CN117251685B - Knowledge graph-based standardized government affair data construction method and device - Google Patents

Abstract

The present invention provides a method and device for constructing standardized government data based on a knowledge graph, which belongs to the field of data processing technology. The method of constructing standardized government data based on a knowledge graph utilizes a feature extraction model through seed words based on government scenarios. , identify multiple initial entities in the government affairs scenario, then use the mutual information value between adjacent initial entities to obtain the first phrase entity, and obtain the second phrase entity by calculating the left and right entropy based on the mutual information value, The scope of phrase entities is further expanded, and the target entity is finally obtained, which realizes the extraction of phrase entities that are nested combinations of multiple words, and then obtains richer and more accurate entities to build the knowledge graph in government affairs scenarios.

Description

A standardized government data construction method and device based on knowledge graph

Technical field

The present invention relates to the field of data processing technology, and in particular to a method and device for constructing standardized government data based on knowledge graphs.

Background technique

Knowledge graphs can formally describe things in the real world and their relationships. The knowledge graph uses entities, relationships, and entity triples to store knowledge, and uses entities as nodes and relationships as edges to build a knowledge network. At present, many well-known knowledge graph projects organize a large amount of data, from which knowledge can be extracted for organization and management, and provide users with high-quality intelligent services, such as understanding the semantics of search and providing more accurate search answers.

In government affairs scenarios, knowledge graphs can play a very important role. The field of government affairs involves a large number of entities and relationships, such as government agencies, departments, officials, regulations, policies, etc. There are complex connections and dependencies between these entities and relationships. Knowledge graphs can model these entities and relationships in the form of graphs to achieve effective management and application of government information. When building a knowledge graph for government affairs scenarios, it is necessary to focus on the standardized representation and processing of data (knowledge) in the process of building a government service knowledge base. The construction process of the knowledge graph includes the acquisition, cleaning and extraction of knowledge (entities, relationships, attributes) of data in government affairs scenarios, thereby achieving knowledge representation, ontology construction, knowledge storage and knowledge fusion.

However, the text data of government affairs scenarios are usually complex and contain a large amount of professional terms and context-related information, and the extraction of entities and relationships may face challenges. Existing entity extraction methods need to segment words first and then obtain entities. Most word segmentation methods use dictionary-based word segmentation methods. The dictionaries used are cross-domain general dictionaries, mostly commonly used vocabulary, and lack of proprietary vocabulary in government affairs scenarios. Moreover, many text entities in government affairs scenarios are nested combinations of multiple words. For nested combinations of multiple words, for example, for the word "house rental party", the entities extracted according to the conventional word segmentation method will be divided into "House rental" and "parties", the effect of this method is not ideal.

Contents of the invention

The present invention provides a standardized government data construction method and device based on knowledge graphs to solve the shortcomings of the existing technology that the entity extraction effect is not ideal, and achieve the effect of accurately extracting entities in government affairs scenarios.

The present invention provides a standardized government data construction method based on knowledge graph, including:

Perform data preprocessing on the corpus of government affairs scenarios;

Based on the seed words of the government affairs scenario, the feature extraction model is used to identify the initial entities in the preprocessed corpus;

Determine mutual information values between adjacent initial entities, and combine adjacent initial entities with mutual information values greater than a first threshold into a first phrase entity;

Determine the left and right entropies of the initial entities whose mutual information values of adjacent initial entities are greater than the second threshold in the preprocessed corpus, and determine the second phrase entity based on the initial entities whose left and right entropies are less than the third threshold; the third The second threshold is smaller than the first threshold;

Determine the initial entity, the first phrase entity and the second phrase entity whose mutual information value with the adjacent initial entity is less than or equal to the first threshold as the target entity;

Based on the corpus of government affairs scenarios and the target entities, extract the relationship between the target entities;

Store the determined relationships between target entities in a relational database to build a government knowledge graph.

According to a knowledge graph-based standardized government data construction method provided by the present invention, determining the mutual information value between adjacent initial entities includes:

Traverse the corpus of government affairs scenes and determine the number of occurrences of two adjacent initial entities and phrases composed of two adjacent initial entities in the corpus of government affairs scenes;

Divide the number of occurrences of two adjacent initial entities by the total number of all words in the corpus to obtain the occurrence probabilities of the two adjacent initial entity words, and use the phrases composed of two adjacent initial entities in the government affairs scene The number of occurrences in the corpus is divided by the total number of all words in the corpus to obtain the occurrence probability of a phrase composed of two adjacent initial entities;

Based on the occurrence probability of two adjacent initial entity words and the occurrence probability of phrases composed of two adjacent initial entities, the mutual information value between adjacent initial entities is determined.

According to a knowledge graph-based standardized government data construction method provided by the present invention, determining the left and right entropy of an initial entity whose mutual information value of adjacent initial entities is greater than the second threshold in the preprocessed corpus includes:

Determine the size of the word selection window, which is used to limit the range of left and right adjacent words of the initial entity in the preprocessed corpus;

Based on the word selection window size, obtain adjacent words in the left and right range of the initial entity whose mutual information value of adjacent initial entities is greater than the second threshold in the corpus of the government affairs scenario;

Determine the number of times each adjacent word appears in all word selection windows, and obtain the frequency of each adjacent word appearing in all left word selection windows and all right word selection windows;

Based on the frequency of occurrence of each adjacent word in each word selection window, the left and right entropy of the initial entity corresponding to each adjacent word in the preprocessed corpus is determined.

According to a knowledge graph-based standardized government data construction method provided by the present invention, the second phrase entity is determined based on the initial entity whose left and right entropy is less than the third threshold, including:

Determine the left entropy value and right entropy value in the left and right entropy of the initial entity whose left and right entropy is less than the third threshold in the preprocessed corpus;

The second phrase entity is composed of the adjacent words corresponding to the smaller one of the left entropy value and the right entropy value and the initial entity.

According to a knowledge graph-based standardized government data construction method provided by the present invention, the feature extraction model is used to identify the initial entities in the preprocessed corpus, including:

Based on the seed words of the government affairs scenario, a text feature template is constructed; the feature template includes the word vector representation, part-of-speech tag, seed word distance and contextual information;

Using initial training corpus and the text feature template to train the feature extraction model;

The corpus in the preprocessed corpus is input into the trained feature extraction model, the features corresponding to the initial entity output by the feature extraction model are obtained, and the initial entity is determined.

According to a knowledge graph-based standardized government data construction method provided by the present invention, the feature extraction model is a BiLSTM-CRF model, and the feature extraction model includes an input layer, a coding layer and a decoding layer;

The input layer is used to convert words or characters in the input corpus text into continuous vector representations;

The encoding layer is used to extract contextual features of the vector sequence of the input corpus text, and generate an encoding vector sequence containing semantic information;

The decoding layer is configured to use a conditional random field to annotate the encoding vector sequence output by the encoding layer, predict the label of each annotated position, and generate and output features corresponding to the initial entity.

According to a knowledge graph-based standardized government data construction method provided by the present invention, the seed words of the government scene include words in the government scene-specific dictionary and words manually annotated in the government scene.

The present invention also provides a standardized government data construction device based on knowledge graph, including:

The first processing module is used for data preprocessing of the corpus of government affairs scenarios;

The second processing module is used to use the feature extraction model to identify the initial entities in the preprocessed corpus based on the seed words of the government affairs scenario;

A third processing module, configured to determine mutual information values between adjacent initial entities, and combine adjacent initial entities with mutual information values greater than the first threshold into the first phrase entity;

The fourth processing module is used to determine the left and right entropy of the initial entity in the preprocessed corpus for which the mutual information value of the adjacent initial entities is greater than the second threshold, and determine the first entity based on the initial entity whose left and right entropy is less than the third threshold. Two phrase entities; the second threshold is smaller than the first threshold;

A fifth processing module, configured to determine the initial entity, the first phrase entity and the second phrase entity whose mutual information value with the adjacent initial entity is less than or equal to the first threshold as the target entity;

The sixth processing module is used to extract the relationship between the target entities based on the corpus of the government affairs scenario and the target entities;

The seventh processing module is used to store the determined relationships between target entities in a relational database to build a government knowledge graph.

The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements any one of the above-mentioned methods based on the knowledge graph. Standardized government data construction method.

The present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the standardized government data construction method based on knowledge graphs as described in any of the above is implemented.

The present invention also provides a computer program product, which includes a computer program. When the computer program is executed by a processor, the computer program implements any one of the above knowledge graph-based standardized government data construction methods.

The present invention provides a method and device for constructing standardized government data based on a knowledge graph, using a feature extraction model based on seed words of a government scenario to identify multiple initial entities in a government scenario, and then using the adjacent initial entities to The first phrase entity is obtained from the mutual information value, and the second phrase entity is obtained by calculating the left and right entropy based on the mutual information value, further expanding the scope of the phrase entity, and finally obtaining the target entity, realizing the nesting of multiple words The combined phrase entities are extracted to obtain richer and more accurate entities to construct the knowledge graph in government affairs scenarios.

Description of the drawings

In order to explain the present invention or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are of the present invention. For some embodiments of the invention, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

Figure 1 is a schematic flow chart of the standardized government data construction method based on knowledge graph provided by the present invention;

Figure 2 is a schematic structural diagram of a standardized government data construction device based on knowledge graphs provided by the present invention;

Figure 3 is a schematic structural diagram of the electronic device provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention more clear, the technical solutions in the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention. , not all examples. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without making creative efforts fall within the scope of protection of the present invention.

The following describes a knowledge graph-based standardized government data construction method and device of the present invention in conjunction with Figures 1-3.

As shown in Figure 1, a standardized government data construction method based on a knowledge graph according to the embodiment of the present invention mainly includes step 110, step 120, step 130, step 140, step 150, step 160 and step 170.

Step 110: Perform data preprocessing on the corpus of government affairs scenarios.

The corpus of government affairs scenarios refers to text data sets used in the fields of government, administrative agencies, public services, etc. These data sets usually contain texts related to policies and regulations, public services, and administrative offices.

The corpus of government affairs scenarios usually contains texts of various policies and regulations, such as constitutions, laws, administrative regulations, rules, normative documents, etc. Public services provided by government agencies cover all areas of social life. The corpus of government affairs scenarios usually contains texts related to public services, such as bidding announcements, budget reports, financial statements, personnel files, social security benefits, etc. The corpus of government affairs scenes may also contain various administrative office-related texts, such as official documents, letters, reports, memos, meeting minutes, work plans, etc.

Data preprocessing of the corpus of government affairs scenarios can include cleaning text data. The data sources in government affairs scenarios are usually relatively fixed, and duplicate data may appear. Therefore, when performing data cleaning, it is necessary to identify and remove duplicate data to avoid interference with subsequent analysis. Data sources in government affairs scenarios may also include some noisy data, such as irrelevant information, mislabeling, etc. When performing data cleaning, these noisy data need to be identified and removed to avoid affecting subsequent analysis.

Since the data sources of government affairs scenarios may come from different agencies or departments, their formats and naming methods may be different. After cleaning is completed, the data can be standardized to ensure that the format and naming of the data are unified.

When there is missing data in government affairs scenarios, such as some fields are not filled in or recorded, these missing data can be identified and processed, such as filling in default values, deleting records, etc.

On this basis, corpus text that is convenient for entity extraction can be obtained.

Step 120: Based on the seed words of the government affairs scenario, use the feature extraction model to identify the initial entities in the preprocessed corpus.

In government affairs scenarios, seed words can be words related to government, administrative management, public affairs, etc.

For example, seed words can be words corresponding to policy documents, laws and regulations (such as the Constitution, criminal law), policy measures, etc.

Seed words can also be words corresponding to public services, such as education, medical care, social security, environmental protection, infrastructure construction, etc.

Seed words can also be words corresponding to government information systems, government activities and meetings, etc. There are no restrictions here.

The seed words of government affairs scenarios can be used as the starting point for relevant information in government affairs scenarios, and are used to search, classify, organize, and extract entities in government affairs scenarios. Of course, the specific seed word list needs to be further customized and expanded based on actual application scenarios and needs.

It can be understood that the seed words of the government affairs scene may include words in the government affairs scene-specific dictionary and words manually annotated in the government affairs scene. The government affairs scenario-specific dictionary is a tool used to store specific terms and nouns in the government affairs field. It can help with tasks such as entity recognition and information extraction from government affairs texts. That is, the seed words of government affairs scenes can be obtained by obtaining existing government affairs scene-specific words or manual annotation.

According to the task requirements of extracting entities, you can choose an appropriate feature extraction model, such as a machine learning-based model (such as conditional random field, maximum entropy model) or a deep learning-based model (such as recurrent neural network, Transformer), to extract government affairs scenes entities in the corpus.

When performing feature extraction, appropriate features can be designed based on the government affairs scenario and seed words. On this basis, the feature extraction model is trained by using the labeled sample data. The sample data can be composed of manually annotated entity data and a corpus of government affairs scenarios to ensure coverage of various entity types in government affairs scenarios.

It can be understood that the trained feature extraction model is used to perform entity recognition on the preprocessed corpus. The feature extraction model can classify each word according to the features obtained by feature extraction, determine whether it is an entity, and obtain each initial entity. .

In some embodiments, based on the seed words of the government affairs scenario, a feature extraction model is used to identify the initial entities in the preprocessed corpus, including the following process.

You can first build a text feature template based on the seed words of the government affairs scene. The feature template includes the word vector representation, part-of-speech tag, seed word distance and contextual information.

On this basis, the initial training corpus and text feature template are used to train the feature extraction model, and then the preprocessed corpus is input to the trained feature extraction model to obtain the features corresponding to the initial entities output by the feature extraction model. Determine the initial entity.

It is understandable that according to the characteristics of government affairs scenarios, such as policies, regulations, public services, etc., some key words can be determined as seed words through expert knowledge or domain analysis, and then combined with word vector representation, part-of-speech tags, seed word distance and context Information and other features are used to construct a text feature template.

On this basis, corpus and text feature templates are input into the feature extraction model for training, with the purpose of learning how to use different features to extract entity information. Input the preprocessed corpus into the feature extraction model, and use the trained model to extract the initial entities in the corpus and obtain the text features corresponding to the entities. According to the features corresponding to the entities output by the feature extraction model, which entities are valid can be determined according to certain rules and thresholds, that is, the initial entities are determined.

In this implementation, seed words are used to construct text feature templates, which can improve the recall rate and precision of the extraction algorithm. Seed words can guide the algorithm to find relevant entities and relationships in the text, while text feature templates can capture key information and help the algorithm accurately identify and extract entities and relationships.

Knowledge in the field of government affairs is constantly changing and expanding, so the knowledge graph needs to be constantly updated and expanded. In the subsequent process, the text feature template can support the dynamic update and expansion of the knowledge graph by adding and modifying seed words.

In some embodiments, the feature extraction model is a BiLSTM-CRF model. The BiLSTM-CRF model is a deep learning model commonly used in sequence labeling tasks. It combines the advantages of BiLSTM and CRF models.

BiLSTM is a bidirectional recurrent neural network that processes input sequences through forward LSTM and backward LSTM respectively to obtain contextual information and can effectively handle long sequence dependencies. In sequence annotation tasks such as entity recognition in government affairs scenarios, BiLSTM can combine words and contextual information to generate a hidden layer state representation corresponding to each word.

CRF is a conditional random field model that takes into account the dependence between adjacent labels and can perform globally optimal annotation of the entire sequence, thereby improving the accuracy of the model. In the entity recognition task, CRF can use the hidden layer state representation generated by BiLSTM to accurately label entities in the sequence.

The feature extraction model includes input layer, encoding layer and decoding layer.

The input layer is used to convert words or characters in the input corpus text into continuous vector representations. The purpose of the input layer is to convert discrete text input into dense vector representation for subsequent processing. For example, pre-trained word embedding models (such as Word2Vec, GloVe) can be used to obtain vector representations of words.

The encoding layer is used to extract contextual features of the vector sequence of the input corpus text and generate an encoding vector sequence containing semantic information. The encoding layer models the input sequence using a bidirectional long short-term memory network. BiLSTM can effectively capture long-distance dependencies in text by considering both forward and backward contextual information. The function of the encoding layer is to extract contextual features of the input sequence and generate a sequence of encoding vectors containing rich semantic information.

The decoding layer is used to use conditional random fields to annotate the encoding vector sequence output by the encoding layer, predict the label of each annotated position, and generate features corresponding to the output initial entities. The decoding layer takes into account the dependencies between adjacent labels and is able to label the entire sequence based on global probabilities rather than just classifying each position independently. The function of the decoding layer is to use the CRF model to predict the label of each position based on the feature sequence output by the encoding layer, and generate the final labeling sequence to obtain the initial entity.

In this implementation, BiLSTM-CRF can automatically learn features and capture contextual information compared to traditional rule-based or feature engineering methods. Traditional rule-based or feature engineering methods require manual design of features, while the BiLSTM-CRF model can automatically learn the feature representation of the input sequence, including contextual information, word vector representation, etc. This makes the model more flexible and versatile. The BiLSTM-CRF model is a sequence-based model that can use contextual information for annotation. Using context information for annotation is very important for entity extraction tasks, because entities composed of nested words in government affairs scenarios are usually associated with their context, and only by considering context information can entities be identified more accurately.

In addition, the BiLSTM-CRF model can uniformly handle multiple entity types, including person names, place names, organization names, etc. This makes the model more versatile and generalizable.

Step 130: Determine mutual information values between adjacent initial entities, and combine adjacent initial entities with mutual information values greater than the first threshold into the first phrase entity.

It can be understood that after the initial entities are identified, the positions between the initial entities can be identified, and then adjacent initial entities can be identified. For example, "house rental" can be identified as being adjacent to "party".

Mutual information value is a metric used to measure the correlation between two events. Determining the mutual information value between adjacent initial entities may include the following process.

You can first traverse the corpus of government affairs scenes and determine the number of occurrences of two adjacent initial entities and phrases composed of two adjacent initial entities in the corpus of government affairs scenes.

Divide the number of occurrences of two adjacent initial entities by the total number of all words in the corpus to obtain the occurrence probabilities of the two adjacent initial entity words, and use the phrases composed of two adjacent initial entities in the government affairs scene The number of occurrences in the corpus is divided by the total number of all words in the corpus to obtain the occurrence probability of a phrase composed of two adjacent initial entities.

Based on the occurrence probability of two adjacent initial entity words and the occurrence probability of phrases composed of two adjacent initial entities, the mutual information value between adjacent initial entities is determined.

For example, assuming that the preprocessed corpus contains two adjacent words A and B, the mutual information value MI(A, B) of words A and B can be expressed as:

MI(A, B) = log ₂ (P(A, B) / (P(A) * P(B)));

Among them, P(A, B) represents the occurrence probability of the phrase composed of word A and word B, and P(A) and P(B) represent the occurrence probabilities of word A and word B respectively.

It can be understood that the first threshold can be set according to actual conditions. On this basis, adjacent initial entities with mutual information values greater than the first threshold can be combined into the first phrase entity.

In this implementation, by calculating the mutual information value of adjacent initial entities, it can be determined whether there is a correlation between two adjacent initial entities, and then the first phrase entity can be formed, thereby constructing a nested composition of multiple words entities, a richer and more accurate initial entity can be obtained.

Step 140: Determine the left and right entropies of the initial entities in the preprocessed corpus of adjacent initial entities whose mutual information values are greater than the second threshold, and determine the second phrase entity based on the initial entities whose left and right entropies are less than the third threshold.

It should be noted that the second threshold is smaller than the first threshold.

In this case, the initial entities whose mutual information values of adjacent initial entities are greater than the second threshold include initial entities that can constitute the first phrase entity, and also include some initial entities that cannot constitute the first phrase entity.

In some embodiments, the left and right entropies of the initial entities whose mutual information values of adjacent initial entities are greater than the second threshold in the preprocessed corpus are determined, and may be specifically determined in the following manner.

First, determine the size of the word selection window. The size of the word selection window is used to limit the range of left and right neighboring words of the initial entity in the preprocessed corpus.

For example, the window size can be set to 2, which means that two adjacent words to the left and two to the right of the target word are considered.

On this basis, based on the word selection window size, obtain adjacent words in the left and right range of the initial entity whose mutual information value is greater than the second threshold in the corpus of the government scene. For example, sliding can be based on a fixed window size. to get nearby words.

Determine the number of times each adjacent word appears in all word selection windows, and obtain the frequency of each adjacent word in all left word selection windows and all right word selection windows; based on the frequency of each adjacent word in each word selection window The frequency of occurrence within the word determines the left and right entropy of the initial entity corresponding to each adjacent word in the preprocessed corpus.

To determine the number of times each initial entity's neighboring words appear in all word selection windows, and to obtain the frequency of each neighboring word appearing in all left word selection windows and all right word selection windows, you can follow the following process .

Traverse the entire corpus, and for each word selection window, count the number of occurrences of each adjacent word in the window, and add them to the corresponding adjacent word counter. On this basis, count the frequency of all adjacent words in the left word selection window, that is, divide the count of each adjacent word by the total number of words in the left word selection window to obtain the frequency. Similarly, count the frequency of all adjacent words in the right word selection window, that is, divide the count of each adjacent word by the total number of words in the right word selection window to get the frequency.

For each neighboring word, calculate its frequency distribution within all left word selection windows. That is, the frequency of the adjacent word in each left word selection window is divided by the total number of words in the window to obtain the frequency distribution. Information entropy can be used to measure the uncertainty of neighboring words in the word selection window on the left. Calculate the probability of each neighboring word based on the frequency distribution, and then calculate its left entropy value. Similarly, for each neighboring word, the above process is repeated to calculate its frequency distribution and right entropy value in all right word selection windows.

After obtaining the left entropy value and the right entropy value, the average of the two can be calculated as the left and right entropy of the initial entity corresponding to each adjacent word in the preprocessed corpus, or the smaller of the two values can be used as The left and right entropy of the initial entity corresponding to each adjacent word in the preprocessed corpus. There is no restriction on the specific calculation method of left and right entropy here.

In this case, the second phrase entity may be determined based on the initial entity whose left and right entropies are less than the third threshold.

In this embodiment, by calculating the left and right entropy, the second phrase entity with a strong sense of boundary after forming a phrase with some initial entities can be further identified, and further entities formed by possible nesting of multiple words can be further identified.

In some embodiments, determining the second phrase entity based on the initial entity whose left and right entropy is less than the third threshold value specifically includes: determining the left and right entropy of the initial entity whose left and right entropy is less than the third threshold value in the preprocessed corpus. value and the right entropy value; the adjacent words corresponding to the smaller one of the left entropy value and the right entropy value and the initial entity form the second phrase entity.

In this implementation, adjacent words with strong boundaries can be clarified, and then combined with the initial entity to form a second phrase entity, thereby improving the accuracy of entity extraction.

Step 150: Determine the initial entity, the first phrase entity and the second phrase entity whose mutual information value with the adjacent initial entity is less than or equal to the first threshold as the target entity.

It can be understood that the second phrase entity may also include phrase entities that are repeated with the first phrase entity. When identifying the target entity, the repeated phrase entities in the second phrase entity may be filtered out to obtain the target entity.

Step 160: Extract the relationship between the target entities based on the corpus of the government affairs scenario and the target entities.

After identifying the target entities in the government scene corpus, the relationship between the target entities can be extracted based on the government scene corpus and the target entities.

In some embodiments, relationships between target entities may be found based on predefined rules or patterns. For example, in a sentence, if there are specific keywords or phrases between two target entities, the corresponding relationship can be defined based on these keywords or phrases.

In some embodiments, machine learning or deep learning technology can also be used to build a relationship classification model, and take the target entity and its context as input to predict the relationship category between them. In this case, you need to prepare a training data set labeled with relationship categories and use a supervised learning algorithm for model training.

It is understood that the appropriate method can be selected based on the specific task and data characteristics. In practical applications, a combination of multiple technologies and methods can also be used to improve the accuracy and effect of relationship extraction.

Step 170: Store the determined relationships between target entities in a relational database to build a government knowledge graph.

Relational databases are used to store relationships between target entities. There are also entities and entity relationships extracted using other entities in the relational database, which are not limited here. It can be understood that after extracting the relationship between the target entities, the determined relationship between the target entities is stored in the relational database to construct the government knowledge graph.

According to a knowledge graph-based standardized government data construction method provided by an embodiment of the present invention, a feature extraction model is used to identify multiple initial entities in the government scenario by using seed words based on the government scenario, and then adjacent initial entities are used The mutual information value between them is used to obtain the first phrase entity, and based on the mutual information value, the second phrase entity is obtained by calculating the left and right entropy, further expanding the scope of the phrase entity, and finally obtaining the target entity, realizing the embedding of multiple words. By extracting a set of combined phrase entities, we can obtain richer and more accurate entities to build a knowledge graph in government affairs scenarios.

The following is a description of a standardized government data construction device based on a knowledge graph provided by the present invention. The standardized government data construction device based on a knowledge graph described below can be used with the standardized government data construction method based on a knowledge graph described above. mutual reference.

As shown in Figure 2, a standardized government data construction device based on a knowledge graph according to the embodiment of the present invention mainly includes a first processing module 210, a second processing module 220, a third processing module 230, a fourth processing module 240, a fifth The processing module 250, the sixth processing module 260 and the seventh processing module 270.

The first processing module 210 is used to perform data preprocessing on the corpus of government affairs scenarios;

The second processing module 220 is used to use the feature extraction model to identify the initial entities in the preprocessed corpus based on the seed words of the government affairs scenario;

The third processing module 230 is configured to determine mutual information values between adjacent initial entities, and combine adjacent initial entities with mutual information values greater than the first threshold into the first phrase entity;

The fourth processing module 240 is used to determine the left and right entropies of the initial entities whose mutual information values of adjacent initial entities are greater than the second threshold in the preprocessed corpus, and determine the first entity based on the initial entities whose left and right entropies are less than the third threshold. Two phrase entities; the second threshold is smaller than the first threshold;

The fifth processing module 250 is configured to determine the initial entity, the first phrase entity and the second phrase entity whose mutual information value with the adjacent initial entity is less than or equal to the first threshold as the target entity;

The sixth processing module 260 is used to extract relationships between target entities based on the corpus of government affairs scenarios and target entities;

The seventh processing module 270 is used to store the determined relationships between target entities in a relational database to build a government knowledge graph.

According to a standardized government data construction device based on a knowledge graph provided by an embodiment of the present invention, a feature extraction model is used through seed words based on government scenarios to identify multiple initial entities in the government scenario, and then adjacent initial entities are used The mutual information value between them is used to obtain the first phrase entity, and based on the mutual information value, the second phrase entity is obtained by calculating the left and right entropy, further expanding the scope of the phrase entity, and finally obtaining the target entity, realizing the embedding of multiple words. By extracting a set of combined phrase entities, we can obtain richer and more accurate entities to build a knowledge graph in government affairs scenarios.

In some embodiments, the third processing module 230 is also used to traverse the corpus of the government affairs scene, and respectively determine the number of occurrences of two adjacent initial entities and a phrase composed of two adjacent initial entities in the corpus of the government affairs scene; Divide the number of occurrences of two adjacent initial entities by the total number of all words in the corpus to obtain the occurrence probabilities of the two adjacent initial entity words, and use the phrases composed of two adjacent initial entities in the government affairs scene The number of occurrences in the corpus is divided by the total number of all words in the corpus to obtain the occurrence probability of a phrase composed of two adjacent initial entities; based on the occurrence probability of two adjacent initial entity words and two adjacent initial entities The occurrence probability of the composed phrases determines the mutual information value between adjacent initial entities.

In some embodiments, the fourth processing module 240 is also used to determine the size of the word selection window. The size of the word selection window is used to limit the range of left and right adjacent words of the initial entity in the preprocessed corpus; based on the size of the word selection window, obtain The initial entity whose mutual information value of adjacent initial entities is greater than the second threshold has adjacent words in the left and right range in the corpus of the government scene; determine the number of times each adjacent word appears in all word selection windows, and obtain each adjacent word respectively. The frequency of occurrence in all left word selection windows and all right word selection windows; based on the frequency of each neighboring word appearing in each word selection window, determine the initial entity corresponding to each neighboring word in the preprocessed corpus The left and right entropy of .

In some embodiments, the fourth processing module 240 is also used to determine the left entropy value and right entropy value in the left and right entropy of the initial entity whose left and right entropy is less than the third threshold in the preprocessed corpus; convert the left entropy value and right entropy value The adjacent word corresponding to the smaller one of the values and the initial entity form the second phrase entity.

In some embodiments, the second processing module 220 is also used to construct a text feature template based on the seed words of the government affairs scenario; the feature template includes the word vector representation of the word, part-of-speech tag, seed word distance and contextual information; using the initial training corpus and The text feature template trains the feature extraction model; input the preprocessed corpus into the trained feature extraction model, obtain the features corresponding to the initial entities output by the feature extraction model, and determine the initial entities.

In some embodiments, the feature extraction model is a BiLSTM-CRF model. The feature extraction model includes an input layer, an encoding layer and a decoding layer; the input layer is used to convert words or characters in the input corpus text into continuous vector representations; encoding The decoding layer is used to extract the contextual features of the vector sequence of the input corpus text and generate a coding vector sequence containing semantic information; the decoding layer is used to use conditional random fields to annotate the coding vector sequence output by the coding layer, and to label each annotated vector sequence. The label of the location is predicted and the features corresponding to the initial entity are generated and output.

In some embodiments, the seed words of the government scene include words in the government scene-specific dictionary and words manually annotated in the government scene.

Figure 3 illustrates a schematic diagram of the physical structure of an electronic device. As shown in Figure 3, the electronic device may include: a processor (processor) 310, a communications interface (Communications Interface) 320, a memory (memory) 330, and a communication bus 340. Among them, the processor 310, the communication interface 320, and the memory 330 complete communication with each other through the communication bus 340. The processor 310 can call the logical instructions in the memory 330 to execute the standardized government data construction method based on the knowledge graph. The method includes: performing data preprocessing on the corpus of the government scene; using the feature extraction model to identify the seed words based on the government scene. Obtain the initial entities in the preprocessed corpus; determine the mutual information value between adjacent initial entities, and combine the adjacent initial entities with mutual information values greater than the first threshold into the first phrase entity; determine the adjacent initial entities The mutual information value of the initial entity is greater than the left and right entropy of the initial entity in the preprocessed corpus of the second threshold, and the second phrase entity is determined based on the initial entity whose left and right entropy is less than the third threshold; the second threshold is less than the first threshold ; Determine the initial entity, the first phrase entity and the second phrase entity whose mutual information value with the adjacent initial entity is less than or equal to the first threshold as the target entity; based on the corpus of the government affairs scenario and the target entity, extract the target entity The relationship between them; store the determined relationship between the target entities in the relational database to build a government knowledge graph.

In addition, the above-mentioned logical instructions in the memory 330 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on this understanding, the technical solution of the present invention essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program code. .

On the other hand, the present invention also provides a computer program product. The computer program product includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can Execute the standardized government data construction method based on the knowledge graph provided by the above methods. The method includes: performing data preprocessing on the corpus of government scenes; based on the seed words of the government scene, using the feature extraction model to identify the preprocessed corpus initial entities; determine the mutual information values between adjacent initial entities, and combine adjacent initial entities with mutual information values greater than the first threshold into the first phrase entity; determine the mutual information values of adjacent initial entities greater than The left and right entropy of the initial entity of the second threshold in the preprocessed corpus is determined, and the second phrase entity is determined based on the initial entity whose left and right entropy is less than the third threshold; the second threshold is less than the first threshold; and the adjacent initial entity is The initial entity, the first phrase entity and the second phrase entity whose mutual information value between entities is less than or equal to the first threshold are determined as target entities; based on the corpus of the government affairs scenario and the target entity, the relationship between the target entities is extracted; the determined The relationships between the target entities are stored in the relational database to build a government knowledge graph.

In another aspect, the present invention also provides a non-transitory computer-readable storage medium on which a computer program is stored. The computer program is implemented when executed by the processor to perform the standardized government data construction based on the knowledge graph provided by the above methods. Method, the method includes: performing data preprocessing on a corpus of government affairs scenarios; using a feature extraction model to identify initial entities in the preprocessed corpus based on the seed words of the government affairs scenario; determining mutual information between adjacent initial entities value, and combine the adjacent initial entities whose mutual information value is greater than the first threshold into the first phrase entity; determine the left and right of the initial entities whose mutual information value is greater than the second threshold of the adjacent initial entities in the preprocessed corpus entropy, and determine the second phrase entity based on the initial entity whose left and right entropy is less than the third threshold; the second threshold is less than the first threshold; the initial entity with the mutual information value between the adjacent initial entity is less than or equal to the first threshold The entity, the first phrase entity and the second phrase entity are determined as the target entity; based on the corpus of the government affairs scenario and the target entity, the relationship between the target entities is extracted; the determined relationship between the target entities is stored in the relational database to Build a government knowledge graph.

The device embodiments described above are only illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in One location, or it can be distributed across multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Persons of ordinary skill in the art can understand and implement the method without any creative effort.

Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the part of the above technical solution that essentially contributes to the existing technology can be embodied in the form of a software product. The computer software product can be stored in a computer-readable storage medium, such as ROM/RAM, magnetic disc, optical disk, etc., including a number of instructions to cause a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that it can still be used Modifications are made to the technical solutions described in the foregoing embodiments, or equivalent substitutions are made to some of the technical features; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims (10)

1. The standardized government affair data construction method based on the knowledge graph is characterized by comprising the following steps of:

carrying out data preprocessing on a corpus of government affair scenes;

based on the seed words of the government affair scene, identifying initial entities in the preprocessed corpus by utilizing a feature extraction model;

determining mutual information values between adjacent initial entities, and combining the adjacent initial entities with the mutual information values larger than a first threshold value into a first phrase entity;

determining left and right entropy of an initial entity with mutual information value larger than a second threshold value in the preprocessed corpus of adjacent initial entities, and determining a second phrase entity based on the initial entity with left and right entropy smaller than a third threshold value; the second threshold is less than the first threshold;

Determining an initial entity, the first phrase entity and the second phrase entity, of which the mutual information values with adjacent initial entities are smaller than or equal to the first threshold value, as target entities;

based on a corpus of government scenes and the target entities, extracting the relation between the target entities;

and storing the determined relation between the target entities in a relation database to construct a government knowledge graph.

2. The knowledge-graph-based standardized government data construction method according to claim 1, wherein the determining the mutual information value between adjacent initial entities includes:

traversing a corpus of the government affair scene, and respectively determining two adjacent initial entities and the occurrence times of phrases formed by the two adjacent initial entities in the corpus of the government affair scene;

dividing the occurrence times of two adjacent initial entities by the total number of all words in the corpus respectively to obtain the occurrence probability of the words of the two adjacent initial entities, and dividing the occurrence times of phrases formed by the two adjacent initial entities in the corpus of the government scene by the total number of all words in the corpus to obtain the occurrence probability of phrases formed by the two adjacent initial entities;

And determining a mutual information value between the adjacent initial entities based on the occurrence probabilities of the two adjacent initial entity words and the occurrence probabilities of phrases formed by the two adjacent initial entities.

3. The knowledge-graph-based standardized government data construction method according to claim 1, wherein the determining the left and right entropy of the initial entity with the mutual information value of the adjacent initial entities greater than the second threshold in the preprocessed corpus comprises:

determining a word selecting window size, wherein the word selecting window size is used for limiting the range of left and right adjacent words of the initial entity in the preprocessed corpus;

based on the size of the word selecting window, acquiring adjacent words of the initial entities, of which the mutual information values are larger than a second threshold, in the left-right range in the corpus of the government affair scene;

determining the frequency of occurrence of each adjacent word in all word selecting windows, and respectively obtaining the occurrence frequency of each adjacent word in all left word selecting windows and all right word selecting windows;

based on the frequency of each adjacent word in each word selecting window, the left and right entropy of the initial entity corresponding to each adjacent word in the preprocessed corpus is determined.

4. The knowledge-graph-based standardized government data construction method of claim 3, wherein the determining the second phrase entity based on the initial entity having the left-right entropy less than the third threshold includes:

determining left entropy values and right entropy values in left entropy and right entropy values of an initial entity with left entropy and right entropy smaller than a third threshold in the preprocessed corpus;

and forming the second phrase entity by the adjacent word corresponding to the smaller one of the left entropy value and the right entropy value and the initial entity.

5. The knowledge-graph-based standardized government affair data construction method according to claim 1, wherein the identifying the initial entity in the preprocessed corpus by using the feature extraction model based on the seed words of the government affair scene comprises:

constructing a text feature template based on the seed words of the government affair scene; the feature templates comprise word vector representations of words, part-of-speech tags, seed word distances and context information;

training the feature extraction model by using an initial training corpus and the text feature template;

inputting the corpus in the preprocessed corpus into the trained feature extraction model to obtain the features corresponding to the initial entity output by the feature extraction model, and determining the initial entity.

6. The knowledge-graph-based standardized government data construction method of claim 5 wherein the feature extraction model is a BiLSTM-CRF model, the feature extraction model including an input layer, an encoding layer and a decoding layer;

the input layer is used for converting words or characters in the input corpus text into continuous vector representations;

the coding layer is used for extracting the context characteristics of the vector sequence of the input corpus text and generating a coding vector sequence containing semantic information;

the decoding layer is used for marking the coded vector sequence output by the coding layer by using a conditional random field, predicting the label of each marked position and generating and outputting the characteristics corresponding to the initial entity.

7. The knowledge-graph-based standardized government affair data construction method according to claim 1, wherein the seed words of the government affair scene include words in a dictionary specific to the government affair scene and words artificially marked in the government affair scene.

8. A standardized government affair data construction device based on a knowledge graph is characterized in that,

the first processing module is used for preprocessing data of a corpus of government affair scenes;

The second processing module is used for identifying an initial entity in the preprocessed corpus by utilizing the feature extraction model based on the seed words of the government affair scene;

the third processing module is used for determining mutual information values between adjacent initial entities and combining the adjacent initial entities with the mutual information values larger than a first threshold value into a first phrase entity;

the fourth processing module is used for determining left and right entropy of the initial entity with the mutual information value larger than a second threshold value in the preprocessed corpus of the adjacent initial entities and determining a second phrase entity based on the initial entity with the left and right entropy smaller than a third threshold value; the second threshold is less than the first threshold;

a fifth processing module, configured to determine an initial entity, the first phrase entity, and the second phrase entity, where a mutual information value between the first phrase entity and an adjacent initial entity is less than or equal to the first threshold, as a target entity;

the sixth processing module is used for extracting the relation between the target entities based on the corpus of government affair scenes and the target entities;

and the seventh processing module is used for storing the determined relation between the target entities in a relation database so as to construct a government knowledge graph.

9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the knowledge-graph-based standardized government data construction method of any of claims 1 to 7 when the program is executed by the processor.

10. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the knowledge-graph-based standardized government data construction method of any of claims 1 to 7.