Domain knowledge question and answer interaction method of expert knowledge base

Abstract

The invention relates to the technical field of knowledge questions and answers, in particular to a domain knowledge question and answer interaction method of an expert knowledge base, and the method comprises the steps: collecting knowledge data of a set domain, and carrying out the triple extraction to obtain a triple data set; carrying out statistics on the distribution of the relational words in each triple in the triple data set, and determining a first frequency characteristic value of each triple; classifying all relational words in the triple data set, counting the relational words in each triple and the distribution of each relational word in the category to which the relational words belong in the triple data set, and determining a second frequency characteristic value of each triple; and combining the first frequency characteristic value and the second frequency characteristic value to obtain the entity relationship anomaly degree of each triad, screening and eliminating abnormal triads in the triad data set, and constructing a knowledge graph by using the remaining triads in the triad data set to perform knowledge question and answer interaction. The accuracy of domain knowledge question-answer interaction is improved.

Description

Technical Field

[0001] This application relates to the field of knowledge question answering technology, specifically to a domain knowledge question answering interaction method for an expert knowledge base. Background Technology

[0002] Knowledge base question answering (KBQA) is a system that uses natural language processing technology to extract information from a knowledge base to answer user questions. Knowledge graphs are a common and effective way to build knowledge bases.

[0003] When using knowledge graph technology to build an expert knowledge base for a specific domain, existing techniques collect multi-source data within that domain and use knowledge extraction tools to extract knowledge from the acquired multi-source data. Based on the extracted triples, a knowledge graph for that domain is then built, resulting in the expert knowledge base. However, the collected multi-source data may contain noisy data due to poor data source quality, inconsistent data formats, and garbled data. This noise can cause abnormal triples in the extracted triples, making the built knowledge graph unable to accurately reflect the true knowledge structure within the domain. Consequently, the reliability of the expert knowledge base is reduced, leading to frequent low-quality answers during the domain-specific question-answering process. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides a domain knowledge question-answering interaction method for an expert knowledge base, thereby resolving the existing issues.

[0005] The domain knowledge question-answering interaction method for an expert knowledge base proposed in this application adopts the following technical solution: One embodiment of this application provides a domain knowledge question-answering interaction method for an expert knowledge base, the method comprising the following steps: Collect knowledge data in a defined domain, extract triples, and obtain a triple dataset. Statistically analyze the distribution of relation words in each triplet in the triplet dataset, and determine the first frequency feature value of each triplet; Obtain the set of synonyms for each relation word in the triple dataset; analyze the semantic similarity between the synonym sets of any two relation words in the triple dataset, and determine the entity relation similarity between the two relation words. Based on the entity relation similarity, all relation words in the triplet dataset are classified, and the distribution of relation words in each triplet and each relation word in its category in the triplet dataset is statistically analyzed to determine the second frequency feature value of each triplet. By combining the first frequency feature value and the second frequency feature value, the entity relationship anomaly degree of each triple is obtained. Abnormal triples are filtered and removed from the triple dataset. The remaining triples in the triple dataset are used to construct a knowledge graph for knowledge question answering interaction.

[0006] In one embodiment, the first frequency feature value is the frequency of occurrence of the relation word in each triple in the triple dataset.

[0007] In one embodiment, the process of determining the entity relationship similarity is as follows: Calculate the similarity between the word vector of each synonym in the synonym set of relation word b1 and the word vector of each synonym in the synonym set of relation word b2. Then, perform a positive fusion of all similarities between relation word b1 and relation word b2 to obtain the entity relationship similarity between relation word b1 and relation word b2.

[0008] In one embodiment, the forward fusion is calculated by averaging.

[0009] In one embodiment, when classifying all relation words in a triplet dataset, the entity relation similarity is used to determine the metric distance between relation words.

[0010] In one embodiment, a negative correlation mapping is performed on the entity relationship similarity to obtain the metric distance between relation words.

[0011] In one embodiment, determining the second frequency characteristic value includes: The positive fusion result of the frequency of occurrence of the relation words in each triple and all relation words in their respective categories in the triple dataset is used as the second frequency feature value of each triple.

[0012] In one embodiment, the entity relationship anomaly degree is negatively correlated with both the first frequency feature value and the second frequency feature value.

[0013] In one embodiment, filtering and removing outlier triples from the triple dataset includes: Obtain triples with and without entity relationship anomalies, calculate the entity relationship anomaly degree of each triple, use it as training data to train a classification model, use the trained classification model to identify each abnormal triple in the triple dataset, and remove them.

[0014] In one embodiment, the input to the trained classification model is the word vectors of each triple in the triple dataset and their entity relation anomalies.

[0015] This application has at least the following beneficial effects: This application, by statistically analyzing the distribution of relational terms in a triplet dataset and calculating the first frequency feature value, can effectively identify which relational terms have higher importance in the knowledge base. This helps in the subsequent identification of triplets with abnormal relations in the triplet dataset. By obtaining a synonym set, it helps to identify and enrich the relational vocabulary used in the domain, avoiding the misidentification of abnormal relational triplets caused by the low frequency of some relational terms due to multiple expressions of relational terms. Furthermore, by combining the first frequency feature value and the second frequency feature value to calculate the entity relation anomaly degree, it can effectively identify and remove abnormal triplets in the knowledge base, enhance the ability of data quality management, avoid erroneous reasoning or inaccurate answers caused by abnormal data, ensure the integrity and consistency of the knowledge graph, and enable users to obtain more accurate and relevant information when interacting with the knowledge base. This high efficiency and targeted nature of the interaction greatly improves the user experience, enhances the practicality of the knowledge base, and achieves high-quality question-and-answer interaction. Attached Figure Description

[0016] To more clearly illustrate the technical solutions and advantages in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart illustrating the steps of a domain knowledge question-answering interaction method for an expert knowledge base provided in this application; Figure 2 Flowchart for determining the anomaly degree of entity relationships in triplet data. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by this application to achieve the intended purpose of the invention, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a domain knowledge question-and-answer interaction method for an expert knowledge base proposed in this application. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0020] The following description, in conjunction with the accompanying drawings, details a specific scheme for a domain knowledge question-answering interaction method for an expert knowledge base provided in this application.

[0021] This application provides an embodiment of a domain knowledge question-answering interaction method for an expert knowledge base. Specifically, it provides the following domain knowledge question-answering interaction method for an expert knowledge base. Please refer to [link / reference]. Figure 1 The method includes the following steps: Step S001: Collect knowledge data in the defined domain, extract triples, and obtain a triple dataset.

[0022] This embodiment collects raw text data from multiple data sources for a specific domain and preprocesses the acquired raw text data to obtain a domain knowledge database formed by the preprocessed text data. In this embodiment, data preprocessing includes data cleaning, format normalization, data fusion, and missing value handling. The specific domain can be the medical field, military field, or public safety field, etc. This embodiment takes the military field as an example. The data sources for obtaining the raw text data are professional data sources such as military equipment magazines, military equipment journals, and academic papers. This is to avoid high-frequency abnormal triplet data with semantic errors (such as common-sense mismatches) when extracting triplet data from entity information and relationships between entities in the subsequent triplet extraction process. The raw text data can be obtained using web crawling technology. Both web crawling technology and text data preprocessing are existing and well-known technologies, and the specific process will not be elaborated further.

[0023] During the data preprocessing of the original text data, all acquired text data undergoes sequential data format unification and cleaning, as well as noise filtering. Data format unification and cleaning includes removing redundant symbols such as special punctuation, garbled characters, and repeated spaces; unifying full-width / half-width characters; and standardizing text case. Noise filtering removes duplicate text, garbled characters, and non-natural language text—invalid content—to prevent high-frequency abnormal triples from being incorrectly identified as entities or relations due to noise introduced by inconsistent data formats and garbled characters from the collected multi-source data when extracting triples from entity information and relationships in the original text data. The data format unification and cleaning, as well as the noise filtering, are well-known techniques, and their specific processes will not be elaborated upon.

[0024] The open-source knowledge extraction tool DeepDive is used to extract entities, relations, and attributes from all text data in the acquired domain knowledge database. This yields information on all entities, relationships between entities, and attribute information within the text data. DeepDive is then used to extract triples from the extracted entity information and relationships, resulting in a triple dataset A, which represents the specific relationships between entities within the text data of the domain knowledge database. Entity extraction, relation extraction, attribute extraction, and triple extraction are all well-known techniques, and their specific processes will not be elaborated upon here.

[0025] Step S002: Statistically analyze the distribution of relation words in each triplet in the triplet dataset, and determine the first frequency feature value of each triplet.

[0026] Because noise filtering of text data cannot completely remove all noise, the acquired domain knowledge database triple dataset will still contain some low-frequency anomalous triples due to residual noise in the text data. Furthermore, the entity relation extraction method used in knowledge graph construction is based on extensive data training and existing knowledge systems, resulting in a generally uniform distribution of normal entity relations in the acquired triple dataset. Therefore, the anomalous entity relations formed by the anomalous triples analyzed in this embodiment typically have a lower frequency of occurrence than normal entity relations in the acquired triple dataset.

[0027] Based on the above analysis, taking any triplet data 'a' in triplet dataset A as an example, we count the frequency of the relation term in triplet data 'a' in all triplet data in triplet dataset A. The statistical result is used as the first frequency feature value of triplet data 'a' to evaluate the occurrence of the entity relation corresponding to triplet data 'a' in triplet dataset A. The smaller the first frequency feature value, the less frequently the entity relation corresponding to triplet data 'a' appears in triplet dataset A. Therefore, triplet data 'a' is more likely to be an abnormal triplet data in triplet dataset A due to noise.

[0028] Step S003: Obtain the set of synonyms for each relation word in the triplet dataset; analyze the semantic similarity between the synonym sets of any two relation words in the triplet dataset, and determine the entity relation similarity between the two relation words.

[0029] Because entity relations with the same meaning may have different verbal expressions in different contexts or data sources—for example, different verbal expressions of instance relations such as "belongs to," "is," and "classifies as" (e.g., aircraft X, belongs to, carrier-based fighter jet), and different verbal expressions of part-whole relations such as "composes," "includes," and "constituts" (e.g., reconnaissance identification system, includes, reconnaissance aircraft)—these different verbal expressions can lead to a lower frequency of occurrence of entity relations corresponding to individual triplet data in the acquired domain knowledge database due to the use of special relation words. Therefore, to reduce the impact of different relation words formed by multiple different verbal expressions of the same entity relation in different contexts or data sources on the evaluation of low-frequency abnormal triplet data, the following processing is performed: Extract all distinct relation words from all triplet data in triplet dataset A to form relation word set B, which contains unique relation words. Obtain the synonym set (including the relation word itself) for each relation word in relation word set B from WordNet. This is used to subsequently evaluate whether the relation words in relation word set B are different verbal expressions of the same entity relation in triplet dataset A. WordNet can obtain the synonym set for Chinese words by installing the OMW component; WordNet is a well-known technology, and the specific process will not be elaborated further. It should be noted that when a relation word in the relation word set does not have a synonym set in WordNet, calculate the cosine similarity between the word vector of this relation word and the word vectors of the remaining relation words in the relation word set. Sort the cosine similarity in descending order, and select the top 20% of the sorted relation words to form the synonym set for that relation word. The word vectors are obtained using the Word2Vec model.

[0030] Taking any two relation words b1 and b2 in relation word set B as an example, each synonym in the synonym set of relation words b1 and b2 is vectorized using a word vector generation model to obtain the word vector of each synonym. Further, the similarity between the word vectors of each synonym in the synonym set of relation word b1 and each synonym in the synonym set of relation word b2 is calculated. All the obtained similarities are positively fused to obtain the entity relation similarity between relation words b1 and b2. This similarity is used to evaluate whether relation words b1 and b2 are different linguistic expressions of the same entity relation in triplet dataset A. The greater the entity relation similarity, the more likely relation words b1 and b2 are different linguistic expressions of the same entity relation.

[0031] It should be noted that positive fusion means combining multiple variables, which can be calculated using methods such as addition, multiplication, or averaging. In this embodiment, the average of all similarities obtained between relation word b1 and relation word b2 is taken as the entity relationship similarity between relation word b1 and relation word b2.

[0032] In this embodiment, the similarity between word vectors is calculated using cosine similarity; the word vector generation model is the Word2Vec model; in other embodiments, the word vector generation model can be the BERT model or the GloVe model, etc. The Word2Vec word vector generation model and the calculation of cosine similarity are well-known techniques, and their specific processes will not be elaborated further.

[0033] Step S004: Classify all relation words in the triplet dataset based on the entity relation similarity, count the distribution of relation words in each triplet and each relation word in its category in the triplet dataset, and determine the second frequency feature value of each triplet.

[0034] Clustering algorithms are used to cluster all relation words in relation word set B, resulting in multiple clusters for relation word set B. Each cluster represents an entity relation pattern in triplet dataset A. Clustering algorithms such as K-Means, DBSCAN, or hierarchical clustering can be used; this embodiment selects K-Means clustering. During clustering, a negative correlation mapping is applied to the entity relation similarity between relation words to obtain a metric distance between them. This embodiment uses the silhouette coefficient method to determine the number of clusters in the K-Means clustering algorithm. K-Means clustering is a well-known technique, and its specific process will not be elaborated further.

[0035] It should be noted that negative correlation mapping represents a mathematical relationship, where an increase in one variable leads to a decrease in another, indicating a negative correlation between the distance measurement and entity relationship similarity. In this embodiment, the negative correlation mapping is performed as follows: the negative of the entity relationship similarity is used as the exponent of an exponential function with the natural constant as the base, and the result of this exponential function is used as the distance measurement between corresponding relational terms. In another embodiment, the negative correlation mapping is performed as follows: the entity relationship similarity is normalized using the Sigmoid function, and the reciprocal of the normalization result is used as the distance measurement between corresponding relational terms.

[0036] Taking triple data a as an example, we select cluster C containing the relation words in triple data a from all clusters in relation word set B. We then count the frequency of each relation word in cluster C in triple dataset A and use the positive fusion result of all the counted frequencies as the second frequency feature value of triple data a.

[0037] In this embodiment, the sum of all occurrence frequencies obtained from the cluster C corresponding to triplet data a is used as the second frequency feature value of triplet data a. This value is used to evaluate the occurrence of all entity relations in triplet dataset A that belong to the same type of entity relation as the entity relation corresponding to triplet data a. The smaller the second frequency feature value, the lower the frequency of occurrence, and the more likely triplet data a is to be a low-frequency abnormal triplet data in triplet dataset A due to noise interference.

[0038] Step S005: Combine the first frequency feature value and the second frequency feature value to obtain the entity relationship anomaly degree of each triple. Filter and remove abnormal triples in the triple dataset. Use the remaining triples in the triple dataset to construct a knowledge graph for knowledge question answering interaction.

[0039] The first and second frequency eigenvalues ​​of all triplet data in triplet dataset A are normalized using the Min-Max normalization method, resulting in the normalized results of the first and second frequency eigenvalues ​​of each triplet data in triplet dataset A. The Min-Max normalization method is a well-known technique, and the specific process will not be described in detail.

[0040] Taking triplet data 'a' as an example, based on the normalized results of the first and second frequency feature values ​​of triplet data 'a', the entity relationship anomaly degree of triplet data 'a' is calculated. This is used to assess whether triplet data 'a' is a low-frequency anomalous triplet data in triplet dataset A caused by data quality issues such as incomplete data and data noise in the collected multi-source data. The smaller the normalized results of the first and second frequency feature values, the greater the entity relationship anomaly degree, and the more likely triplet data 'a' is to be an anomalous triplet data. In this embodiment, the calculation method for the entity relationship anomaly degree of triplet data 'a' is as follows: In the formula, W represents the entity relation anomaly degree of triple data a, and S represents the mean of the normalized results of the first frequency feature value and the normalized results of the second frequency feature value of triple data a. This embodiment sets a preset minimum positive number greater than 0 to prevent the denominator from being 0. The implementer can set the parameters according to the actual situation; this embodiment does not impose any restrictions. The flowchart for determining the anomaly degree of entity relationships in triplet data is as follows: Figure 2 As shown.

[0041] Furthermore, all triplet data from multiple different knowledge databases within the same domain are obtained through manual sampling and annotation. This includes low-frequency anomalous triplet data and correct triplet data. The entity relation anomaly degree of each obtained triplet data is calculated using the same calculation method as for triplet data 'a'. In this embodiment, the number of low-frequency anomalous triplet data and correct triplet data in the obtained triplet data is 200 each; the implementer can set this number according to actual conditions. The obtained low-frequency anomalous triplet data, correct triplet data, and the corresponding entity relation anomaly degrees are all used as the training dataset.

[0042] A binary classification model is trained based on the training dataset. This trained model is then used as the entity relationship anomaly detection model for triplet data in triplet dataset A, determining whether triplet data in dataset A are low-frequency anomalous triplet data. In the training dataset, the label for the entity relationship anomaly degree of low-frequency anomalous triplet data is set to 1, while the label for the entity relationship anomaly degree of correct triplet data is set to 0. The binary classification model is trained using a Support Vector Machine (SVM) model in this embodiment. Implementers can choose other feasible binary classification models, such as decision tree models or logistic regression models. The training of the SVM model is a well-known technique, and the specific process will not be elaborated further.

[0043] The word vectors and entity relation anomaly scores of each triplet in triplet dataset A are used as inputs to the trained entity relation anomaly detection model. The output is a label for the entity relation anomaly score of each triplet in triplet dataset A. All triplet datasets in triplet dataset A with a label of 1 for entity relation anomaly score are removed, resulting in the removed triplet dataset A1. This dataset represents all triplet datasets in triplet dataset A that contain normal entity relation data. The word vectors of the triplets are obtained using the Word2Vec model.

[0044] The protege open-source software was used to sequentially perform knowledge fusion and knowledge graph construction on all triples in the triple dataset A1, resulting in the domain knowledge graph corresponding to the domain knowledge database. The knowledge fusion and knowledge graph construction using protege are well-known techniques, and the specific process will not be elaborated further. The Neo4j graph database was then used to store the constructed domain knowledge graph as a knowledge base, resulting in the expert knowledge base.

[0045] In the process of domain knowledge question answering, the user's input text is acquired, and text relationship classification and named entity recognition are performed on the input text to achieve text intent parsing. The text is extracted into triples, and a fuzzy matching algorithm is used to match the entities in the triples, avoiding the uncertainty caused by the user's inconsistent names. Through query strategies and rule-based methods, the triples processed by the fuzzy matching algorithm are transformed into a database query language that the graph database used by the expert knowledge base can understand. The transformed database query language is then used to query the expert knowledge base, and the query results are fed back to the user, completing the knowledge question answering interaction. The text relationship classification, named entity recognition, and fuzzy matching algorithms are all existing well-known technologies, and their specific processes are not elaborated upon.

[0046] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, specific embodiments of this specification have been described above. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0047] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0048] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them; modifications to the technical solutions described in the foregoing embodiments, or equivalent substitutions of some of the technical features, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A domain knowledge question-answering interaction method for an expert knowledge base, characterized in that, The method includes the following steps: Collect knowledge data in a defined domain, extract triples, and obtain a triple dataset. Statistically analyze the distribution of relation words in each triplet in the triplet dataset, and determine the first frequency feature value of each triplet; Obtain the set of synonyms for each relation word in the triple dataset; analyze the semantic similarity between the synonym sets of any two relation words in the triple dataset, and determine the entity relation similarity between the two relation words. Based on the entity relationship similarity, all relation words in the triplet dataset are classified, and the distribution of relation words in each triplet and each relation word in its category in the triplet dataset is statistically analyzed to determine the second frequency feature value of each triplet. By combining the first frequency feature value and the second frequency feature value, the entity relationship anomaly degree of each triple is obtained. Abnormal triples are filtered and removed from the triple dataset. The remaining triples in the triple dataset are used to construct a knowledge graph for knowledge question answering interaction.

2. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, The first frequency feature value is the frequency of occurrence of the relation word in each triplet in the triplet dataset.

3. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, The process for determining the similarity of the entity relationships is as follows: Calculate the similarity between the word vector of each synonym in the synonym set of relation word b1 and the word vector of each synonym in the synonym set of relation word b2. Then, perform a positive fusion of all similarities between relation word b1 and relation word b2 to obtain the entity relationship similarity between relation word b1 and relation word b2.

4. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 3, characterized in that, The forward fusion is calculated by taking the average value.

5. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, When classifying all relation words in the triplet dataset, the entity relation similarity is used to determine the metric distance between relation words.

6. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 5, characterized in that, By performing a negative correlation mapping on the entity relationship similarity, the metric distance between relational terms is obtained.

7. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, The determination of the second frequency characteristic value includes: The positive fusion result of the frequency of occurrence of the relation words in each triple and all relation words in their respective categories in the triple dataset is used as the second frequency feature value of each triple.

8. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, The anomaly degree of the entity relationship is negatively correlated with both the first frequency feature value and the second frequency feature value.

9. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 1, characterized in that, The process of filtering and removing outlier triples from the triple dataset includes: Obtain triples with and without entity relationship anomalies, calculate the entity relationship anomaly degree of each triple, use it as training data to train a classification model, use the trained classification model to identify each abnormal triple in the triple dataset, and remove them.

10. The domain knowledge question-answering interaction method for an expert knowledge base as described in claim 9, characterized in that, The input to the trained classification model is the word vectors of each triple in the triple dataset and its entity relation anomaly score.

Images