Information project book entity and relationship joint extraction method and system

Abstract

The application provides an informationization project book entity and relationship joint extraction method and system, each word in the informationization project book is context-aware coded to obtain a word vector containing context information, a high-order tensor representation of the word vector is projected by using a task field vector to obtain entity task feature representation and relationship task feature representation respectively, and uniqueness between different tasks is ensured; the entity and relationship extraction tasks are modeled as task nodes in a graph to realize efficient interaction between tasks, an adaptive boundary guide factor and a multiplication attention mechanism are used to capture head and tail positions of the entity and the relationship, accurate calculation of relationship types between word pairs is improved, and finally the relationship types between the word pairs are decoded to accurately extract entity and relationship information in the informationization project book.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, and in particular relates to a method and system for jointly extracting entities and relationships from an information project document. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Joint entity and relationship extraction from IT project documents refers to automatically identifying key project-related entities and their relationships from the project document, converting the text data in the project document into structured data. Analyzing only text data can easily lead to one-sided and inaccurate information extraction, increasing the burden of manual analysis and the risk of errors. By extracting entities and relationships and converting the IT project document into structured data, the accuracy and efficiency of information extraction can be improved. It can also help project managers fully grasp the relationship between project content and key elements, thereby optimizing the decision-making process and reducing risks.

[0004] Current research on joint entity and relation extraction from IT project documents mainly employs pipelined and joint extraction methods. Pipeline methods treat joint entity and relation extraction as two independent tasks, lacking coordination and mutual optimization within the same model framework, making it difficult to simultaneously identify and model entities and relations. Joint extraction methods, on the other hand, output all entities and their relations at once through a single model, effectively utilizing the feature associations between the two tasks. However, existing research has failed to fully explore the intrinsic connections between entities and relations, making it impossible to achieve more accurate and comprehensive extraction from complex structured documents like IT project documents.

[0005] In addition, the IT project proposal contains a large number of technical terms and domain-specific expressions, requiring the model not only to understand the entities and relationships in the text, but also to handle industry-specific contexts. Relationships are often based on the contextual definition of entities and are implicit in some cases. Therefore, it is necessary to fully explore the intrinsic connections between entities and relationships to improve the accuracy of extraction. Summary of the Invention

[0006] To overcome the shortcomings of the prior art, the present invention provides a method and system for jointly extracting entities and relationships from an information project document, which can accurately extract entity and relationship information from the information project document.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for jointly extracting entities and relationships from an information-based project proposal, comprising: Each word in the IT project document is context-aware encoded to obtain a word vector containing context information; The word vectors are mapped to higher-order tensor representations, and the higher-order tensor representations of the word vectors are projected onto the task field vectors to obtain entity task feature representations and relation task feature representations, respectively. All representations in the entity task feature representation and relation task feature representation are regarded as the entity task node set and relation task node set, respectively. A graph structure is constructed, and the nodes interact and learn through graph convolution to obtain the entity task node set and relation task node set of task adaptability. For byte pairs in the entity task node set and relation task node set of task adaptability, the relationship type score between the corresponding word pairs is calculated using the adaptive boundary guidance factor and multiplicative attention mechanism, respectively. Based on the calculated relationship type scores between word pairs, the entities and relationships in the information technology project document are decoded.

[0008] Secondly, the present invention provides an information-based project document entity and relation joint extraction system, comprising: An information-based project proposal entity and relation extraction system, characterized in that it includes: The extraction module is configured to perform context-aware encoding on each character in the information technology project document to obtain a character vector containing context information. The projection module is configured to: map the word vectors to higher-order tensor representations, and project the higher-order tensor representations of the word vectors using task field vectors to obtain entity task feature representations and relation task feature representations, respectively. The graph structure module is configured to treat all representations in the entity task feature representation and the relation task feature representation as sets of entity task nodes and sets of relation task nodes, construct a graph structure, and enable interaction learning between nodes through graph convolution to obtain the set of entity task nodes and set of relation task nodes for task adaptability. The computation module is configured to: calculate the relation type score between corresponding word pairs in the set of entity task nodes and the set of relation task nodes that are adaptive to the task, using an adaptive boundary guiding factor and a multiplicative attention mechanism. The decoding module is configured to decode the entities and relationships in the information project document based on the calculated relationship type scores between word pairs.

[0009] Thirdly, the present invention provides an electronic device including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the method described in the first aspect.

[0010] Fourthly, the present invention provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in the first aspect.

[0011] Fifthly, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the method described in the first aspect.

[0012] The above one or more technical solutions have the following beneficial effects: This invention performs context-aware encoding on each character in an information technology project document to obtain character vectors containing contextual information. It then projects the higher-order tensor representation of the character vectors using task field vectors to obtain entity task feature representations and relation task feature representations, ensuring the uniqueness between different tasks. By modeling entity and relation extraction tasks as task nodes in a graph, it achieves efficient interaction between tasks. Adaptive boundary guidance factors and multiplicative attention mechanisms are used to capture the head and tail positions of entities and relations, improving the accurate calculation of relation types between character pairs. Finally, it decodes the relation types between character pairs to accurately extract entity and relation information from the information technology project document.

[0013] This invention first uses a linear mapping function to map word vectors into higher-order tensor representations. Then, it introduces a task field vector and projects each word's higher-order tensor through the task field vector to dynamically segment it into entity task features and relation task features. This approach allows each word vector to naturally expand in both entity and relation extraction task directions, adapting to the needs of different tasks and thus better capturing task-specific contextual information.

[0014] This invention utilizes cosine similarity to measure the relevance between entity task nodes and relation task nodes, while introducing task-aware weighting coefficients to achieve effective propagation and fusion of information between different tasks. By obtaining the connection weights between nodes in the same task through normalized cosine similarity, interactive learning between nodes is achieved through graph convolution, ultimately generating more discriminative and task-adaptive feature representations. Through task-aware information flow propagation, it can flexibly combine the needs of multiple tasks, improving the performance of entity and relation extraction tasks, especially in complex multi-task learning scenarios, providing better feature representation capabilities.

[0015] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart illustrating the steps of a method for jointly extracting entities and relationships from an information project document according to an embodiment of the present invention. Figure 2 In this embodiment of the invention, a word vector graph is generated using the BERT model and a bidirectional long short-term memory network. Figure 3 This is a diagram of the task field information expansion module in an embodiment of the present invention; Figure 4 This is a diagram of the information flow propagation module for task awareness in an embodiment of the present invention; Figure 5 This is a diagram showing the results of the joint extraction of entities and relationships from an information technology project document in an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0020] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0021] Example 1 This embodiment discloses a method for jointly extracting entities and relationships from an information technology project proposal, including: Each word in the IT project document is context-aware encoded to obtain a word vector containing context information; By mapping word vectors to higher-order tensor representations and projecting the higher-order tensor representations of word vectors onto task field vectors, we obtain entity task feature representations and relation task feature representations, respectively. All representations in the entity task feature representation and relation task feature representation are regarded as the entity task node set and relation task node set, respectively. A graph structure is constructed, and the nodes interact and learn through graph convolution to obtain the entity task node set and relation task node set of task adaptability. For byte pairs in the entity task node set and relation task node set of task adaptability, the relationship type score between the corresponding word pairs is calculated using the adaptive boundary guidance factor and multiplicative attention mechanism, respectively. Based on the calculated relationship type scores between word pairs, the entities and relationships in the information technology project document are decoded.

[0022] This embodiment proposes a joint entity and relation extraction method for information technology project documents. This method employs a BERT model and a bidirectional long short-term memory network to perform context-aware encoding on the information technology project document text, generating word vectors containing contextual information. Task field vectors are introduced, and geometric transformations are performed on the word vectors using different task field direction vectors to obtain task-specific word vectors. Subsequently, the entity and relation extraction tasks are modeled as task nodes in a graph, and information is transmitted using the graph structure to ensure effective interaction between tasks. An adaptive boundary guidance factor is introduced, and a multiplicative attention mechanism is used to accurately capture the head and tail positions of entities and relations, calculate the relation types between word pairs, and use a joint decoding module to decode the relation types between word pairs in the entity and relation tasks. This accurately extracts entities and relations from the information technology project document, optimizes the accuracy and efficiency of the extraction process, and improves the overall effect of information extraction.

[0023] The following is combined Figure 1 This embodiment provides a detailed description of a method for jointly extracting entities and relationships from an information technology project proposal: Step 1: Perform context-aware encoding on each character in the IT project document to obtain a character vector containing context information.

[0024] Specifically, the BERT model and bidirectional long short-term memory network are used to perform context-aware encoding on each character in the information project proposal text, generating character vectors containing context information.

[0025] The input information technology project proposal contains A sentence with one character , ,in, For the first sentence Each word, if Then fill in the missing characters with 0s. Then delete the redundant words, keeping only the words in the sentence. Each character, after processing, the sentence , the sentence The input is fed into the pre-trained language model BERT to generate character vectors containing contextual semantic representations. Each character vector is further processed using a bidirectional long short-term memory network to enhance its global contextual information, such as... Figure 2 As shown, the specific mathematical model is as follows: ; ; In the formula, This is a sequence of character vectors processed by the pre-trained language model BERT. ,in The first sentence in the IT project proposal The character vectors obtained after processing each character by BERT Output the dimension of the word vectors for BERT. For the pre-trained language model BERT, This is a sequence of character vectors processed by a bidirectional long short-term memory network. ,in For the first The word vectors obtained after processing each word vector through a bidirectional long short-term memory network. It is a bidirectional long short-term memory network.

[0026] BERT utilizes the pre-training mechanism of deep learning to capture contextual information of words, so that each word vector depends not only on the word itself, but also on the context of its surrounding environment. The bidirectional long short-term memory network further processes these word vectors, thereby enhancing the contextual understanding of each word in the whole sentence. The combination of BERT and bidirectional long short-term memory network can effectively improve the representational power of word vectors, so that semantic representation can not only consider the meaning of a single word, but also reflect its role and relationship in the context. This helps to process complex and specialized information project document text data and provides more accurate feature representations for subsequent joint entity and relation extraction tasks.

[0027] In this embodiment, text data from the information technology project proposal is collected, and the entities and relationships in the text are labeled to construct a joint entity and relationship extraction dataset from the information technology project proposal.

[0028] Specifically, text data from the IT project proposal was collected, and the text content was segmented according to periods to obtain 1033 sentences. Entities in the sentences were labeled, including project name, project goals, tasks, personnel, time, budget, resources, and system modules. Relationships between entities were also labeled, including goal achievement relationships, responsibility allocation relationships, task timeline relationships, resource requirement relationships, funding allocation relationships, potential risk relationships, project and personnel relationships, and module dependency relationships. A joint entity and relationship extraction dataset from the IT project proposal was constructed, and the dataset was divided into a training set and a validation set in an 8:2 ratio for training and validation of the entire model.

[0029] S2: Map the word vectors to higher-order tensor representations, and project the higher-order tensor representations of the word vectors onto the task field vectors to obtain entity task feature representations and relation task feature representations, respectively.

[0030] like Figure 3 As shown, a task field information expansion module is constructed, a task field vector is introduced, and two types of extraction task direction vectors, namely entity and relation, are constructed in the task field. Geometric transformations are performed on the word vectors along different task field direction vectors to obtain task-specific word vectors.

[0031] Word vectors are treated as multidimensional tensor entities in the task field. First, a linear mapping function is used to map the word vectors into higher-order tensor representations. Then, a task field vector is introduced, and each word's higher-order tensor is hyperbolic projected through the task field vector to dynamically segment it into entity task features and relation task features. This approach allows each word vector to naturally expand in both entity and relation extraction directions, adapting to the needs of different tasks and thus better capturing task-specific contextual information, including: S21. Using a linear mapping function, the word vector is mapped to a higher-order tensor representation. The specific mathematical model is as follows: ; In the formula, It is a sequence of higher-order tensors mapped from a sequence of character vectors. ,in For the first The higher-order tensor representation of a character For a learnable parameter matrix, For learnable bias terms, It represents the higher-order dimension of the tensor; S22. Introduce task field vectors and project them onto the higher-order tensor representations of word vectors to obtain task-specific feature representations for entities and relations, respectively. The specific mathematical model is as follows: ; ; In the formula, The sequence represents the features of the entity task. ,in The first sentence in the IT project proposal Entity task feature representation of a single character, For the entity task's task field vector, For a 10-dimensional unit sphere in hyperbolic space, The tanh activation function is used. The sequence is used to represent the features of relational tasks. ,in The first sentence in the IT project proposal Character relational task feature representation This is the task field vector for the relational task. and The two task field vectors are first obtained through random initialization, then optimized as part of the model parameters, and adjusted through backpropagation.

[0032] S3: Treat all representations in the entity task feature representation and relation task feature representation as sets of entity task nodes and relation task nodes, construct a graph structure, and enable interactive learning between nodes through graph convolution to obtain the entity task node set and relation task node set of task adaptability.

[0033] like Figure 4 As shown, a task-aware information flow propagation module is proposed, which models entity and relation extraction tasks as task nodes in a graph, and transmits information between task nodes and task-specific word vectors through the graph structure.

[0034] Specifically, building The layer task-aware information flow propagation module, in this embodiment, is used as... Taking an example, the entity task node representation is randomly initialized. Relationship task node representation Furthermore, all representations in the entity and relation task feature representation sequence are treated as byte points to construct an entity task node set. , and the set of relational task nodes , In the The layer uses fully connected connections to link all nodes of the same task. Entities and relationships between tasks are connected through task nodes, forming a graph structure. This module includes the following steps: S31. Calculate the first using cosine similarity. The connection weights between entities and relational task nodes in the layer are represented by the following mathematical model: ; In the formula, For the first The connection weights between layer entities and relational task nodes. For transpose operation, , The first Layered entity and relation task node representation, for ,have , , The Euclidean norm is used to measure the size of a vector.

[0035] Cosine similarity is used to measure the correlation between nodes in a set of entity task nodes. After normalization, the connection weights between nodes are obtained. The specific mathematical model is as follows: ; ; In the formula, For the first Layer nodes and The correlation, For the first Layer nodes and Connection weights between them , , The first Node representation in the layer entity task node set. For the first The set of entity task nodes in the layer, for ,have , , , ; Similarly, calculate the first... Connection weights between nodes in a hierarchical task node set ,in , The first The node representation in the layer's relational task node set. For the first The set of entity task nodes in the layer, for ,have , , ; S32. Introduce task-aware weighting coefficients to dynamically update the entity and relation task node representations based on connection weights. The specific mathematical model is as follows: ; ; ; In the formula, For the first The task-aware weighting coefficient of the layer, , For the updated version Layered entity and relation task node representation, To control the hyperparameter of connection weight adjustment based on task node correlation, it was set to 0.001 during implementation; S33. Perform interactive learning on the byte points in the entity and relation task node sets respectively to update the byte point representations. The specific mathematical model is as follows: ; ; In the formula, , Each is the updated version of the first. Byte point characteristics in the set of layer entity and relation task nodes. For the first Layer nodes With the Connection weights between layer entity task nodes For the first Layer nodes With the Connection weights between task nodes in a hierarchical relationship; S34. Repeat S31-S33 until the first... Layer entity task node set and the set of relational task nodes , , ,in For the first Layer entity task node set Each byte is represented by a dot. For the first Layered relationship task node set Each byte is represented by a dot.

[0036] This embodiment introduces entity and relation task node representations to capture the unique task features of each task. Cosine similarity is used to measure the correlation between entity and relation task nodes. At the same time, task-aware weighting coefficients are introduced to achieve effective propagation and fusion of information between different tasks. The connection weights between nodes in the same task are obtained through normalized cosine similarity, and interactive learning between nodes is achieved through graph convolution. Finally, more discriminative and task-adaptive feature representations are generated. Through task-aware information flow propagation, the needs of multiple tasks can be flexibly combined, improving the performance of entity and relation extraction tasks, especially in complex multi-task learning scenarios, providing better feature representation capabilities.

[0037] S4: For byte pairs in the entity task node set and relation task node set of task adaptability, calculate the relation type score between the corresponding word pairs using the adaptive boundary guiding factor and multiplicative attention mechanism.

[0038] This embodiment constructs an adaptive global pointer module, introduces an adaptive boundary guiding factor, and combines a multiplicative attention mechanism to capture the head and tail positions of entities and relations respectively, and calculates the relation type between word pairs.

[0039] Specifically, it includes: S41. Using linear mapping, the byte-point representation in the entity task node set is mapped to the entity header and tail representation. The specific mathematical model is as follows: ; ; In the formula, For sentences The Middle The character as the first The representation of the entity head, For sentences The Middle The character as the first The representation of the tail of the species. The dimension after linear mapping, during implementation , , These are the learnable weight matrices used to generate the head and tail representations, respectively. , These are the bias terms for generating the head and tail representations, respectively; S42. Introducing an adaptive boundary guiding factor Calculate the first word in the sentence. The word and the first The character as the first The mathematical model for scoring the head and tail of each entity type is as follows: ; In the formula, For the first The word and the first The character as the first Adaptive boundary guiding factors for the head and tail boundaries of various entity types. The weight matrix is ​​a learnable matrix. The tanh activation function is used. The weight matrix is ​​a learnable matrix. For bias terms, For splicing operations, For sentences The Middle The character as the first Representation of the tail of the species entity; S43. Combining adaptive boundary guidance factors and multiplicative attention mechanisms to calculate segment scores, relative position encoding is introduced to better capture the boundary information of the span, calculating the score from the first segment in the sentence. One to the first The segment consisting of 10 characters belongs to the 10th rank. The scoring for each entity type, specifically the mathematical model, is as follows: ; In the formula, For the sentence from the first One to the first The segment consisting of 10 characters belongs to the 10th rank. Scores for each entity type Encoding for relative positions; S44. Similarly, apply steps S41 to S43 to the byte pairs in the relational task node set, and calculate the word pair belonging to the first... Score of each type of relationship .

[0040] This embodiment maps byte point representations to head and tail representations through linear mapping, ensuring that the model can capture deep relationships between words. An adaptive memory boundary factor is introduced to further calculate the relevance scores between word pairs, optimizing the model's understanding of fine-grained relationships between word pairs. This allows the model to dynamically adjust the scores between word pairs based on context to identify entity boundaries. The adaptive memory boundary factor, combined with multiplicative attention and relative position encoding, calculates the score of segments composed of multiple byte points as entities, enabling the model to flexibly consider the relative positions of byte points, thereby improving its ability to model long-distance dependencies. By combining the adaptive memory boundary factor, multiplicative attention, and relative position encoding, the ability to capture relationships between word pairs is significantly improved.

[0041] S5. Decode the relationship types between word pairs in the entity and relationship task to obtain the entities and relationships in the information project document text.

[0042] This embodiment constructs a joint entity and relation decoding module, which jointly decodes the entities and relations in the information technology project document based on the entity type score and relation type score between word pairs. Specifically: S51, traverse the entity type scores between word pairs, if Then the sentence From the middle One to the first The characters constitute the first A type of entity is defined, and that entity is added to the entity set. This is the entity threshold, set to 1 during implementation; S52. Iterate through the scores of all word pair relationship types. Then, find the entity set starting with the first [entity name]. The character is the head entity and the first The two entities, each with a head character, constitute... Add the relationship to the relationship set. This is the relationship threshold, which is set to 1 during implementation.

[0043] This embodiment introduces a threshold to flexibly control the accuracy of entities and relationships during the decoding process, avoiding the problem of over- or under-decoding. By traversing and decoding the entity type and relationship type scores between word pairs, the model can effectively extract meaningful entities and the relationships between them from the text.

[0044] This embodiment's method is implemented using the Python 3.8 programming language and the PyTorch framework. To improve training efficiency, the method runs in a CUDA 11 environment and uses an NVIDIA 3090 GPU for computation. During training, the initial learning rate is set to... Batch size is ,use The optimizer performs gradient updates, and the loss function chosen is the cross-entropy loss function.

[0045] The partial results obtained from the joint extraction of entities and relations in this embodiment are as follows: Figure 5 As shown, in Entity, Text represents the list of words contained in the entity, Index represents the list of indices of the words contained in the entity in the sentence, EntityType represents the entity type, and in Relation, Head represents the first entity in the relation, Tail represents the second entity in the relation, and RelationType represents the relation type. Figure 5 It can be seen that this method is able to effectively extract entities and their relationships from the information technology project proposal.

[0046] Example 2 The purpose of this embodiment is to provide an information-based project proposal entity and relation joint extraction system, including: The extraction module is configured to perform context-aware encoding on each character in the information technology project document to obtain a character vector containing context information. The projection module is configured to: map word vectors to higher-order tensor representations, and project the higher-order tensor representations of word vectors onto task field vectors to obtain entity task feature representations and relation task feature representations, respectively. The graph structure module is configured to treat all representations in the entity task feature representation and the relation task feature representation as sets of entity task nodes and sets of relation task nodes, construct a graph structure, and enable interaction learning between nodes through graph convolution to obtain the set of entity task nodes and set of relation task nodes for task adaptability. The computation module is configured to: calculate the relation type score between corresponding word pairs in the set of entity task nodes and the set of relation task nodes that are adaptive to the task, using an adaptive boundary guiding factor and a multiplicative attention mechanism. The decoding module is configured to decode the entities and relationships in the information project document based on the calculated relationship type scores between word pairs.

[0047] In further embodiments, the following is also provided: An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When executed by the processor, the computer instructions perform the method described in Embodiment 1. For brevity, further details are omitted here.

[0048] It should be understood that in this embodiment, the processor can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc.

[0049] Memory may include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory may also include non-volatile random access memory. For example, memory may also store information about the device type.

[0050] A computer-readable storage medium for storing computer instructions, which, when executed by a processor, perform the method described in Embodiment 1.

[0051] The method in Embodiment 1 can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor. The software modules can reside in readily available storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, a detailed description is not provided here.

[0052] A computer program product includes a computer program that, when executed by a processor, implements the method described in Embodiment 1.

[0053] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer-executable instructions, such as instructions included in program modules, which execute in a device on a target real or virtual processor to perform the processes / methods described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc., that perform specific tasks or implement specific abstract data types. In various embodiments, the functionality of program modules can be combined or divided among program modules as needed. The machine-executable instructions for the program modules can execute within a local or distributed device. In a distributed device, the program modules can reside in both local and remote storage media.

[0054] The computer program code used to implement the methods of the present invention may be written in one or more programming languages. This computer program code may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the computer or other programmable data processing device, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer, or entirely on a remote computer or server.

[0055] In the context of this invention, computer program code or related data may be carried by any suitable carrier to enable a device, apparatus, or processor to perform the various processes and operations described above. Examples of carriers include signals, computer-readable media, and the like. Examples of signals may include electrical, optical, radio, sound, or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0056] Those skilled in the art will recognize that the units and algorithm steps described in conjunction with the embodiments herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0057] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for jointly extracting entities and relationships from an information technology project proposal, characterized in that, include: Each word in the IT project document is context-aware encoded to obtain a word vector containing context information; The word vectors are mapped to higher-order tensor representations, and the higher-order tensor representations of the word vectors are projected onto the task field vectors to obtain entity task feature representations and relation task feature representations, respectively. Specifically: The word vectors are mapped to higher-order tensor representations using a linear mapping function; By projecting the higher-order tensor representation of word vectors onto the task field vector, and unfolding different aspects of the higher-order tensor along different geometric flow directions, the entity task feature representation and relation task feature representation are obtained respectively. All representations in the entity task feature representation and the relation task feature representation are treated as sets of entity task nodes and sets of relation task nodes, respectively. A graph structure is constructed, and the interaction and learning between nodes are achieved through graph convolution to obtain the entity task node set and relation task node set for task adaptability; specifically: Calculate the first using cosine similarity respectively The connection weights between entity task nodes and relational task nodes in the layer, the connection weights between nodes in the entity task node set, and the connection weights between nodes in the relational task node set; According to the In the layer, the connection weights between entity task nodes and relational task nodes introduce task-aware weighting coefficients. The layer is updated based on these task-aware weighting coefficients and the connection weights between nodes in the entity task node set. Layer entity task node representation; update the first layer based on the task-aware weighting coefficient and the connection weights between nodes in the relational task node set. Hierarchical relationship task node representation; The byte points in the set of entity task nodes and the set of relation task nodes are interactively learned and updated to update the byte point representation until the set of entity task nodes and the set of relation task nodes with task adaptability are obtained. For byte pairs in the entity task node set and relation task node set of task adaptability, the relationship type score between the corresponding word pairs is calculated using the adaptive boundary guidance factor and multiplicative attention mechanism, respectively. Based on the calculated relationship type scores between word pairs, the entities and relationships in the information project document are decoded.

2. The method for jointly extracting entities and relationships from an information technology project proposal as described in claim 1, characterized in that, Each character in the IT project document is coded using context-aware encoding to obtain a character vector containing contextual information, specifically: The pre-trained BERT model is used to process sentences in the information technology project proposal to generate word vectors containing contextual semantic representations. Each character vector is processed using a bidirectional long short-term memory network to enhance its global contextual information.

3. The method for jointly extracting entities and relationships from an information technology project proposal as described in claim 1, characterized in that, The process of interactively learning the byte points in the entity task node set and the relation task node set, updating the byte point representations, until the task-adaptive entity task node set and relation task node set are obtained, specifically involves: ； ； in, 、 Each is the updated version of the first. Byte point characteristics in the set of layer entity task nodes and the set of relational task nodes. For the first Layer nodes With the Connection weights between layer entity task nodes For the first Layer nodes With the The connection weights between task nodes in a layered relationship. For the first Layered task node set For the first Layer entity task node set, For the first Layer Entity Task Node and Connection weights between them For the first Hierarchical relationship task nodes and Connection weights.

4. The method for jointly extracting entities and relationships from an information technology project proposal as described in claim 1, characterized in that, For byte pairs in the entity task node set and relation task node set of task adaptability, the relation type score between the corresponding word pairs is calculated using the adaptive boundary guidance factor and multiplicative attention mechanism, respectively: Map the byte-point representations in the set of entity task nodes or relation task nodes for task adaptability to entity header and tail representations, and calculate the first byte representation in the sentence. The word and the first The character as the first Type of entity or the first The head and tail scores of each relation type are used to obtain an adaptive boundary guiding factor; For the The word and the first Each character is encoded using relative position encoding, and combined with an adaptive boundary guiding factor, the result is obtained from the first character. One to the first The segment consisting of 10 characters belongs to the 10th rank. Type of entity or the first Scores for each type of relationship.

5. The method for jointly extracting entities and relationships from an information technology project proposal as described in claim 4, characterized in that, Based on the calculated relationship type scores between word pairs, the entities and relationships in the information technology project document are decoded, specifically: Iterate through the entity type scores between word pairs. If the score is greater than the entity threshold, then the sentence... From the middle One to the first The characters constitute the first The first entity will be the first Add the type of entity to the entity set; Iterate through the scores of all word pairs' relation types. If a score is greater than a relation threshold, find the word pair with the specified relation type from the entity set. The character is the head entity and the first... The character is the head of an entity, which is composed of two entities. Add the relation to the relation set.

6. A system for jointly extracting entities and relationships from an information-based project proposal, characterized in that, include: The extraction module is configured to perform context-aware encoding on each character in the information technology project document to obtain a character vector containing context information. The projection module is configured to: map the word vectors to higher-order tensor representations, and project the higher-order tensor representations of the word vectors using task field vectors to obtain entity task feature representations and relation task feature representations, respectively. Specifically: The word vectors are mapped to higher-order tensor representations using a linear mapping function; By projecting the higher-order tensor representation of word vectors onto the task field vector, and unfolding different aspects of the higher-order tensor along different geometric flow directions, the entity task feature representation and relation task feature representation are obtained respectively. The graph structure module is configured to: treat all representations in the entity task feature representation and the relation task feature representation as sets of entity task nodes and sets of relation task nodes, construct a graph structure, and enable interaction learning between nodes through graph convolution to obtain the entity task node set and relation task node set for task adaptability; specifically: Calculate the first using cosine similarity respectively The connection weights between entity task nodes and relational task nodes in the layer, the connection weights between nodes in the entity task node set, and the connection weights between nodes in the relational task node set; According to the In the layer, the connection weights between entity task nodes and relational task nodes introduce task-aware weighting coefficients. The layer is updated based on these task-aware weighting coefficients and the connection weights between nodes in the entity task node set. Layer entity task node representation; update the first layer based on the task-aware weighting coefficient and the connection weights between nodes in the relational task node set. Hierarchical relationship task node representation; The byte points in the set of entity task nodes and the set of relation task nodes are interactively learned and updated to update the byte point representation until the set of entity task nodes and the set of relation task nodes with task adaptability are obtained. The computation module is configured to: calculate the relation type score between corresponding word pairs for byte pairs in the entity task node set and the relation task node set, respectively, using an adaptive boundary guiding factor and a multiplicative attention mechanism. The decoding module is configured to decode the entities and relationships in the information project document based on the calculated relationship type scores between word pairs.

7. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

Images