A radar-oriented joint entity relation extraction method

Abstract

The application discloses a radar-oriented entity relationship joint extraction method, first, Chinese word segmentation and vectorization processing are performed on radar text corpus, then the number of entities in the radar text corpus and the start / stop position of a single entity are acquired based on the word vector, and the entity vector of each entity is calculated; then, the relationship vector between any two entities and the relationship vector based on the attention weight are calculated, the two relationship vectors are spliced and input into a full connection neural network, and finally, the entity relationship is extracted through the full connection neural network, so that the entity relationship about the radar feature is quickly extracted from the radar field sample corpus.

Description

A Joint Entity Relationship Extraction Method for Radar Technical Field

[0001] This invention belongs to the field of radar technology, and more specifically, relates to a method for joint extraction of entity relationships for radar applications. Background Technology

[0002] With the rapid development of computer technology in modern society, coupled with the widespread use of artificial intelligence and knowledge graph technologies, the traditional method of manually judging possible radar models based on certain known parameters lacks theoretical basis and often relies on the potential judgment rules of experts, which has certain limitations and is difficult to meet the requirements of improving work efficiency and judgment accuracy.

[0003] The types of radars are constantly emerging, and they are updated and replaced rapidly, with diverse radar parameters. The method of manually predicting radar models based on partial observation parameters and relying on expert experience has reached its limit, and there is vast room for improvement in the era of big data.

[0004] Entity extraction and relation extraction are tasks that intersect natural language processing and knowledge graph artificial intelligence technologies. Their objective is to automatically extract triplet knowledge from unstructured text data, i.e., data in the form of <entity-relation-entity> or <entity-attribute-attribute value>. The high-quality data extracted in a fixed format is a crucial component in building a knowledge base for the radar domain. Furthermore, the accuracy of the data directly impacts the accuracy of downstream tasks; therefore, entity extraction and relation extraction are critical steps before practical applications.

[0005] Currently, most radar knowledge bases rely on manual construction, which is insufficient in scale to support downstream AI-based tasks and makes it difficult to extract key information from big data. Researching entity extraction and relation extraction techniques for the radar field can directly guide this task towards the big data era, enabling the automated construction of radar knowledge bases. This ensures both efficient knowledge extraction and reliable data quality, which is crucial for carrying out downstream tasks. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a joint entity relation extraction method for radar, which can quickly extract entity relations related to radar features from radar domain sample corpora.

[0007] To achieve the above-mentioned objectives, the present invention provides a radar-based entity relationship joint extraction method, characterized by comprising the following steps:

[0008] (1) Perform Chinese word segmentation on the radar text corpus, and then vectorize the Chinese words according to the Chinese word vectors in the training library FastText to obtain the vectors corresponding to the radar text corpus. ,in, Indicates the first Vector representation of Chinese words, Represents the vector dimension. This indicates the number of Chinese words obtained after word segmentation of the radar text corpus.

[0009] (2) Obtain the number of entities in the radar text corpus and the start or end position of a single entity;

[0010] (2.1) Transform the vector The input is fed into the Transformer model, where the encoder extracts a vector incorporating contextual information. , Indicates the first Vector representation of each Chinese word combined with contextual information;

[0011] (2.2) Set up two parallel fully connected neural networks, where the first fully connected neural network is used to determine the starting position of the entity and the other fully connected neural network is used to determine the ending position of the entity.

[0012] vector The input is simultaneously fed into two parallel fully connected neural networks, and the output of the fully connected neural networks is a sequence of "0" or "1", where "1" represents the start or end position of an entity, and "0" represents a position that is not the start or end position of an entity.

[0013] In the sequence output by the first fully connected neural network, the position of all "1"s is marked as the starting position of the entity. In the sequence output by the second fully connected neural network, the position of all "1"s is marked as the ending position of the entity. Then the number of entities is counted.

[0014] Finally, following the principle of proximity, the position of the first occurrence of "1" in the first sequence and the position of the first occurrence of "1" in the second sequence are taken as the start or end position of the first entity, and so on.

[0015] (3) Calculate the entity vector of each entity;

[0016] Based on the start or end position of each entity, for an entity composed of a single Chinese word, the vector of the Chinese word combined with the context information is used to represent the entity vector; for an entity composed of multiple Chinese words, the vectors corresponding to the multiple Chinese words are added together and then normalized to serve as the entity vector.

[0017] (4) Calculate the relationship vector between any two entities;

[0018] Two entities are randomly selected, one denoted as the head entity, and the corresponding entity vector is denoted as... The other is denoted as the tail entity, and the corresponding entity vector is denoted as... ;

[0019] Using tail entity vectors Subtract head entity vector This yields the relationship vector between entities. ;

[0020] (5) Calculate the relationship vector between any two entities based on attention weights;

[0021]

[0022] in, Indicates the first Attention weights for each Chinese word. Indicates the first Vector representation of each Chinese word combined with contextual information Indicates the number of Chinese words between two entities. A relationship vector between two entities based on attention weights;

[0023] (6) Extract entity relations;

[0024] (6.1) Transfer the relation vector splicing relation vectors After that, I got one dimensional relation vector ;

[0025] (6.2) Transfer the relation vector Input to dimension The hidden layer dimension is The output dimension is A fully connected neural network, thus obtaining We calculate a probability value and then use the Softmax function to select the relationship corresponding to the highest probability as the relationship between entity pairs.

[0026] The objective of this invention is achieved as follows:

[0027] This invention provides a joint entity relation extraction method for radar. First, the radar text corpus is segmented into Chinese and vectorized. Then, the number of entities in the radar text corpus and the start / end position of each entity are obtained based on the word vectors, and the entity vector of each entity is calculated. Next, the relation vector between any two entities and the relation vector based on attention weights are calculated. The two relation vectors are concatenated and input into a fully connected neural network. Finally, the entity relations are extracted through the fully connected neural network.

[0028] Furthermore, the entity relationship joint extraction method for radar in this invention also has the following beneficial effects:

[0029] (1) The present invention provides a joint entity relationship extraction method for radar, which can effectively handle the entity overlap problem. Compared with the traditional extraction method of first extracting entities and then extracting relationships, it utilizes the interaction and inherent connection between the two tasks to alleviate the error accumulation and transmission of the two independent tasks.

[0030] (2) The present invention provides a joint entity relationship extraction method for radar, proposing a relationship feature extraction method based on the TransE model. This method effectively extracts the relationship features between entity pairs by modeling the relationship as the subtraction of two entity vectors.

[0031] (3) The present invention provides a joint entity relationship extraction method for radar, which proposes a relationship feature extraction method based on an attention mechanism. This method utilizes the characteristic that the relationship between entity pairs is related to the text between the two entities. By judging the similarity between the Chinese words between entities and the head and tail entities, a certain relationship weight is assigned to the words. This weight is multiplied by its corresponding word vector and accumulated, thereby extracting the relationship features between entity pairs, which improves the accuracy of joint entity relationship extraction in the radar field to a certain extent. Attached Figure Description

[0032] Figure 1 is a flowchart of a radar-oriented entity relationship joint extraction method according to the present invention;

[0033] Figure 2 is a schematic diagram of the start and end positions of the entity. Detailed Implementation

[0034] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0035] Example

[0036] Figure 1 is a flowchart of a radar-oriented entity relationship joint extraction method according to the present invention.

[0037] In this embodiment, as shown in Figure 1, the present invention provides a radar-oriented entity relationship joint extraction method, which includes the following steps:

[0038] S1. Perform Chinese word segmentation on the radar text corpus, and then vectorize the Chinese words according to the Chinese word vectors in the training library FastText to obtain the vectors corresponding to the radar text corpus. ,in, Indicates the first Vector representation of Chinese words, Represents the vector dimension. This represents the number of Chinese words obtained after word segmentation of the radar text corpus. In this embodiment, it is initialized to... =300-dimensional word vectors;

[0039] In this embodiment, taking the radar text corpus “AN / SPS-64 series radar is developed by Raytheon and commonly operates in the I band” as an example: Since the radar text corpus is an unstructured text field, it needs to be segmented before being input into the model. After segmentation, the original sentence is processed into 14 words: “AN / SPS-64”, “series”, “radar”, “by”, “Raytheon”, “company”, “developed”, “,”, “commonly used”, “operating”, “band”, “for”, “I”, “band”.

[0040] Subsequently, the text words are mapped to corresponding 300-dimensional word vectors using the FastText Chinese word vector library. For words not in the FastText library, they are randomly initialized as 300-dimensional word vectors. Finally, each word in the sample corpus generates a 300-dimensional word vector, represented as follows: .

[0041] S2. Obtain the number of entities in the radar text corpus and the start or end position of a single entity;

[0042] S2.1, Word vectors The input is fed into the Transformer model, where the encoder extracts a vector incorporating contextual information. ;

[0043] In this embodiment, word vectors After interaction with the Transformer model, information between words and context in the sample corpus is extracted, which helps to improve the prediction accuracy of entity start and end positions.

[0044] S2.2 Set up two parallel fully connected neural networks with the same input dimension as the word vector dimension, both being 300-dimensional, the hidden layer dimension being 512-dimensional, and the output dimension being the same as the number of word vectors, which is 14-dimensional in this embodiment; wherein, the first fully connected neural network is used to determine the starting position of the entity, and the other fully connected neural network is used to determine the ending position of the entity.

[0045] vector The input is simultaneously fed into two parallel fully connected neural networks, and the output of the fully connected neural networks is a sequence of "0" or "1", where "1" represents the start or end position of an entity, and "0" represents a position that is not the start or end position of an entity.

[0046] In the sequence output by the first fully connected neural network, the position of all "1"s is marked as the starting position of the entity. In the sequence output by the second fully connected neural network, the position of all "1"s is marked as the ending position of the entity. Then the number of entities is counted.

[0047] Finally, following the principle of proximity, the position of the first occurrence of "1" in the first sequence and the position of the first occurrence of "1" in the second sequence are taken as the start or end position of the first entity, and so on.

[0048] In this embodiment, as shown in Figure 2, the sequences output by the two parallel fully connected neural networks are: "1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 0" and "1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1". The first sequence represents the words at positions 1, 5, and 13, namely "AN / SPS-64", "Thor", and "I", corresponding to the starting positions of the entities. The second sequence represents the words at positions 1, 6, and 14, namely "AN / SPS-64", "Company", and "Band", corresponding to the ending positions of the entities. Since the ending position of an entity follows its starting position, matching backwards from the starting position and forwards from the ending position, after processing the outputs of the two fully connected neural networks, three possible entities are obtained: "AN / SPS-64", "Thor Company", and "I Band".

[0049] S3. Calculate the entity vector of each entity;

[0050] Based on the start / end positions of each entity, for an entity composed of a single Chinese word, the entity vector is represented by a vector combining the Chinese word with contextual information. For example, in two binary sequences, the word at position 1 constitutes an entity, and the corresponding entity vector is... For entities composed of multiple Chinese words, the vectors corresponding to each word are summed and then normalized to obtain the entity vector. For example, if the two Chinese words at positions 5 and 6 constitute an entity, the corresponding entity vector needs to be summed and normalized first. and Add them together, then normalize them to obtain the entity vector of the entity;

[0051] S4. Calculate the relationship vector between any two entities;

[0052] Based on the TransE model, for a relation triple <head entity - relation - tail entity>, the relation can be represented by the vector difference between the two entities, i.e., "head entity vector + relation vector = tail entity vector". Two entities are randomly selected; one is denoted as the head entity, and its corresponding entity vector is denoted as... The other is denoted as the tail entity, and the corresponding entity vector is denoted as... ;

[0053] Therefore, based on the above relational triples, we utilize the tail entity vector Subtract head entity vector This yields the relationship vector between entities. ;

[0054] S5. Calculate the relationship vector between any two entities based on attention weights;

[0055]

[0056] in, Indicates the first Attention weights for each Chinese word. This represents the dot product. Indicates the first Vector representation of each Chinese word combined with contextual information Indicates the number of Chinese words between two entities. A relationship vector between two entities based on attention weights;

[0057] In this embodiment, taking the entity pair "AN / SPS-64, Raytheon" as an example, the vectors corresponding to the two entities and the words between them are... ,in, This represents the entity vector corresponding to the entity "AN / SPS-64". This represents the entity vector corresponding to the entity "Raytheon". These represent the vectors corresponding to the words "series", "radar", and "by" between the two entity pairs.

[0058] S6. Extract entity relationships;

[0059] S6.1, Transfer the relation vector splicing relation vectors After that, I got one dimensional relation vector ;

[0060] S6.2, Transfer the relation vector Input to dimension The hidden layer dimension is The output dimension is A fully connected neural network, thus obtaining We calculate a probability value and then use the Softmax function to select the relationship corresponding to the highest probability as the relationship between entity pairs.

[0061] In this embodiment, the input dimension of the fully connected neural network is twice the word vector dimension, which is 600, the hidden layer dimension is 1024, and the output dimension is set according to the needs of the actual radar text corpus. Typically, when the entity relation type is set to... When, the corresponding output dimension is In this embodiment, eight entity labels characterizing different features of radar entities are established based on the radar text corpus: "Radar Model," "Function," "Country," "Developing Unit," "Radar System," "Operating Frequency Band," "Repetition Frequency," and "Pulse Width." Seven relationships describing the relationships between entity pairs are established: "Radar Model - Application - Function," "Radar Model - Belonging to - Country," "Radar Model - Developed by - Developing Unit," "Radar Model - Radar System - Radar System," "Radar Model - Operating Frequency Band - Operating Frequency Band," "Radar Model - Repetition Frequency - Repetition Frequency," and "Radar Model - Pulse Width - Pulse Width." Therefore, the output dimension of the fully connected neural network is 8.

[0062] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for joint extraction of entity relationships for radar, characterized in that, The steps include: (1) Performing Chinese word segmentation on the radar text corpus, and then vectorizing the Chinese words according to the Chinese word vectors in the training library FastText to obtain the vectors corresponding to the radar text corpus. ,in, Indicates the first Vector representation of Chinese words, Represents the vector dimension. (2) The number of Chinese words obtained after the radar text corpus word segmentation operation; (3) The number of entities in the radar text corpus and the start or end position of a single entity; (4) The vector The input is fed into the Transformer model, where the encoder extracts a vector incorporating contextual information. ， Indicates the first (2.2) Set up two parallel fully connected neural networks, where the first fully connected neural network is used to determine the starting position of the entity and the other fully connected neural network is used to determine the ending position of the entity; then, the vector is... Simultaneously input to two parallel fully connected neural networks, and output a sequence of "0" or "1" through the fully connected neural network, where "1" represents the start or end position of the entity, and "0" represents that it is not the start or end position of the entity; in the sequence output by the first fully connected neural network, all positions where "1" is located are marked as the start position of the entity, and in the sequence output by the second fully connected neural network, all positions where "1" is located are marked as the end position of the entity, and then the number of entities is counted; finally, according to the principle of proximity, the first occurrence of "1" in the first sequence and the first occurrence of "1" in the second sequence are taken as the start or end position of the first entity, and so on; (3) Calculate the entity vector of each entity; according to the start or end position of each entity, for an entity composed of a single Chinese word, the vector of the Chinese word combined with the context information is used to represent the entity vector; for an entity composed of multiple Chinese words, the vectors corresponding to the multiple Chinese words are added together and normalized to be used as the entity vector of the entity; (4) Calculate the relationship vector between any two entities; randomly select two entities, one is denoted as the head entity, and the corresponding entity vector is denoted as The other is denoted as the tail entity, and the corresponding entity vector is denoted as... ; Using tail entity vectors Subtract head entity vector This yields the relationship vector between entities. (5) Calculate the relationship vector between any two entities based on attention weights; ;in, Indicates the first Attention weights for each Chinese word. Indicates the first Vector representation of each Chinese word combined with contextual information Indicates the number of Chinese words between two entities. (6) Extract entity relations; (6.1) Transfer the relation vector between two entities based on attention weights; splicing relation vectors After that, I got one dimensional relation vector (6.2) Transfer the relation vector Input to dimension The hidden layer dimension is The output dimension is A fully connected neural network, thus obtaining We calculate a probability value and then use the Softmax function to select the relationship corresponding to the highest probability as the relationship between entity pairs.

2. The radar-oriented entity relationship joint extraction method according to claim 1, characterized in that, The vector dimension It is 300.

Images