2.1

Knowledge map

Knowledge mapping is composed of entities, the relationship between entities and entities, and the attributes of entities. The knowledge graph is composed of pieces of knowledge, that is, SPO triples (subject-predicate-object).

Knowledge mapping is to represent knowledge, convert knowledge that people can understand into the form of mapping, so that machines can also understand natural language. The concept of knowledge map can be traced back to the semantic network proposed in the 1950s and 1960s. Semantic network is a form of knowledge representation, which is composed of interconnected nodes and edges. Nodes represent concepts or objects, and edges represent relationships between them.

The construction of knowledge mapping starts from the most original data (including structured data, semi-structured data and unstructured data), and adopts the technical means in the fields of natural language processing and data mining to extract knowledge facts from the original database and store them into the knowledge base. This process includes three main processes: knowledge extraction, knowledge fusion and knowledge processing. Each update iteration contains these three phases.

2.2

Graph database

Through the steps of knowledge extraction and knowledge fusion in the process of building knowledge map, the required knowledge is obtained, and then the knowledge needs to be persisted for use. At present, in the knowledge map, the storage form of knowledge map is divided into two types according to the storage type: storage based on table structure (RDF) and storage based on graph structure (graph database). RDF (resourse description framework), or Resource Description Framework, is a standard data model developed by W3C (World Wide Web Consortium) for describing entities/resources. In addition, relational databases can also store knowledge maps, which are actually composed of triples. Relational databases pay more attention to the internal attributes of entities, and the relationships between entities are usually realized by foreign keys. The relational database based on the table structure often requires time-consuming join operations.For most of the existing terabyte-level data, the relational database is often unable to meet the speed requirements, while the graph database technology makes up for this shortcoming, showing amazing performance when dealing with massive data, and plays an important role in the knowledge graph storage.

Graph database is widely used in various fields. It is a kind of non-relational database, which stores a graph structure. Graph is a complex nonlinear structure. Graph database consists of two important parts: node and relation, and each entity is a node. Nodes can have many attributes, and nodes are connected by edges (relationships), which are directional. In the graph structure, the relationship between nodes does not have only one direct predecessor and direct successor as in the linear structure, nor does it contain only hierarchical relationships as in the tree structure. The relationship between nodes is arbitrary, and any two data elements in the graph may be related.

Unlike relational databases, graph database relationships are not represented using foreign keys, but rather using edges for relational representation. For graph database, it can support knowledge storage and query with large amount of data and complex association. At present, the mainstream graph databases are Neo4j, Orient DB, Microsoft Azure, Cosmos DB and so on. There are many differences between graph database and RDF, as shown in Table 1:

Table 1.

Comparison between RDF and graph database.

Unlike relational databases, graph database relationships are not represented using foreign keys, but rather using edges for relational representation. For graph database, it can support knowledge storage and query with large amount of data and complex association. At present, the mainstream graph databases are Neo4j, Orient DB, Microsoft Azure, Cosmos DB and so on. There are many differences between graph database and RDF, as shown in Table 1:

2.3

Word2Vec

In terms of language processing, the most granular unit is the word. Sentences are formed from words, and then paragraphs and chapters are formed from sentences. Therefore, to deal with natural language, we must first deal with words. There is a task to determine whether a word is a verb or a noun. If we use the method of machine learning, we already have a series of labeled samples (x, y), where x represents the word and y represents the part of speech corresponding to the word. We want to identify the part of speech of the unlabeled word. The idea is to build a mapping of f (x)-> y, that is, to create a neural network whose input is the word and its output is the corresponding part of speech of the word. Train this model to get the ability to identify the part of speech of words. But there is a problem here, usually the input words are natural language, which is an abstract summary of human beings and can not be understood by machines, so we need to find ways to convert natural language into a form that machines can understand, that is, numerical form. This way of converting natural language into numerical form is called word embedding. Word 2vec is a kind of word embedding.

The following is a brief introduction to the implementation process of Word2vec. Taking Chinese as an example, a high-frequency dictionary is constructed according to all the text data in the data set, that is, the sentence is segmented at the character level to form a dictionary.For example, given the sentence:

“This,this,this,is,is,is,power,power,power,voltage,voltage,voltage,transformer,transformer,transformer.” The dictionary constructed is shown in the Table 2:

Table 2.

High frequency dictionary.

Each word corresponds to an index, indicating the arrangement position of the current word. Take “Power voltage transformer” as an example to perform the Word2vec operation steps:

that length of the dictionary in Table 2 is 5, and an all-zero vector VEC = [0, 0, 0, 0, 0] with a dimension of 5 is initialize;

2.4

Search engine

2.5

Semantic search

Tim Berners-Lee, the father of the World Wide Web, explained that “the essence of semantic search is to use mathematics to get rid of the guesses and approximations used in today’s search, and to introduce a clear understanding of the meaning of words and how they relate to what we find in search engine input boxes”.

With the emergence of the concept of Semantic Web, more and more open linked data and user-generated content are published on the Internet, and the Internet has been transformed into a data network that contains a large number of entities and relationships between entities. In this context, Google proposed the concept of knowledge graph in May 2012, which aims to describe the relationship between various entities in the real world, so as to improve search results. Following that, Sogou put forward “knowledge cube”, Microsoft put forward “Probase” and Baidu put forward “intimate”.

The working principle of semantic search engine is that it not only pays attention to the user input content, but also pays attention to the meaning expressed by the user input content, accurately understands the real intention of the user, searches with this semantic information, and can return more accurate search results to the user. Compared with the traditional keyword-based search engine, there is a great progress.

3.1

Ontology library

Ontology is the base of knowledge base. Ontology is at the conceptual level, which is similar to the class in programming language. Instance is a concrete presentation of ontology, which is similar to the instantiation in programming language.

When establishing a knowledge base, an ontology base should be established first. For example: I establish an ontology base, in which only one class is established: equipment, which has attributes such as service life, date of manufacture, model, etc., and other relationships are not considered for the time being; supplementary example data: transformer, transmission line, transformer, etc.

Therefore, a knowledge base is established to store the power equipment ontology and the instance data of the power equipment.

3.2

Construction of Corpus for Power Equipment Knowledge Annotation

In this paper, a corpus of power equipment knowledge annotation is built, which contains 4000 power equipment knowledge related data from provincial power grid companies, as well as unstructured or semi-structured data related to power equipment on the Internet. Due to the differences in the structure of the data obtained from various sources, after data cleaning work such as desensitization and non-text content processing, the regular matching method based on rules is used to extract the power equipment knowledge text of the data and store it in MYSQL database. We have also worked with power industry experts to develop a set of marking rules, using the YEDDA marking tool for manual marking. The marking personnel are all composed of power industry workers. The marking work is carried out in a back-to-back manner, and the final review is completed by power industry experts.

5.1

Necessity of introducing knowledge map

The traditional keyword-based search system has the problem of low precision and recall, which is due to the fact that the machine can not understand the meaning of the user’s input content, resulting in the search answer returned by the machine often can not meet the user’s needs. Based on the knowledge map of power equipment constructed above, the power equipment knowledge is stored in the graph database Neo4j, and the traditional keyword-based power equipment knowledge search system is improved through word segmentation of user input sentences, entity relationship extraction, and graph database query technology.

As the core of knowledge map, the establishment process of knowledge base is a continuous process. The content of knowledge base is not immutable, but needs to be updated and integrated iteratively. Whenever there is a user searching, through the intelligent semantic understanding of the query sentence input by the user, the stored content is automatically retrieved and matched in the knowledge base, and the results are presented in a visual way.

The disadvantages of the traditional search model are that it is difficult to understand the user’s intention, to match accurately, and to provide personalized services. The introduction of knowledge graph can solve the following reasons: knowledge graph can express the association between query and answer, and provide interpretable basis for search.

5.2

Semantic search method based on knowledge map

This paper combines the traditional keyword search method and the entity query method based on knowledge map, first uses the traditional keywords to roughly determine the search scope, and then uses the knowledge map subgraph query to accurately search the semantic. The basic process comprises the following steps of: 1) identifying keywords of the content input by a user, and positioning a candidate knowledge map subgraph matched with the search content, thereby accelerating the efficiency of entity search in the knowledge map; 2) identifying an entity of the content input by the user, generating a Neo4j Cypher query statement, and performing entity search on the positioned knowledge map subgraph; 3) discover the relationship between entities in the user’s query content through the graph database query, and then understand the user’s search intention; 4) sort the search results (by the importance in the knowledge graph structure, by the popularity of entities, and by the relevance to the query). The semantic search process based on knowledge mapping is shown in Figure 3.

Fig. 3.

Semantic Search Process Based on Knowledge Map

The traditional keyword search method, because the machine can not understand the user’s search intention, so the search results usually have a large error, while the semantic search method based on knowledge map, when querying, because it needs to traverse the entire knowledge map, it will take a lot of time. Combining the advantages of the two, in the initial stage of search, the keyword search is used to quickly locate the target search area, and then the knowledge map subgraph query is carried out in this area, which not only ensures the accuracy of the search results, but also improves the overall search efficiency.