A question answering method and system based on deep learning and a knowledge graph

Abstract

The application provides a question and answer method and system based on deep learning and a knowledge graph, which comprises the following steps: constructing a question and answer training corpus containing a question of a head entity and a relationship path from the question to an answer for a triple (e head , r, e tail ) in the knowledge graph, training a question-relation relationship identification model using the corpus, wherein e head represents the head entity, e tail represents the tail entity, and r represents the relationship between the head entity and the tail entity; inputting a question of a user into the relationship identification model, selecting the relationship r q→a with the highest probability as the path from the question to the answer; obtaining an entity e q in the question using an entity matching model comprising a direct matching mode and a fuzzy matching mode; obtaining the entity e q from the entity matching model and the relationship r q→a from the relationship matching model; finding the tail entity e a of the corresponding triple in the knowledge graph according to the entity obtained from the entity matching model and the relationship r q→a obtained from the relationship matching model; and determining the final answer and returning the final answer to the user. The application can effectively improve the accuracy of artificial intelligence question and answer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning and question answering system, and particularly relates to a question answering method and system based on deep learning and knowledge graph. BACKGROUND

[0002] Question answering system is a high-level form of information retrieval system, aiming at answering the question raised by user in natural language with accurate and concise natural language. However, the traditional search engine requires high information retrieval ability of user, and the result returned by the traditional search engine is usually unsatisfactory. With the continuous development of computer technology, it is inevitable that the traditional search engine is replaced by higher-level technology.

[0003] Deep learning is a new research direction in the field of machine learning, which is introduced into machine learning to make it closer to the original goal-artificial intelligence: enabling machines to have analysis and learning ability like human beings. Knowledge graph technology describes the concept entity and its relationship in the objective world in a structured form, expresses the information on the Internet in a form closer to human cognition of the world, and provides a better ability to organize, manage and understand the massive information on the Internet. Knowledge graph, together with big data and deep learning, has become one of the core driving forces for the development of Internet and artificial intelligence.

[0004] With the increase of data volume, the question answering system that can help people quickly and accurately obtain information is emerging in the face of the combination of deep learning technology and knowledge graph technology. Nowadays, more and more colleges and companies are involved in the research of question answering system technology, for example, Microsoft's chat robot Xiaoice, Ali's intelligent customer service Xiaohumi, Apple's voice assistant Siri, etc. have gradually entered people's daily life. At present, question answering system has become a research direction in the field of artificial intelligence and natural language processing, which is widely concerned and has broad development prospects. SUMMARY

[0005] The problem to be solved by the present application is to improve the accuracy of question answering by using deep learning technology and knowledge graph technology.

[0006] The present application provides a question answering method based on deep learning and knowledge graph, comprising the following steps,

[0007] Step 1, for the (e head ,r,e tail ) triple in the knowledge graph, constructing a question answering training corpus containing the question of the head entity and the relationship path from the question to the answer, and using the corpus to train the "question-relation" relation identification model, wherein e head represents the head entity, e tail represents the tail entity, and r represents the relationship between the head entity and the tail entity;

[0008] Step 2, input the user's question to the relation recognition model, select the highest probability relation r q→a As the path from the question to the answer;

[0009] Step 3, use the entity matching model including direct matching method and fuzzy matching method to obtain the entity e in the question q ;

[0010] Step 4, according to the entity matching model e q The entity obtained by the relation matching model r q→a Find the tail entity e a of the corresponding triple in the knowledge graph, determine the final answer and return it to the user.

[0011] Moreover, the implementation process of step 1 is as follows,

[0012] Create a knowledge graph, pre-set different entity types under the knowledge graph, and construct entity relationship triples under different types (e head ,r,e tail );

[0013] According to the relationship between entity types in the knowledge graph, construct (Q, R) corpus templates, where the question Q contains an entity type in the knowledge graph, and the relationship R is the relationship path from the question Q to the correct answer A;

[0014] Extract (e head ,r,e tail ) triples in the knowledge graph, fill the head entity e head under the same entity type and the relationship r corresponding to the head entity into the corpus template (Q, R), and finally obtain the question and answer dialogue training data set;

[0015] Use the data set to train the "question-relation" relation recognition model.

[0016] Moreover, the implementation process of step 2 is as follows,

[0017] Carry out preprocessing operations such as word segmentation on the user's input question, and then input it to the relation recognition model;

[0018] Select the most probable candidate relationship path r q→a as the real relationship path from the question to the answer.

[0019] Moreover, the implementation process of step 3 is as follows,

[0020] Directly match the entity e qThe implementation is to extract all entities in the knowledge graph, traverse each entity, and determine whether the entity is a substring of the question. If one of the entities is a substring of the question, the entity is used as the question entity. If the direct matching is successful, step 4 is directly executed.

[0021] The entity e in the fuzzy matching question q The implementation is as follows,

[0022] The relationship r obtained according to step 2 q→a Query the triples in the knowledge graph to obtain all head entity sets in the triples that meet the condition of the relationship being r q→a in the knowledge graph.

[0023] The question is segmented.

[0024] For each segment of the question, it is determined whether the segment is a substring of any entity in the head entity set. If a segment is determined to be a substring of an entity, the entity is used as the question entity.

[0025] Moreover, the implementation process of step 4 is as follows,

[0026] The question entity e obtained according to step 3 q and the relationship path r obtained according to step 2 q→a Query the corresponding triples in the knowledge graph, and the query method is as follows:

[0027] e a =KG(e q ,r q→a )

[0028] Wherein, KG refers to the knowledge graph constructed in the relationship recognition model construction process, in the knowledge graph KG, according to the given head entity and relationship, the tail entity e a is queried, and then the final reply is generated in combination with the template and returned to the user.

[0029] Moreover, it is used to support human-computer question and answer or information retrieval services.

[0030] On the other hand, the application provides a question and answer system based on deep learning and knowledge graph, which is used to realize the question and answer method based on deep learning and knowledge graph as described above.

[0031] Moreover, it includes a processor and a memory, the memory is used to store program instructions, and the processor is used to call the storage instructions in the memory to execute the question and answer method based on deep learning and knowledge graph as described above.

[0032] Alternatively, a readable storage medium is included, and the readable storage medium stores a computer program. The computer program, when executed, implements the deep learning and knowledge graph-based question answering method described above.

[0033] The application provides a deep learning and knowledge graph-based question answering method and system. First, knowledge in a knowledge graph is extracted and filled into a pre-constructed "question-relation" template to obtain a question answering training data set. A "question-relation" relation identification model is trained using the data set. Then, a question is input into the relation identification model to obtain a relation path from the question to an answer. Second, an entity matching model combining a direct matching method and a fuzzy matching method is used to obtain entities in the question. Finally, the entities matched by the entity matching model and the relations identified by the relation identification model are used to query corresponding triples in the knowledge graph to obtain a final answer and return the final answer to a user. According to the question answering method provided by the application, a deep learning model is used to predict the relation from the question to the answer. That is, triples in the knowledge graph are filled into a question answering corpus template to expand the scale of the question answering corpus at a small cost, avoid the problem that a deep learning model cannot be fully learned due to data sparseness, and effectively improve the accuracy of the relation identification model in predicting the relation path from the question to the answer. In addition, the entity matching model combining the direct matching method and the fuzzy matching method is used to obtain entities in the question. The direct matching is responsible for matching entities in a complete form, and the fuzzy matching is responsible for matching entities in a brief form. The entity matching model can effectively complete the entity matching task and plays a crucial role in successfully replying to the system. The entities and relations obtained by the entity matching model and the relation identification model are used to query corresponding triples in the knowledge graph, and the answer can be quickly and accurately obtained. The application can effectively improve the accuracy of the intelligent reply of the system, can be applied to the market, can abandon the time-consuming information retrieval process, can reply to the inquiry of the user with high efficiency and high accuracy, and has great market value and research significance. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 A construction flowchart of the relation identification model provided by the embodiment of the application is shown in the figure.

[0035] Figure 2 A step flowchart of the deep learning and knowledge graph-based question answering method provided by the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0036] In order to more clearly and explicitly highlight the technology and advantages of the application, the application is described more completely and in detail in combination with the drawings and embodiments.

[0037] The embodiment of the application provides a deep learning and knowledge graph-based question answering method, which comprises the following steps.

[0038] (1) Relationship identification model training:

[0039] This step is aimed at the (e head ,r,e tail ) triples in the knowledge graph, and builds a question-answer training corpus containing the question of the head entity e head and the relationship path r from the question to the answer, and uses the corpus to train the "question-relation" relationship identification model.

[0040] Further, the implementation of this step is as follows:

[0041] Create a knowledge graph, pre-set different entity types under the knowledge graph, and build entity relationship triples (e head ,r,e tail ) under different types, where e head represents the head entity, e tail represents the tail entity, and r represents the relationship between the head entity and the tail entity.

[0042] According to the relationship between the entity types in the knowledge graph, build a (Q, R) corpus template, where the question Q contains an entity type in the knowledge graph, and the relationship R is the relationship path from the question Q to the correct answer A.

[0043] Extract (e head ,r,r tail ) triples from the knowledge graph, fill the head entity e head under the same entity type and the relationship r corresponding to the head entity into the corpus template (Q, R), and finally obtain the question-answer dialogue training dataset.

[0044] Use the dataset to train the "question-relation" relationship identification model.

[0045] (2) Relationship identification:

[0046] This step inputs the user's question into the relationship identification model, and selects the most probable relationship r q→a as the path from the question to the answer.

[0047] Further, the implementation of this step is as follows:

[0048] Perform preprocessing operations such as word segmentation on the user's input question, and then input it into the relationship identification model.

[0049] Select the most probable candidate relationship path r q→a as the true relationship path from the question to the answer.

[0050] (3) Entity matching:

[0051] The entity matching model including the direct matching mode and the fuzzy matching mode is used to obtain the entity e in the question in this step q .

[0052] Further, the implementation of this step is proposed as follows:

[0053] The entity matching model combining the direct matching mode and the fuzzy matching mode is used to extract the entity e in the question q , that is, the direct matching mode is used first, and then the fuzzy matching mode is used to obtain the entity in the question if the matching entity fails.

[0054] The direct matching mode is as follows:

[0055] All entities in the knowledge graph are extracted, and each entity is traversed to determine whether the entity is a substring of the question to obtain the entity in the question;

[0056] The entity e in the question is directly matched q , that is, all entities in the knowledge graph are extracted, and each entity is traversed to determine whether the entity is a substring of the question, if one of the entities is a substring of the question, the entity is taken as the question entity. If the direct matching is successful, step (4) is directly executed.

[0057] The fuzzy matching mode is as follows:

[0058] According to the relationship r identified in (2) q→a , the triplets in the knowledge graph are queried to obtain all head entity sets in the triplets with the relationship r q→a in the knowledge graph;

[0059] The question input by the user is segmented;

[0060] For each segmented word, it is determined whether the segmented word is a substring of any entity in the head entity set, so as to obtain the entity in the question;

[0061] (4) Answer determination:

[0062] The entity e obtained by the entity matching model q and the relationship r obtained by the relationship matching model q→a are used to find the tail entity e of the corresponding triplet in the knowledge graph a , and the final answer is determined and returned to the user.

[0063] Further, the implementation of this step is proposed as follows:

[0064] The question entity e obtained in step 3 q and the relationship path r obtained in step 2 q→aIn the knowledge graph, the corresponding triple is queried in the following way:

[0065] e α =KG(e q ,r q→a )

[0066] Wherein, KG refers to the knowledge graph constructed in the process of constructing the relationship recognition model, in the knowledge graph KG, according to the given head entity and relationship, the tail entity e a is queried to obtain, then the final reply is generated in combination with the template and returned to the user.

[0067] That is, according to the question entity e q obtained by the entity matching model and the relationship path r q→a obtained by the relationship recognition model, the triple in the knowledge graph is queried, which is consistent with the head node e q , the relationship r q→a , and the answer entity e a is obtained, then the final reply is generated in combination with the template and returned to the user.

[0068] As shown in Figure 1 , for the purpose of implementation reference, the construction process of the relationship recognition model of the embodiment is provided, including the following steps:

[0069] Step 1, constructing a knowledge graph:

[0070] Step 1.1, classifying the entities in the experimental data set, entities of the same type have the same category label, for example, under the entity type of "course name", there are entities such as "discrete mathematics" and "college English"; under the entity type of "teacher", there are entities such as "Zhang San" and "Li Si"; under the entity type of "online learning resources", there are entities such as "discrete mathematics excellent course" and "college English excellent course";

[0071] Step 1.2, creating an entity relationship triple in the form of "head entity-relation-tail entity" according to the format (e head , r, e tail ), which represents the connection between entities and entities. For example, there is a relationship "teaching teacher" between the entity "discrete mathematics" and the entity "Zhang San", that is, it constitutes an entity relationship triple (discrete mathematics, teaching teacher, Zhang San);

[0072] Step 2, filling the knowledge in the knowledge graph into the template to obtain training corpus:

[0073] Step 2.1: Based on the relationships between entity types in the knowledge graph obtained in Step 1, construct a (Q,R) corpus template, where question Q contains an entity type from the knowledge graph, and relation R is the path from that entity type to the correct answer. A partial template of the question-and-answer corpus using the course name "Course Name" and the relation "Instructor" as an example is shown in Table 1:

[0074] Table 1

[0075]

[0076] Step 2.2: Extract all entities from the knowledge graph, and replace the <entity type> in the construction template with all entities under that entity type, finally obtaining the question-answering dialogue training dataset. Taking the entity and relation "instructor" under the course name "course name" as an example, a partial dataset is shown in Table 2:

[0077] Table 2

[0078]

[0079] Step 3, Corpus Preprocessing:

[0080] Step 3.1: Using the word segmentation tool provided by BERT, first segment the question Q to obtain the word segmentation list Q. x =[[CLS],x1,x2,...,x i ,...,x m [SEP]], where m represents the total number of words in the question text, and x i The i-th word in the text is represented by [CLS], where i = 1, 2, ..., m. The [CLS] flag is added at the beginning of the sentence and the representation vector obtained by the BERT model is used for subsequent relation classification tasks. The [] flag is added at the end of the sentence to separate sentences.

[0081] BERT stands for Bidirectional Encoder Representation from Transformers. It is a pre-trained language representation model. For specific implementation details of BERT, please refer to existing technology: Devlin, J., Chang, MW, Lee, K., & Toutanova, K. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805. This invention will not elaborate further.

[0082] Step 3.2, process the segmented text Qx Using word embedding technology to process, get the input BERT encoding vector sequence Q e = [e cls ,e1,e2,...,e i ,...,e m ,e sep ],e i i represents the word vector of the i-th word segmentation, e cls and e sep represent the word vectors corresponding to the identifier [CLS] and the separator [].

[0083] Step 4, input into the BERT model:

[0084] Input the word vector Q e into the BERT model, and obtain the feature vector representation of each word in the text Q h = [h cls ,h1,h2,…h i ,…,h m ,h sep ],h cls , h sep respectively represent the outputs corresponding to the identifier [CLS] and the separator [SEP], and h i represent the output of the word vector x i through BERT;

[0085] Step 5, through the full connection layer and the sigmoid function, the relationship probability distribution is obtained according to the representation of [CLS], and the specific implementation process is as follows:

[0086] p(r q→a ) = sigmoid(Wh cls )

[0087] Where W is a matrix representing the trainable parameters of the full connection layer, and sigmoid represents the activation function, which can map a numerical value to (0, 1). The relationship probability distribution p(r q→a ) is obtained by using the activation function, and the relationship r q→a with the maximum probability is selected as the predicted relationship path of the question Q to the answer A.

[0088] The full connection layer, i.e. each node of the current layer is connected with each node of the last layer, fuses all the features extracted by the last layer. The implementation of the full connection network layer can be referred to Svozil D, Kvasnicka V, Pospichal J. Introduction to multi-layer feed-forward neural networks[J]. Chemometrics and intelligent laboratory systems, 1997, 39(1): 43-62. The present application does not elaborate.

[0089] Step 6, judging whether the predicted relationship path is the real relationship path from the question to the answer:

[0090] For each question-answer data (Q, R) in the obtained question-answer dialogue training data of step 2, steps 3-5 are performed on the question Q to obtain the predicted relationship r q→a , which is compared with the real relationship R to iteratively optimize the parameters of the relationship identification model, and finally obtain the relationship identification model capable of predicting the relationship path from the input user question to the answer.

[0091] As Figure 2 shown, for the purpose of implementation and reference, the specific process of the embodiment of the deep learning and knowledge graph-based question-answering method is provided, including the following steps:

[0092] Step 1, relationship identification:

[0093] Step 1.1, receiving the input question Q of the user, performing steps 3-5 of the construction process of the relationship identification model on Q to obtain the probability distribution of the relationship from the question to the answer. At this time, the relationship identification model used is a deep learning model optimized by the question-answer training data set;

[0094] Step 1.2, selecting the candidate relationship path r q→a with the largest probability as the real relationship path from the question Q to the answer A;

[0095] Step 2, entity identification:

[0096] Step 2.1, directly matching the entity e q in the question, i.e. extracting all the entities in the knowledge graph, traversing each entity to determine whether the entity is a substring of the question, if one of the entities is a substring of the question, the entity is the question entity. If the direct matching is successful, step 3 is directly performed;

[0097] The purpose of direct matching is to match the entity in the question that exists in the knowledge graph in its entirety, for example, in the question "Who is the teacher of Discrete Mathematics?", the direct matching method can obtain the entity "Discrete Mathematics" that exists in the knowledge graph in its entirety.

[0098] Step 2.2, fuzzy matching entity e in the question q The specific implementation steps are as follows:

[0099] Step 2.2.1, the relationship path from the question to the answer obtained in step 1 is r q→a The query range of the question entity is limited to all head entities in the triplets in the knowledge graph that satisfy the relationship r q→a .

[0100] The purpose of this step is to avoid incorrect query ranges of the question entity, because only when the relationship r q→a exists between the question entity and other entities in the knowledge graph, can the corresponding triplets be queried in the knowledge graph; otherwise, the query cannot be performed. For example, in the question "Who is the teacher of Mathematics?", the entity "Discrete Mathematics Course" is not within the query range of the question entity, because the relationship path from the question to the answer obtained through the relationship matching model is "teacher", and in the knowledge graph, the entity "Discrete Mathematics Course" does not have the relationship "teacher" with other entities, so it is not within the query range, avoiding the matching to an incorrect question entity.

[0101] Step 2.2.2, tokenizing the question Q;

[0102] Step 2.2.3, using the set of head entities obtained in step 2.2.1, for each token of the question, it is determined whether the token is a substring of any entity in the set of head entities, if a token is determined to be a substring of an entity, the entity is taken as the question entity;

[0103] The purpose of fuzzy matching is to solve the situation that direct matching mode cannot match: the entity appearing in the question in a brief form. For example, in the question "Who is the teacher of mathematics?", "mathematics" is a brief form of the entity "discrete mathematics" in the knowledge graph, and the entity "discrete mathematics" expected by the user cannot be matched using the direct matching mode; in the fuzzy matching mode, first, the relationship path to the answer is obtained through the relationship matching model, which is "teacher", second, the search range of the question entity is determined: all head entities in the triple that satisfy the relationship "teacher" in the knowledge graph, and because the entity "discrete mathematics" exists in the head entity set, the entity "discrete mathematics" is preliminarily determined as a candidate question entity; then the question is segmented, and the segmentation "mathematics" is obtained; finally, it is judged that the segmentation "mathematics" is a substring of the entity "discrete mathematics", so that the question entity "discrete mathematics" is obtained.

[0104] Step 3, answer determination:

[0105] According to the question entity e obtained in step 2 q and the relationship path r obtained in step 1 q→a , the corresponding triple in the knowledge graph is queried, and the query method is as follows:

[0106] e a =KG(e q ,r q→a )

[0107] Wherein, KG refers to the knowledge graph constructed in the relationship recognition model construction process, in the knowledge graph KG, according to the given head entity and relationship, the tail entity e a is queried, and then the final reply is generated combined with the template and returned to the user.

[0108] The above method can realize a question and answer method and system based on deep learning and knowledge graph.

[0109] In specific implementation, the method proposed in the technical scheme of the application can realize automatic operation process by computer software technology, and the system device of the method, such as computer readable storage medium storing the corresponding computer program of the technical scheme of the application and computer equipment and server including running the corresponding computer program, should also be within the protection scope of the application.

[0110] In some possible embodiments, a question and answer system based on deep learning and knowledge graph is provided, including a processor and a memory, the memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute the question and answer method based on deep learning and knowledge graph as described above.

[0111] In some possible embodiments, a deep learning and knowledge graph based question answering system is provided, comprising a readable storage medium, and a computer program is stored on the readable storage medium, and the computer program is executed to implement the deep learning and knowledge graph based question answering method as described above.

[0112] The specific embodiments described herein are merely illustrative of the spirit of the application. Various modifications or supplements or replacements of the described specific embodiments or similar ways can be made by those skilled in the art in the technical field of the application, without departing from the spirit of the application or exceeding the scope defined by the appended claims.

Claims

1.A method for question answering based on deep learning and a knowledge graph, characterized in that, The method comprises the following steps, Step 1, for the triples in the knowledge graph , construct a question and answer training corpus containing the question of the head entity and the relationship path of the question to the answer, and train the "question-relation" relation identification model using the corpus, wherein represents the head entity, represents the tail entity, represents the relationship between the head entity and the tail entity; The implementation process is as follows, Creating a knowledge graph, pre-setting different entity types under the knowledge graph, and constructing different types of entity relationship triplets; According to the relationship between the entity types in the knowledge graph, a corpus template is constructed , wherein the question contains an entity type in the knowledge graph, the relationship is the relationship path from the question to the correct answer . Extract the triples in the knowledge graph , fill the head entity of the same entity type , and the relationship corresponding to the head entity into the corpus template , and finally obtain the question and answer dialogue training data set; The "question-relation" relation recognition model is trained by using the data set; Step 2, input user's question to relation identification model, select the highest probability relation As a path from question to answer; Step 3, obtaining the entity in the question using an entity matching model including a direct matching mode and a fuzzy matching mode ; The implementation process is as follows, Entities in direct matching problems The implementation is to extract all entities in the knowledge graph, traverse each entity, judge whether the entity is a substring of the question, if one of the entities is a substring of the question, the entity is taken as a question entity, and if the direct matching is successful, step 4 is directly executed. Entities in fuzzy match problems The implementation is as follows, According to the relationship obtained in step 2 , query the triples in the knowledge graph to obtain all head entity sets in the triples with the relationship in the knowledge graph; The question is segmented; For each segmented word of the question, it is determined whether the segmented word is a substring of any entity in the head entity set, and if the segmented word is determined to be a substring of an entity, the entity is taken as a question entity; Step 4, entities obtained according to the entity matching model and relations obtained according to the relation matching model In the knowledge graph, find the tail entity corresponding to the triple Determine the final answer and return it to the user. 2.The method of claim 1, wherein: The implementation process of step 2 is as follows, The segmented pre-processing operation is performed on the question input by the user, and then the question is input into the relation recognition model; selecting the candidate relation path with the highest probability as the true relation path from the question to the answer , as the true relation path from the question to the answer 3.The method of claim 1, wherein: The implementation process of step 4 is as follows, Question entity according to step 3 and the relationship path obtained in step 2 Query the corresponding triple in the knowledge graph, and the query method is as follows: Wherein, The knowledge graph is constructed in the relationship identification model construction process, and the knowledge graph is used to query the tail entity according to the given head entity and relationship Then, the final reply is generated in combination with the template and returned to the user.​ 4.The method of claim 1 or 2 or 3, characterized in that: The method is used for supporting man-machine question and answer or information retrieval service. 5.A question answering system based on deep learning and knowledge graph, characterized in that: The method is used for implementing the question and answer method based on deep learning and knowledge graph. 6.The question answering system based on deep learning and knowledge graph of claim 5, characterized in that: The method comprises a processor and a memory, the memory is used for storing program instructions, and the processor is used for calling the stored instructions in the memory to execute the question and answer method based on deep learning and knowledge graph. 7.The question answering system based on deep learning and knowledge graph of claim 5, characterized in that: The method comprises a readable storage medium, and the readable storage medium stores a computer program, and the computer program is executed to implement the question and answer method based on deep learning and knowledge graph.

Images