In the embodiment of the present invention, the identifying an entity relation between key entities in the key entity set and a first relation weight corresponding to the entity relation includes:

- extracting, from the enterprise relation instance, an entity relation corresponding to key entities in the key entity set; and
- acquiring an occurrence frequency of the entity relation within a preset period of time, and using the occurrence frequency as a first relation weight corresponding to the entity relation.

Specifically, the enterprise relation instance contains an enterprise relation type, for example, competition, cooperation, holding, supply, merger, merger and acquisition or the like. The extracted entity relation corresponding to the key entities is the enterprise relation type. The obtained occurrence frequency of the entity relation within the preset period of time refers to the ratio of the number of occurrences of a certain business relation to the number of occurrences of all business relations between two enterprise entities, i.e., the occurrence probability. The occurrence probability is used as the first relation weight corresponding to the entity relation.

Specifically, the performing, by using a preset temporal order interpolation algorithm, weight calculation on the enterprise text information after deletion of the enterprise relation instance to obtain a second relation weight includes:

- when the enterprise text information does not contain enterprise relation instances, backtracking historical text information corresponding to the enterprise text information; and
- extracting a historical entity relation in the historical text information, calculating the weight of the historical entity relation according to the temporal order interpolation algorithm, and using the weight of the historical entity relation as a second relation weight.

Specifically, the enterprise text information not containing enterprise relation instances means that there is no relation instance in the enterprise text information, and the entity relation in the relation instance refers to a continuous entity relation between enterprises. The continuous entity relation means that the relation will last for a period of time but is not a one-time event relation. For example, competition, cooperation, holding and supply are all continuous competitive relations.

When the enterprise text information does not contain enterprise relation instances, it is impossible to directly acquire the corresponding weight, and it is necessary to backtrack the historical text information corresponding to the enterprise text information. For example, if there was no competitive relation between enterprise A and enterprise B in June 2000, that is, there was no relation instance in June 2000, in this case, it is impossible to calculate the corresponding weight according to the enterprise text information. Thus, it is necessary to backtrack the historical text information of this relation in January 2000 and calculate the weight corresponding to the historical entity relation in the historical text information, and the weight corresponding to the historical entity relation is used as the second relation weight.

Preferably, a weight of an entity relation when no enterprise business relation occurs within a specified period of time is calculated by using the temporal order interpolation algorithm, wherein the temporal order interpolation algorithm may be a linear interpolation algorithm or a polynomial interpolation algorithm. Thus, any continuous relation between any enterprises within the specified period of time has a weight at any time, thereby ensuring the accuracy of data.

In S4, an entity relation having the first relation weight or the second relation weight satisfying a preset reference condition is used as a standard entity relation.

In the embodiment of the present invention, both the first relation weight and the second relation weight are used for describing the occurrence frequency of an entity relation, wherein the first relation weight is calculated when the enterprise relation instance contains entity relations, and the second relation weight is obtained by calculating the historical text information using the preset temporal order interpolation algorithm when the enterprise relation instance does not contain entity relations. The first relation weight or the second relation weight is compared with the preset reference condition, and an entity relation satisfying the preset reference condition is used as a standard entity relation.

In the embodiment of the present invention, an enterprise relation instance in enterprise text information is extracted to obtain data after preliminary screening of text information, and key entities in the enterprise relation instance are extracted by using a graph sorting algorithm, so that the key entities can be obtained accurately; a first relation weight of the key entities and a second relation weight of the enterprise text information after deletion of the enterprise relation instance are identified, respectively; and an entity relation satisfying a preset reference condition is screened out according to the first relation weight and the second relation weight. The occurrence frequency of the entity relation is analyzed from the perspective of weight by using the first relation weight and the second relation weight, so the screened standard entity relation is more accurate and more efficient. Therefore, the entity relation mining method provided by the present invention can solve the problem of low accuracy in entity relation mining.

As shown in FIG. 3, a functional module diagram of an entity relation mining apparatus according to an embodiment of the present invention is shown.

The entity relation mining apparatus 100 described herein may be installed in an electronic device. Depending upon the realized functions, the entity relation mining apparatus 100 may include an instance extraction module 101, an entity extraction module 102, a weight calculation module 103 and a relation mining module 104. The modules described herein may also called units, which refer to a series of computer program segments that can be executed by a processor of the electronic device and can complete fixed functions, and are stored in a memory of the electronic device.

In this embodiment, the function of each module/unit is described below:

- the instance extraction module 101 is configured to acquire enterprise text information and extract an enterprise relation instance in the enterprise text information;
- the entity extraction module 102 is configured to perform key entity extraction on the enterprise relation instance by using a preset graph sorting algorithm to obtain a key entity set;
- the weight calculation module 103 is configured to identify an entity relation between key entities in the key entity set and a first relation weight corresponding to the entity relation, delete the enterprise relation instance from the enterprise text information, and perform, by using a preset temporal order interpolation algorithm, weight calculation on the enterprise text information after deletion of the enterprise relation instance to obtain a second relation weight; and
- the relation mining module 104 is configured to use an entity relation having the first relation weight or the second relation weight satisfying a preset reference condition as a standard entity relation.

Specifically, the specific implementation modes of the modules in the entity relation mining apparatus 100 are described below.

In step 1, enterprise text information is acquired, and an enterprise relation instance in the enterprise text information is extracted.

In the embodiment of the present invention, the enterprise text information refers to various forms of texts related to enterprise information in the field of enterprise business relation, for example, news articles, industry reports, technical articles or the like.

Specifically, the extracting an enterprise relation instance in the enterprise text information includes:

- acquiring a preset data structure template, and searching structure data corresponding to the preset data structure template from the enterprise text information; and
- performing data combination on the structure data to obtain an enterprise relation instance.

Specifically, the preset data structure template contains different template entries, such as entity name, relation type, data and source. Text information data corresponding to the template entries is searched from the enterprise text information, and the text information data is combined to obtain an enterprise relation instance.

For example, the enterprise relation instance may be that the enterprise name is company A and company B, and the relation type may be competition, cooperation, holding, supply, merger, merger and acquisition or the like. The date may be a specific time limit, and the source refers to a website or document to which the enterprise text information belongs.

In step 2, key entity extraction is performed on the enterprise relation instance by using a preset graph sorting algorithm to obtain a key entity set.

In the embodiment of the present invention, the performing key entity extraction on the enterprise relation instance by using a preset graph sorting algorithm to obtain a key entity set includes:

- performing instance analysis on the enterprise relation instance to obtain instance analysis data;
- performing tokenization and screening on the instance analysis data to obtain a standard participle set;
- constructing a directed weighted graph according to standard tokens in the standard token set, and calculating node weights corresponding to a plurality of nodes in the directed weighted graph; and
- using the nodes having node weights exceeding a preset threshold in the directed weighted graph as key entities, and performing entity collection on the plurality of key entities to obtain a key entity set.

Specifically, the entity analysis refers to the extraction of data under different template entries in the enterprise relation instance, and the tokenization may be implemented by a benchmark tokenizer. The graph sorting algorithm may be a TextRank model that can extract key entities in a text.

The TextRank model is a graph-based sorting model for text. Its basic idea is to divide a text into several constituent units (words or sentences) and establish a graph model which is essentially a directed weighted graph.

Specifically, the performing tokenization and screening on the instance analysis data to obtain a standard token set includes:

- performing sentence segmentation on the instance analysis data according to a preset symbol to obtain a segmented sentence set;
- deleting stop words in the segmented sentence set, and performing tokenization on segmented sentences after deletion of the stop words to obtain an initial token set; and
- performing part-of-speech tagging on a plurality of initial tokens in the initial token set, and reserving initial tokens consistent with a preset specified part of speech as standard tokens to obtain a standard token set.

Specifically, the preset symbol may be a full stop, and the deletion of the stop words in the segmented sentence set may be performed by comparison with reference to the existing stop word removal list. The tokenization may adopt a trained tokenization model or a benchmark tokenizer, and the preset specified part of speech may be noun, verb and adjective.

In the embodiment of the present invention, the node weight is calculated by the following formula:

$WS (V_{i}) = (1 - d) + d * \sum_{V_{j} \in In (V_{i})} \frac{W_{ji}}{\sum_{V_{k} \in Out (V_{j})} W_{jk}} WS (V_{j})$

where WS(V_i) represents a node weight of a node V_i, d is a damping coefficient, ln(V_i) is a first node set pointing to the node V_i, Out(V_j) is a second node set pointed by the node V_i, W_jiis a connection weight between nodes V_iand V_j, and W_jkis a connection weight between nodes V_kand V_j.

The damping coefficient d represents the probability of pointing from a specific point to any other point in the directed weighted graph, and the damping coefficient preferably has a value of 0.85.

In step 3, an entity relation between key entities in the key entity set and a first relation weight corresponding to the entity relation are identified, the enterprise relation instance is deleted from the enterprise text information, and weight calculation is performed on the enterprise text information after deletion of the enterprise relation instance by using a preset temporal order interpolation algorithm to obtain a second relation weight.

In the embodiment of the present invention, the identifying an entity relation between key entities in the key entity set and a first relation weight corresponding to the entity relation includes:

- extracting, from the enterprise relation instance, an entity relation corresponding to key entities in the key entity set; and
- acquiring an occurrence frequency of the entity relation within a preset period of time, and using the occurrence frequency as a first relation weight corresponding to the entity relation.

- when the enterprise text information does not contain enterprise relation instances, backtracking historical text information corresponding to the enterprise text information; and
- extracting a historical entity relation in the historical text information, calculating the weight of the historical entity relation according to the temporal order interpolation algorithm, and using the weight of the historical entity relation as a second relation weight.

Preferably, the weight of an entity relation when no enterprise business relation occurs within a specific period of time is calculated by using the temporal order interpolation algorithm, wherein the temporal order interpolation algorithm may be a linear interpolation algorithm or a polynomial interpolation algorithm. Thus, any continuous relation between any enterprises within the specified period of time has a weight at any time, thereby ensuring the accuracy of data.

In step 4, an entity relation having the first relation weight or the second relation weight meeting a preset reference condition is used as a standard entity relation.

In the embodiment of the present invention, both the first relation weight and the second relation weight are used for describing the occurrence frequency of an entity relation, wherein the first relation weight is calculated when the enterprise relation instance contains entity relations, and the second relation weight is obtained by calculating the historical text information using the preset temporal order interpolation algorithm when the enterprise relation instance does not contain entity relations. The first relation weight or the second relation weight is compared with the preset reference condition, and an entity relation meeting the preset reference condition is used as a standard entity relation.

In the embodiment of the present invention, an enterprise relation instance in enterprise text information is extracted to obtain data after preliminary screening of text information, and key entities in the enterprise relation instance are extracted by using a graph sorting algorithm, so that the key entities can be obtained accurately; a first relation weight of the key entities and a second relation weight of the enterprise text information after deletion of the enterprise relation instance are identified, respectively; and an entity relation satisfying a preset reference condition is screened out according to the first relation weight and the second relation weight. The occurrence frequency of the entity relation is analyzed from the perspective of weight by using the first relation weight and the second relation weight, so the screened standard entity relation is more accurate and more efficient. Therefore, the entity relation mining apparatus provided by the present invention can solve the problem of low accuracy in entity relation mining.

As shown in FIG. 4, a schematic structure diagram of an electronic device for implementing an entity relation mining method according to an embodiment of the present invention is shown.

The electronic device 1 may include a processor, a memory 11, a communication bus 12 and a communication interface 13, and may further include computer programs, such as an entity relation mining program, which are stored in the memory 11 and can be run on the processor 10.

In some embodiments, the processor 10 may be composed of integrated circuits, for example, being composed of a single packaged integrated circuit, or being composed of a plurality of integrated circuits packaged with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, combinations of various control chips or the like. The processor 10, as a control unit of the electronic device, connects the components of the whole electronic device by using various interfaces and lines, and executes various functions of the electronic device and processes data by running or executing the programs or modules (e.g., executing the entity relation mining program, etc.) stored in the memory 11 and calling the data stored in the memory 11.

The memory 11 includes at least one type of readable storage medium, and the readable storage mediums include flash memories, mobile hard disks, multimedia cards, card-type memories (SD or DX memories, etc.), magnetic memories, magnetic disks, optical disks, etc. In some embodiments, the memory