A complex question-answering method and system applied to mineral knowledge graphs

Abstract

The present invention discloses a complex question-answering method and system for mineral knowledge graphs, comprising: constructing a complex mineral question-answering dataset based on an existing simple mineral question-answering dataset; constructing a text representation model based on the existing mineral knowledge graph and dataset; utilizing the text representation model and dataset to construct a mineral question center word recognition, entity disambiguation, and reasoning model; using the trained mineral question center word recognition and entity disambiguation model to perform center word recognition and entity disambiguation on user-input questions, locate the reasoning starting entity, and use the reasoning model to infer the answer from the mineral knowledge graph using the entity as the starting point and return it to the user. The method and system provided by the present invention solve the problem that the current mineral knowledge graph can only perform simple question-answering with single-hop reasoning and cannot answer complex questions involving multi-hop relationships. It achieves accurate answers to mineral questions input in natural language and can be used for intelligent query of mineral knowledge graphs.

Description

Technical Field

[0001] The present invention relates to the field of data processing technology, and in particular to a complex question-answering method and system applied to a mineral knowledge graph. Background Art

[0002] The mineral knowledge graph stores a vast amount of mineralogical knowledge. It consists of multiple facts, each composed of a head entity, a relationship, and a tail entity. These facts effectively demonstrate the connections and property characteristics between minerals. Currently, common mineral knowledge graph question-and-answer systems can only answer simple questions about a single fact. However, with the growth and complexity of mineral data, this simple approach can no longer meet the needs of geoscientists for mineral knowledge acquisition. Complex question-and-answer systems are now required, requiring multiple intermediate steps and reasoning across multiple facts to provide accurate answers. Summary of the Invention

[0003] In view of this, it is necessary to build a mineral knowledge question-answering system based on the latest knowledge graph used to organize and store mineral knowledge, which can perform complex reasoning and answer questions input by users in natural language, so as to solve the current situation where the mineral question-answering system can only answer simple questions involving single-hop reasoning but cannot answer complex questions involving multi-hop reasoning.

[0004] A first aspect of the present invention provides a multi-hop reasoning-based question-answering method and system for a mineral knowledge graph, the method comprising:

[0005] Step 1: Construct a complex mineral question-answering dataset based on the existing simple mineral question-answering dataset;

[0006] Step 2: Represent the mineral text as a word vector model and a sentence vector model. The word vector model uses the ComplEx model to embed entities and relations into a d-dimensional complex space. The sentence vector model first uses the Bert model, adds [CLS] at the beginning of the question, and then uses the Bert output corresponding to the question as the initial sentence vector of the question. The initial sentence vector is input into the stacked fully connected layer and projected into the complex space to obtain the final sentence embedding vector W. q ;

[0007] Step 3: Perform BIO annotation on the mineral question-answering dataset. Each natural language question is converted into a sentence vector and mapped into a label sequence consisting of the three letters B, I, and O with the same length as the sentence vector. The Bert-BiLSTM-CRF model is trained using the negative log-likelihood loss function to obtain a model that can identify the central word consisting of B and I in the question.

[0008] Step 4: For the identified central word of the question, use the two methods of candidate entity generation based on edit distance algorithm and word segmentation to jointly filter out the candidate entities most relevant to the central word in the knowledge graph entity set;

[0009] Step 5: Record the candidate entity most relevant to the central word as the initial entity e0, and obtain all the first-degree outgoing relationship sets of e0 from the knowledge graph. Find the relationship r1 that best reflects the relationship between the question and entity e0 from the relationship set, and then locate entity e1 from the knowledge graph based on entity e0 and relationship r1. Then repeat the above process starting from entity e1 until r n = <stop>, stop reasoning;

[0010] Preferably, step 1 "constructing a complex mineral question-answering dataset based on an existing simple mineral question-answering dataset" specifically includes:

[0011] Based on the existing single-hop dataset, a multi-hop question dataset is constructed by expanding the single-hop question content. The details are as follows: first, the entity e0 contained in the single-hop question is retrieved in the knowledge graph, and then the retrieved related triples are<e0,r1,e1> The e0 in the question is replaced with the corresponding question words such as "what" and "where" to modify the single-hop question to obtain a two-hop question, and then the triples can be further retrieved.<e0,r1,e1> Another entity e1 in the<e1,r2,e2> The e1 in the sentence is replaced with the corresponding question words such as "what" and "where", and then the two-hop question is modified to obtain a three-hop question. In order to enable the question answering system to handle more diverse questions, the generated questions are processed through synonym conversion, sentence reconstruction, and Chinese-English translation to increase the number of questions and improve the generalization ability of the model. Each data contains the question text and the triple sequence e0,r1,e1,r2,e2,...,e required to answer this question. i ,r i+1 ,e i+1 ,...,e n ;

[0012] Preferably, step 2 represents the mineral text as a word vector model and a sentence vector model. The word vector model uses the ComplEx model to embed entities and relations into a d-dimensional complex space. The sentence vector model first uses the Bert model, adds [CLS] at the beginning of the question, and then uses the Bert output corresponding to the question as the initial sentence vector of the question. The initial sentence vector is input into the stacked fully connected layer and projected into the complex space to obtain the final sentence embedding vector W. q Specifically include:

[0013] Entity embedding is a d-dimensional vector, and relation embedding is a d-dimensional matrix. The knowledge triple (h, r, t) obtained from the knowledge graph during training is called a positive triple, where h is the head entity, r is the relation, and t is the tail entity. Then, a random entity t* is obtained from the knowledge triple, and it is combined with h and r to form a negative triple (h, r, t*), where t* is required to be not equal to t. The embedding vectors / matrices corresponding to h, r, t, and t* are e respectively. h , W r , e t , e t* , use the ComplEx scoring formula shown in formula (1) to calculate the scores of positive and negative triples and use the loss function shown in formula (2) for training, that is, maximize the score of positive triples as much as possible and minimize the score of negative triples, so that the model can better distinguish between positive and negative triples. In formula (1), Re represents the real part of the complex number. Represents e t The complex conjugate operation of ;

[0014]

[0015] According to the inference path e0,r1,e1,r2,e2,...,e i ,r i+1 ,e i+1 ,...,e n Extract e0 as e h , e1,...,e n As positive examples e t , and then randomly obtain entities from the knowledge graph as negative examples e t* and e t* Not equal to e t , define the constraint functions shown in formulas (3) and (4) and the loss function shown in formula (5) for training, where ζ represents the set of all entities on the reasoning path belonging to the question, y is the target value, the target value of the positive example is set to 1, and the target value of the negative example is set to 0, σ is the sigmoid function, Re represents the real part of the complex number, Represents e t The complex conjugate operation of ;

[0016]

[0017] Preferably, step 4 "using two methods, one based on edit distance algorithm and one based on word segmentation, to jointly screen out candidate entities related to the central word in the knowledge graph entity set for the identified central word of the question" specifically includes:

[0018] The candidate entity vectors are concatenated with the question vector and then fed into a fully connected layer. The Softmax activation function is used to output the probability that each candidate entity belongs to the set of entities truly related to the central word. The entity with the highest probability is the disambiguated entity and serves as the starting point for relational reasoning. The cross-entropy loss function is used during training.

[0019] Preferably, step 5 records the candidate entity most relevant to the central word as the initial entity e0, obtains all first-degree outgoing relationship sets of e0 from the knowledge graph, finds the relationship r1 that best reflects the relevance between the question and entity e0 from the relationship set, and then locates entity e1 from the knowledge graph based on entity e0 and relationship r1, and then repeats the above process starting from entity e1 until r n = <stop>, stop reasoning "specifically includes:

[0020] During training, for the training problem q in the data set and its reasoning path e0,r1,e1,r2,e2,...,e i ,r i+1 ,e i+1 ,...,e n , the mineral problem q, r i The pre-inference relation r1-r i-1 With entity e i-1 Use word embedding model to convert to vector representation and <stop>The identifier is added to all one-hop outbound relations of the current entity as candidate relations. The question vector Q is concatenated with the relation vector r1 and then input into the fully connected layer to output a new vector Q'. Q' is concatenated with the relation vector r2 and then input into the fully connected layer to output a new vector Q", and so on. The question vector Q is concatenated with all relations up to r i-1 The preceding reasoning relationship is concatenated one by one and processed by the fully connected layer before being combined with the entity e i-1 and relationship r i After splicing, input the next fully connected layer to get r i The semantic matching probability that can be correctly inferred is that the relationship with the highest matching probability is the direction of the next hop. If <stop>If the identifier has the highest matching probability, it means that the final answer has been inferred, and further reasoning is stopped. The cross entropy loss function is used during training;

[0021] After training the above model, the user's input question is first converted into a sentence vector, and then the mineral question center word recognition model parses the center word in the question. After disambiguation by the mineral entity disambiguation model, the initial entity for reasoning is obtained. After multi-hop reasoning by the reasoning model, the reasoning answer is obtained and returned to the user.

[0022] The above-mentioned complex question-answering method and system applied to the mineral knowledge graph solves the problem that the current mineral knowledge graph can only perform simple question-answering with single-hop reasoning and cannot answer complex questions involving multi-hop relationships. It improves the ability to answer various forms of natural language mineral questions, realizes accurate answers to mineral questions input in natural language, and can be used for intelligent query of mineral knowledge graphs. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a flow chart of a complex question-answering method applied to a mineral knowledge graph provided by a preferred embodiment of the present invention.

[0025] Figure 2 This is the text representation word vector model training process provided by the preferred embodiment of the present invention

[0026] Figure 3 It is a text representation sentence vector model provided by the preferred embodiment of the present invention

[0027] Figure 4 The mineral question core word recognition model provided by the preferred embodiment of the present invention

[0028] Figure 5 The mineral entity disambiguation model provided by the preferred embodiment of the present invention is

[0029] Figure 6 The mineral reasoning model provided by the preferred embodiment of the present invention

[0030] Figure 7 This is the interface of the mineral question-answering system provided by the preferred embodiment of the present invention DETAILED DESCRIPTION

[0031] In order to make the purpose, features, and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described below are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, an embodiment of the present invention provides a complex question-answering method and system applied to a mineral knowledge graph, comprising the following steps:

[0033] S1 builds a complex mineral question-answering dataset based on the existing simple mineral question-answering dataset

[0034] Based on the existing single-hop dataset, a multi-hop question dataset is constructed by expanding the single-hop question content. The details are as follows: first, the entity e0 contained in the single-hop question is retrieved in the knowledge graph, and then the retrieved related triples are<e0,r1,e1> The e0 in the question is replaced with the corresponding question words such as "what" and "where" to modify the single-hop question to obtain a two-hop question, and then the triples can be further retrieved.<e0,r1,e1> Another entity e1 in the<e1,r2,e2> The e1 in the sentence is replaced with the corresponding question words such as "what" and "where", and then the two-hop question is modified to obtain a three-hop question. In order to enable the question answering system to handle more diverse questions, the generated questions are processed through synonym conversion, sentence reconstruction, and Chinese-English translation to increase the number of questions and improve the generalization ability of the model. Each data contains the question text and the triple sequence e0,r1,e1,r2,e2,...,e required to answer this question. i ,r i+1 ,e i+1 ,...,e n , divided into training set, validation set and test set in a ratio of 7:1:2;

[0035] S2 represents mineral text as a word vector model and a sentence vector model. The word vector model uses the ComplEx model to embed entities and relations into a d-dimensional complex space. The sentence vector model first uses the Bert model, adds [CLS] at the beginning of the question, and then uses the Bert output corresponding to the question as the initial sentence vector of the question. The initial sentence vector is input into the stacked fully connected layer and projected into the complex space to obtain the final sentence embedding vector W q

[0036] The specific training process of the word vector model is as follows Figure 2 As shown in the figure, the knowledge triple (h, r, t) obtained from the knowledge graph during training is called a positive triple, where h is the head entity, r is the relationship, and t is the tail entity; then a random entity t* is obtained from the knowledge triple, and it is combined with h and r to form a negative triple (h, r, t*). Here, t* is required to be unequal to t. Then h, r, t, t* are converted into corresponding embedding vectors e through the entity embedding vector matrix E and the relationship embedding vector matrix R. h , W r , e t , e t* , use the ComplEx scoring formula shown in formula (1) to calculate the scores of positive and negative triples and use the loss function shown in formula (2) for training, that is, maximize the score of positive triples as much as possible and minimize the score of negative triples, so that the model can better distinguish between positive and negative triples. In formula (1), Re represents the real part of the complex number. Represents e t The complex conjugate operation of ;

[0037]

[0038] like Figure 3 As shown, the sentence vector model is based on the inference path e0,r1,e1,r2,e2,...,e in the data set. i ,r i+1 ,e i+1 ,...,e n The e0 extracted from it is used as e h , e1,...,e n As positive examples e t , and then randomly obtain entities from the knowledge graph as negative examples e t* and e t* Not equal to e t , define the constraint functions shown in formulas (3) and (4) and the loss function shown in formula (5) for training, where ζ represents the set of all entities on the reasoning path belonging to the question, y is the target value, the target value of the positive example is set to 1, and the target value of the negative example is set to 0, σ is the sigmoid function, Re represents the real part of the complex number, Represents e t The complex conjugate operation of ;

[0039]

[0040] S3 performs BIO annotation on mineral questions, where B indicates the beginning of an entity, I indicates the middle of an entity, and O indicates an irrelevant word. Each natural language question is converted from a sentence vector model into a sentence vector and mapped into a label sequence consisting of three letters B, I, and O with the same length as the original question. The Bert-BiLSTM-CRF model is trained using a negative log-likelihood loss function to obtain a central word recognition model that can recognize the central word consisting of B and I in the question.

[0041] like Figure 4 As shown;

[0042] S4 uses two methods to generate candidate entities based on edit distance and word segmentation for the identified question center word, and jointly screens out candidate entities that may be most relevant to the center word in the knowledge graph entity set.

[0043] The candidate entity vectors are concatenated with the question vectors and input into the fully connected layer. The Softmax activation function is used to output the probability that each candidate entity belongs to the entity set that is truly related to the central word. The entity with the highest probability is the disambiguated entity and will serve as the starting point for relational reasoning. For example, Figure 5 As shown, the cross entropy loss function is used in the training process;

[0044] S5 records the entity e that is most relevant to the central word as the initial entity e0, and then obtains all the first-degree outgoing relationship sets of e0 from the knowledge graph, finds the relationship r1 that best reflects the relevance between the question and entity e0 from the relationship set, and then obtains the entity e0 from r1 based on the entity e0 and the relationship r1. i Entity e1 is located in the knowledge graph, and then the above process is repeated starting from entity e1 until r n = <stop>Stop reasoning

[0045] During training, for the training problem q in the data set and its reasoning path e0,r1,e1,r2,e2,...,e i ,r i+1 ,e i+1 ,...,e n , the mineral problem q, r i The pre-inference relation r1-r i-1 With entity e i-1 Use word embedding model to convert to vector representation and <stop>The identifier is added to all one-hop outbound relations of the current entity as candidate relations. The question vector Q is concatenated with the relation vector r1 and then input into the fully connected layer to output a new vector Q'. Q' is concatenated with the relation vector r2 and then input into the fully connected layer to output a new vector Q", and so on. The question vector Q is concatenated with all relations up to r i-1 The preceding reasoning relationship is concatenated one by one and processed by the fully connected layer before being combined with the entity e i-1 and relationship r i After splicing, input the next fully connected layer to get r i The semantic matching probability that can be correctly inferred is that the relationship with the highest matching probability is the direction of the next hop. If <stop>If the identifier has the highest matching probability, it means that the final answer has been inferred, and further reasoning is stopped. The cross entropy loss function is used during training;

[0046] S6 uses the trained central word recognition model and entity disambiguation model to perform central word recognition and knowledge graph entity positioning on the user's input question, obtains the reasoning starting entity, and uses the reasoning model and multi-hop reasoning to gradually obtain the reasoning answer from the mineral knowledge graph and return it to the user.

[0047] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features thereof can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.< / stop> < / stop> < / stop> < / stop> < / stop> < / stop> < / stop>

Claims

1. A complex question-answering method and system applied to a mineral knowledge graph, comprising: S1: Construct a complex mineral question-answering dataset based on the existing simple mineral question-answering dataset; S2: Represent the mineral text as a word vector model and a sentence vector model. The word vector model uses the ComplEx model to embed entities and relations into a d-dimensional complex space. The sentence vector model first uses the Bert model, adds [CLS] at the beginning of the question, and then uses the Bert output corresponding to the question as the initial sentence vector of the question. The initial sentence vector is input into the stacked fully connected layer and projected into the complex space to obtain the final sentence embedding vector W q ; S3: Perform BIO annotation on the mineral question-answering dataset. Each natural language question is converted into a sentence vector and mapped into a label sequence consisting of the three letters B, I, and O with the same length as the sentence vector. The Bert-BiLSTM-CRF model is trained using the negative log-likelihood loss function to obtain a model that can identify the central word consisting of B and I in the question. S4: For the identified central word of the question, two methods, based on edit distance algorithm and word segmentation, are used to jointly select the candidate entities most relevant to the central word in the knowledge graph entity set; S5: Record the candidate entity most relevant to the central word as the initial entity e0, and obtain all the first-degree outgoing relationship sets of e0 from the knowledge graph. Find the relationship r1 that best reflects the relevance between the question and entity e0 from the relationship set, and then locate entity e1 from the knowledge graph based on entity e0 and relationship r1. Then repeat the above process starting from entity e1 until r1 is found. n = <stop> , stop reasoning.< / stop> 2. The complex question-answering method applied to mineral knowledge graph according to claim 1, characterized in that: Step 1, "Constructing a complex mineral question-answering dataset based on an existing simple mineral question-answering dataset," specifically includes: Based on the existing simple single-hop mineral question answering dataset, a complex multi-hop question dataset is constructed by expanding the single-hop question content. The details are as follows: first, the entity e0 contained in the single-hop question is retrieved in the knowledge graph, and then the retrieved related triples are<e0,r1,e1> The e0 in the question is replaced with the corresponding question words such as "what" and "where" to modify the single-hop question to obtain a two-hop question, and then the triples can be further retrieved.<e0,r1,e1> Another entity e1 in the<e1,r2,e2> The e1 in the sentence is replaced with the corresponding question words such as "what" and "where", and then the two-hop question is modified to obtain a three-hop question. In order to enable the question answering system to handle more diverse questions, the generated questions are processed through synonym conversion, sentence reconstruction, and Chinese-English translation to increase the number of questions and improve the generalization ability of the model. Each data contains the question text and the triple sequence e0,r1,e1,r2,e2,...,e required to answer this question. i ,r i+1 ,e i+1 ,...,e n ; 3. The complex question-answering method applied to mineral knowledge graph according to claim 1, characterized in that: Step 2: Represent the mineral text as a word vector model and a sentence vector model. The word vector model uses the ComplEx model to embed entities and relations into a d-dimensional complex space. The sentence vector model first uses the Bert model, adding [CLS] at the beginning of the question. The Bert output corresponding to the question is then used as the initial sentence vector of the question. The initial sentence vector is input into the stacked fully connected layer and projected into the complex space to obtain the final sentence embedding vector W. q Specifically include: Entity embedding is a d-dimensional vector, and relation embedding is a d-dimensional matrix. During training, the knowledge triple (h, r, t) obtained from the knowledge graph is called a positive triple, where h is the head entity, r is the relation, and t is the tail entity. Then, a random entity t* is obtained from the knowledge triple, and it is combined with h and r to form a negative triple (h, r, t*), where t* is required to be not equal to t, and the embedding vectors corresponding to h, r, t, and t* are e respectively. h , W r , e t , e t* , use the ComplEx scoring formula shown in formula (1) to calculate the scores of positive and negative triples and use the loss function shown in formula (2) for training, that is, maximize the score of positive triples as much as possible and minimize the score of negative triples, so that the model can better distinguish between positive and negative triples. In formula (1), Re represents the real part of the complex number. Represents e t The complex conjugate operation of ; According to the inference path e0,r1,e1,r2,e2,...,e i ,r i+1 ,e i+1 ,...,e n The e0 extracted from it is used as e h , e1,...,e n As positive examples e t , and then randomly obtain entities from the knowledge graph as negative examples e t* and e t* Not equal to e t , define the constraint functions shown in formulas (3) and (4) and the loss function shown in formula (5) for training, where ζ represents the set of all entities on the reasoning path belonging to the question, y is the target value, the target value of the positive example is set to 1, and the target value of the negative example is set to 0, σ is the sigmoid function, Re represents the real part of the complex number, Represents e t The complex conjugate operation of ; 4. The complex question-answering method applied to mineral knowledge graph according to claim 1, characterized in that: Step 4, "Using two methods, namely, the edit distance algorithm and the word segmentation algorithm, to jointly select candidate entities related to the central word in the knowledge graph entity set for the identified central word of the question," specifically includes: The candidate entity vectors are concatenated with the question vector and then fed into a fully connected layer. The Softmax activation function is used to output the probability that each candidate entity belongs to the set of entities truly related to the central word. The entity with the highest probability is the disambiguated entity and serves as the starting point for relational reasoning. The cross-entropy loss function is used during training.

5. The complex question-answering method applied to mineral knowledge graph according to claim 1, characterized in that: Step 5: Record the candidate entity most relevant to the central word as the initial entity e0, and obtain all the first-degree outgoing relationship sets of e0 from the knowledge graph. Find the relationship r1 that best reflects the relevance between the question and entity e0 from the relationship set. Then, locate entity e1 from the knowledge graph based on entity e0 and relationship r1. Then repeat the above process starting from entity e1 until r1 is found. n = <stop> , stop reasoning "specifically includes:< / stop> During training, for the training problem q in the data set and its reasoning path e0,r1,e1,r2,e2,...,e i ,r i+1 ,e i+1 ,...,e n , question q, r i The pre-inference relation r1-r i-1 With entity e i-1 Use word embedding model to convert to vector representation, and then <stop>The identifier is added to all one-hop outbound relations of the current entity as candidate relations. The question vector Q is concatenated with the relation vector r1 and then input into the fully connected layer to output a new vector Q'. Q' is concatenated with the relation vector r2 and then input into the fully connected layer to output a new vector Q", and so on. The question vector Q is concatenated with all relations up to r i-1 The preceding reasoning relationship is concatenated one by one and processed by the fully connected layer before being combined with the entity e i-1 and relationship r i After splicing, input the next fully connected layer to get r i The semantic matching probability that can be correctly inferred is that the relationship with the highest matching probability is the direction of the next hop. If <stop> If the identifier has the highest matching probability, it means that the final answer has been inferred, and further reasoning is stopped. The cross entropy loss function is used during training;< / stop> < / stop> 6. The complex question-answering method applied to mineral knowledge graph according to claim 1, characterized in that: First, the user's input question is converted into a sentence vector using the Mineral Text Representation Model. Then, the Mineral Question Central Word Recognition Model is used to parse the central word, and the entity disambiguation model is used to obtain the initial entity for reasoning. The multi-hop reasoning of the reasoning model is used to gradually obtain the reasoning answer from the Mineral Knowledge Graph and return it to the user.

7. The complex question-answering method applied to a mineral knowledge graph according to claim 1, characterized in that: The complex question-answering method applied to the mineral knowledge graph is used to answer complex mineral questions involving multi-hop relationships.

Images