0032] The acquisition module obtains entity relationship combinations from the knowledge graph;

[0033] A processing module randomly masks the entities in the entity relationship combination, extracts some of the masked entity relationship combinations, and generates fine-tuning data;

[0034] A fine-tuning module, using the fine-tuning data to fine-tune the pre-trained model obtained by pre-training to obtain a question generation model;

[0035] The generation module generates a question-answering dataset based on the masked full entity relationship combination and the question generation model.

[0036] A fourth aspect of the present invention provides an electronic device, characterized in that it includes a processor and a memory, the memory stores a plurality of instructions, and the processor is used to read the instructions and execute the method as described in the first aspect above.

[0037] A fifth aspect of the present invention provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor and executed as the method of the first aspect described above.

[0038] The beneficial effects of the present invention are:

[0039] First, the question-and-answer database of the present invention is automatically generated based on the model. Compared with traditional solutions, the process of manually inputting question-and-answer data is omitted, which significantly reduces the cost of data set construction.

[0040] Secondly, the questions generated in the question-answering dataset of the present invention are derived from the existing knowledge graph, which ensures the accuracy and high quality of the dataset.

[0041] Third, the question-answering corpus constructed based on the method of the present invention can realize the knowledge base-based question-answering (KBQA) task only through simple database retrieval and matching. Compared with the traditional KBQA method, it can improve the efficiency and the accuracy of question-answering. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flowchart of the method for generating a question and answer database based on a knowledge graph described in the present invention.

[0043] Figure 2 This is a flowchart of three heuristic rules for obtaining knowledge groups from a knowledge graph as described in the present invention.

[0044] Figure 3 This is a flowchart of the method of constructing a question-answer database using entity masks according to the present invention.

[0045] Figure 4 A schematic diagram of the effects of two pre-training tasks on the pre-training model described in the present invention.

[0046] Figure 5 This is a flowchart of the method for generating a question and answer database based on a knowledge graph described in the present invention.

[0047] Figure 6 This is a block diagram of the device for generating a question and answer database based on a knowledge graph according to the present invention. DETAILED DESCRIPTION

[0048] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0049] The method provided by the present invention can be implemented in the following terminal environment, and the terminal may include one or more of the following components: a processor, a memory, and a display screen. The memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the method described in the following embodiment.

[0050] The processor may include one or more processing cores. The processor uses various interfaces and lines to connect various parts of the entire terminal, and executes various functions of the terminal and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory, and calling data stored in the memory.

[0051] The memory may include random access memory (RAM) or read-only memory (ROM). The memory may be used to store instructions, programs, codes, code sets or instructions.

[0052] The display screen is used to display the user interface of each application.

[0053] In addition, those skilled in the art can understand that the structure of the above terminal does not constitute a limitation on the terminal, and the terminal may include more or fewer components, or combine certain components, or arrange the components differently. For example, the terminal also includes components such as a radio frequency circuit, an input unit, a sensor, an audio circuit, and a power supply, which will not be described in detail here.

[0054] Embodiment 1

[0055] In view of the shortcomings of traditional methods, the present invention first provides a method for generating a question and answer library based on a knowledge graph in a first aspect, randomly masks the entity relationship combinations extracted from the knowledge graph, obtains questions using a pre-trained question generation model, and constitutes a question and answer pair with the masked entities or relationships.

[0056] See also Figure 1 The method for generating a question-answer library based on a knowledge graph of the present invention specifically includes:

[0057] Step S1, obtaining entity relationship combinations from the knowledge graph;

[0058] After receiving the knowledge graph, the triples in the knowledge graph are automatically processed to obtain knowledge groups, i.e. entity relationship combinations;

[0059] In a further embodiment, the present invention processes triples in the knowledge graph according to a series of heuristic rules to obtain knowledge groups, and the heuristic rules specifically include one or more of the following three types:

[0060] 1) Templates that match first-degree relationships of entities,

[0061] 2) Templates that match the second-degree relations of entities,

[0062] 3) A template that joins two entities with a first-degree relationship.

[0063] The template represented by the heuristic rule is as follows Figure 2 As shown. Taking the entity "Zhang San" as an example, with the help of rule 1), the first-degree relationship knowledge related to Zhang San can be extracted, such as "Zhang San-nationality-country A" and so on; with the help of rule 2), the second-degree relationship knowledge related to Zhang San can be extracted, such as "Zhang San-wife-Li Si-occupation-actor" and so on; with the help of rule 3), the knowledge that combines two entities and has a first-degree relationship can be extracted, such as "Zhang San-wife-Li Si and Zhang San-occupation-actor, director" and so on.

[0064] Step S2, randomly masking the entities in the entity relationship combination, extracting some of the masked entity relationship combinations, and generating fine-tuning data.

[0065] First, randomly mask the entities in the above full knowledge group, then randomly select a part of the masked knowledge group, and get questions by manually asking users. Finally, use each masked knowledge group and the question asked to it as a piece of data to form fine-tuning data in this way.

[0066] For example Figure 3 As shown, for the knowledge group "Zhang San-Nationality-Country A", during the random masking process, the entity "Country A" is masked to obtain the masked knowledge group "Zhang San-Nationality-?X", and the question "Which country is Zhang San from?" is obtained by manually asking the user, thus forming a fine-tuning data including "Zhang San-Nationality-?X" and "Which country is Zhang San from?".

[0067] Step S3, fine-tuning the pre-trained model obtained by pre-training using the fine-tuning data to obtain a question generation model;

[0068] Specifically, the pre-trained model includes an encoder and a decoder. Based on the generated fine-tuning data, the masked knowledge group in the fine-tuning data is used as the input of the encoder end, and the questions in the fine-tuning data are used as the output of the decoder end, and the pre-trained model is end-to-end fine-tuned to obtain a question generation model.

[0069] For example Figure 3 As shown, the masked knowledge group "Zhang San-nationality-?X" is input into the encoder of the question generation model, and the question "Which country is Zhang San from?" in the fine-tuning data is used as the output of the decoder end of the question generation model. The pre-trained model is end-to-end fine-tuned to obtain the question generation model.

[0070] For the second-degree relationship knowledge in the knowledge graph and the first-degree relationship knowledge that unites two entities, random masking can also be performed in a similar manner to the above to obtain other items of the question generation model.

[0071] Since the questions generated by the question generation model of the present invention are derived from the existing knowledge graph, the data set is ensured to have high accuracy and quality.

[0072] The pre-trained model is based on an existing specific model and uses large-scale corpus to perform entity prediction training to enhance the model's word knowledge (language knowledge) and question generation capabilities. Figure 4 As shown, the present invention designs two pre-training tasks.

[0073] The first task is to generate questions from paragraphs. By using the paragraphs and questions in the reading comprehension corpus, a specific model is trained to generate corresponding questions based on the input paragraphs, so as to improve the question generation ability of the pre-trained model. The reading comprehension corpus contains a large number of paragraph-question pair data sets, where each piece of data contains a paragraph of text and a question that can be answered based on the content of the paragraph of text.

[0074] The second task is to use encyclopedia corpus for pre-training, mask entities marked with hyperlinks in the encyclopedia corpus, and train a specific model to obtain the masked part according to the masked paragraphs, so as to enhance the wordknowledge (language knowledge) ability of the pre-trained model. The encyclopedia corpus is a data set composed of a large number of text paragraphs in the encyclopedia, and each piece of data is selected from the introduction and description of a certain entity in the encyclopedia.

[0075] Specifically, the training method of the pre-training model is as follows:

[0076] Get reading comprehension corpus and encyclopedia corpus;

[0077] Generate a first training data set using paragraphs and questions in the reading comprehension corpus;

[0078] Masking entities marked with hyperlinks in the encyclopedia corpus to generate a second training data set, wherein the second training data set includes the masked paragraphs of the encyclopedia corpus and the masked entities;

[0079] The pre-trained model is obtained by training a specific model using the first training data set and the second training data set, including taking the paragraphs in the first training data set as the model input and the questions as the expected output of the model, and pre-training the model for the paragraph-to-question generation task; taking the masked paragraphs in the second training data set as the input and the masked entities as the expected output, and pre-training the model for the masked entity prediction task.

[0080] The specific model may optionally be a pre-trained language model, such as a BART, T5, GLM or other pre-trained model.

[0081] Step S4, generating a question-answering dataset based on the masked full entity relationship combination and the question generation model.

[0082] According to all the masked knowledge groups obtained in step S2, they are used as input data and input into the encoder end of the question generation model obtained in step S3. For each input masked knowledge group, the decoder end of the question generation model outputs the corresponding question, and the masked entity in the knowledge group is used as the answer to form a question and answer data with the output question. The collection of all the above question and answer data is used as the question and answer data set.

[0083] For example Figure 3 As shown, the masked knowledge group "Zhang San - wife - Li Si - occupation -? X" is input into the encoder of the question generation model, and the decoder of the question generation model outputs "What does Zhang San's wife Li Si do?", thereby obtaining a question-answer data {"id", "1024", "question", "What does Zhang San's wife Li Si do?", "answer", "actor"}.

[0084] Since the question-and-answer dataset is automatically generated based on the model, there is no need for manual input of question-and-answer data, which significantly reduces the cost of dataset construction.

[0085] Embodiment 2

[0086] On this basis, the embodiment of the present invention also proposes a KBQA (knowledge base question answering) method using the above question and answer library. That is, for new questions input by users, it is only necessary to search in the question and answer library generated based on the knowledge graph, and by using the constructed large-scale question and answer corpus, the knowledge-based question and answer task is converted into a corpus retrieval task, and the answer to the question can be obtained without the need for entity recognition, relationship extraction, entity linking and other steps.

[0087] See also Figure 5 The question-answering method based on knowledge graph of the present invention is implemented based on the question-answering library generated by the above embodiment, and specifically includes:

[0088] Step S101, receiving a question input by a user;

[0089] Step S102, searching in a question-answer corpus according to the question to obtain question-answer pairs, wherein the question-answer corpus is generated according to the aforementioned method for generating a question-answer corpus based on a knowledge graph;

[0090] According to the questions input by the user, the question-answering system calculates the sentence similarity through the ES search engine (Elastic Search), performs search and matching in the question-answer corpus generated based on the knowledge graph, and first recalls the K question-answer pairs with the highest similarity; K is an integer and can be adjusted arbitrarily according to user needs;

[0091] In an optional manner, after obtaining K question-answer pairs, the question-answering system reorders the K question-answer pairs using a preset ranking model, such as a ranking model with strong semantic representation capability (such as BERT);

[0092] Step S103: Return the answer to the question and answer pair to the user.

[0093] Preferably, the question-answering system may return the answer to the first-ranked question to the user end as the answer to the user's question.

[0094] It can be seen that the question-answering corpus constructed based on the method of the present invention can realize the KBQA task only through simple database retrieval and matching, which can improve the accuracy of question-answering while improving efficiency.

[0095] Embodiment 3

[0096] like Figure 6 As shown, another aspect of the present invention also relates to a functional device that corresponds to the above-mentioned question-answer library generation method process, that is, an embodiment of the present invention further provides a question-answer library generation device based on a knowledge graph, characterized in that it includes:

[0097] An acquisition module 201 acquires entity relationship combinations from a knowledge graph;

[0098] The processing module 202 randomly masks the entities in the entity relationship combination, extracts some of the masked entity relationship combinations, and generates fine-tuning data;

[0099] A fine-tuning module 203, using the fine-tuning data to fine-tune the pre-trained model obtained by pre-training to obtain a question generation model;

[0100] The generation module 204 generates a question-answering dataset based on the masked full entity relationship combination and the question generation model.

[0101] In a preferred embodiment, the acquisition module 201 further comprises:

[0102] The heuristic rule module is used to process triples in the knowledge graph based on heuristic rules as the entity relationship combination.

[0103] In a preferred embodiment, the heuristic rules further include:

[0104] Templates matching first-degree relationships of entities;

[0105] Templates that match second-degree relationships of entities; and / or

[0106] A template that joins two entities to a first degree relationship.

[0107] In a preferred embodiment, the pre-trained model is pre-trained in the following manner:

[0108] Generate corresponding questions using the mapping relationship between paragraphs and questions in the dialogue question-answering corpus;

[0109] Encyclopedia corpus is used for pre-training, and entities marked with hyperlinks in the encyclopedia corpus are masked.

[0110] In a preferred embodiment, the processing module 202 further includes:

[0111] The question acquisition module is used to obtain questions by manually asking questions for the extracted masked entity relationship combinations;

[0112] The fine-tuning data generation module takes each masked entity relationship combination and the question asked about it as a question-answering data to generate the fine-tuning data.

[0113] In a preferred embodiment, the fine-tuning module 203 further includes:

[0114] The fine-tuning module is used to perform end-to-end fine-tuning on the pre-trained model based on the generated fine-tuning data, using the masked entity relationship combination as the input of the encoder end of the pre-trained model, and using the questions in the fine-tuning data as the output of the decoder end of the pre-trained model to obtain a question generation model that generates questions from the masked entity relationship combination.

[0115] In a preferred embodiment, the generating module 204 further comprises:

[0116] An input and output module, used to input the masked full entity relationship combination into the encoder end of the question generation model, and for each input masked entity relationship combination, the decoder end of the question generation model outputs the corresponding question;

[0117] The data set forming module takes the masked entity in each entity relationship combination as the answer, associates the question and the answer, and forms the question and answer data set.

[0118] The present invention also provides a memory storing a plurality of instructions, wherein the instructions are used to implement the method described in the first embodiment.

[0119] The present invention also provides an electronic device, comprising a processor and a memory connected to the processor, wherein the memory stores a plurality of instructions, and the instructions can be loaded and executed by the processor so that the processor can execute the method described in the first embodiment.

[0120] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for generating a question-answer database based on a knowledge graph, characterized in that: include: Obtain entity relationship combinations from the knowledge graph; Randomly masking the entities in the entity relationship combination, extracting some of the masked entity relationship combinations, and generating fine-tuning data; Using the fine-tuning data, fine-tuning the pre-trained model obtained by pre-training to obtain a question generation model; Generate a question-answering dataset based on the masked full entity relationship combination and the question generation model; The method of fine-tuning the pre-trained model using the fine-tuning data to obtain a question generation model further includes: Based on the generated fine-tuning data, using the masked entity relationship combination as an input of the encoder end of the pre-trained model, using the questions in the fine-tuning data as the output of the decoder end of the pre-trained model, and performing end-to-end fine-tuning on the pre-trained model to obtain a question generation model that generates questions from the masked entity relationship combination; The step of generating a question-answering dataset based on the masked full entity relationship combination and the question generation model further includes: The masked full entity relationship combination is input into the encoder end of the question generation model. For each input masked entity relationship combination, the decoder end of the question generation model outputs the corresponding question, using the masked entity in each entity relationship combination as the answer, and associating the question and answer to form the question and answer dataset.

2. The method according to claim 1, characterized in that The obtaining of entity relationship combinations from the knowledge graph further includes: Processing triples in the knowledge graph based on heuristic rules as entity relationship combinations; The heuristic rules include: Templates matching first-degree relationships of entities; Templates that match second-degree relationships of entities; and / or A template that joins two entities to a first degree relationship.

3. The method according to claim 1, characterized in that The pre-trained model is pre-trained in the following manner: Get reading comprehension corpus and encyclopedia corpus; Generate questions using the paragraphs and answers in the reading comprehension corpus to obtain a first training data set; Performing mask prediction on entities marked with hyperlinks in the encyclopedia corpus to obtain a second training data set; The pre-trained model is obtained by pre-training a specific model using the first training data set and the second training data set; the specific model is BART, T5 or GLM.

4. The method according to claim 1, characterized in that: Generating fine-tuning data further includes: For the extracted masked partial entity relationship combinations, questions are obtained by manually asking questions, and then each masked entity relationship combination and the question asked about it are used as a question-and-answer data to generate the fine-tuning data.

5. A question-answering method based on knowledge graph, characterized in that: include: Receive user input questions; Searching a question-answer corpus according to the question to obtain a question-answer pair, wherein the question-answer corpus is generated based on a knowledge graph according to any one of the methods described in claims 1 to 4; The answer in the question-answer pair is returned to the user.

6. A device for generating a question-answer database based on a knowledge graph, characterized in that: include: The acquisition module obtains entity relationship combinations from the knowledge graph; A processing module randomly masks the entities in the entity relationship combination, extracts some of the masked entity relationship combinations, and generates fine-tuning data; A fine-tuning module, using the fine-tuning data to fine-tune the pre-trained model obtained by pre-training to obtain a question generation model; A generation module generates a question-answering dataset based on the masked full entity relationship combination and the question generation model; The fine-tuning module further comprises: A fine-tuning module is used to perform end-to-end fine-tuning on the pre-trained model based on the generated fine-tuning data, using the masked entity relationship combination as an input of the encoder end of the pre-trained model, and using the questions in the fine-tuning data as the output of the decoder end of the pre-trained model to obtain a question generation model that generates questions from the masked entity relationship combination; The generating module further comprises: An input and output module, used to input the masked full entity relationship combination into the encoder end of the question generation model, and for each input masked entity relationship combination, the decoder end of the question generation model outputs the corresponding question; The data set forming module takes the masked entity in each entity relationship combination as the answer, associates the question and the answer, and forms the question and answer data set.

7. An electronic device, characterized in that: The method comprises a processor and a memory, wherein the memory stores a plurality of instructions, and the processor is used to read the instructions and execute the method according to any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be read by a processor to execute the method according to any one of claims 1 to 5.

Images