A method and system for constructing a relationship graph based on conversation content

Abstract

This application relates to a method for constructing a relationship graph based on conversation content. It addresses common issues in spoken text such as homophones and multiple referents through multi-dimensional feature fusion-based entity alignment, ensuring the accuracy and uniqueness of graph nodes. Secondly, it utilizes a large language model with fine-tuning to extract deep relationships, upgrading the traditional single person-relationship-person triples in the graph into dynamic relationships with rich attributes such as time, location, event, and evidence source. Furthermore, it forcibly binds relationships to original evidence, resolving issues of information traceability and interpretability. This enables the construction of a rich, traceable deep relationship graph, improving the efficiency and depth of conversation content analysis and providing reliable technical support for the construction of evidence chains.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method, system, computer device, and computer-readable storage medium for constructing a relational graph based on conversation content. Background Technology

[0002] In certain specialized professional settings, conversations with relevant individuals are central to obtaining crucial information. These conversations often feature long time spans, high information density, and complex interpersonal relationships. Currently, relationship mapping technology is being applied to sift through these massive amounts of unstructured conversation texts, identifying relationships and reconstructing the sequence of events.

[0003] Existing relationship graph technologies primarily focus on extracting and displaying basic interpersonal relationship triples (e.g., "Zhang Moumou - Colleague - Li Moumou"). However, this traditional graph technology has significant limitations in practice: First, the relationship information is too simplistic and lacks depth. Existing relationship graphs can only identify the existence and basic type of relationships, but cannot depict key attributes such as the time, place, specific event background, and frequency of interaction. As a result, the graphs become highly simplified relationship skeletons rather than rich and detailed networks. Summary of the Invention

[0004] This application provides a method, system, computer device, and computer-readable storage medium for constructing a relationship graph based on conversation content, in order to at least solve the problem that relationship graph construction methods in related technologies cannot meet the backtracking requirements in specific scenarios.

[0005] In a first aspect, embodiments of this application provide a method for constructing a relationship graph based on conversation content, the method comprising: Acquire and process conversation texts from specific professional scenarios to generate clean texts to be analyzed. The processing includes preprocessing and standardizing the conversation texts. Based on the clean text, entity recognition is performed to determine preliminary entities. Through a multi-dimensional feature fusion model, a comprehensive similarity score between the preliminary entities is calculated. Based on the comprehensive similarity score, the preliminary entities in the clean text are aligned to determine globally unique entity objects. Using a large language model finely tuned by instructions, attributed relationships between entity objects are extracted from clean text that has undergone preliminary entity alignment, and the original evidence fragments of the attributed relationships in the clean text are associated with them. The attributed relationships include a preset relationship type and attribute information associated with the attributed relationships. Based on the attributed relationships between the entity objects and the associated original evidence fragments, a conversation content relationship graph is constructed.

[0006] In some embodiments, preprocessing the conversation text includes: Using a language model or pre-defined error correction rules, the homophonic errors of proper nouns in the conversation text are initially corrected, and text normalization processing is performed. The text normalization processing includes: removal of colloquial words, merging of repetitive expressions, and regularization of sentence logic. Furthermore, the normalized conversation text is segmented into semantic units centered on specific events and with limited context, according to dialogue rounds, and / or time stamps, and / or topic clustering, to obtain the clean text.

[0007] In some embodiments, calculating the comprehensive similarity score between the preliminary entities using a multi-dimensional feature fusion model includes: Obtain the multidimensional features of the preliminary entity, wherein the multidimensional features include: literal features, contextual semantic features, and entity attribute features; By introducing a negative sample learning mechanism into the machine learning model, the comprehensive similarity score is obtained by performing a weighted calculation based on the multidimensional features and preset weights. The literal features are obtained by calculating the literal edit distance of the entity name and / or the similarity of the pinyin sequence, and the contextual semantic features are obtained by converting the context fragments around the initial entity into vector representations and calculating the cosine similarity between the vector representations.

[0008] In some embodiments, the entity attribute features are obtained by extracting attribute information associated with the entity object from the clean text of the initial entity alignment and performing similarity comparison on the attribute information, wherein the attribute information includes job information, affiliated unit information, place of origin information or known associated person information.

[0009] In some embodiments, attributed relations are extracted using a large language model that has been fine-tuned by instructions, including: By using instruction fine-tuning technology, a basic large language model is fine-tuned and trained based on a predefined structured instruction dataset that meets the needs of specific professional scenarios, resulting in a fine-tuned model that can execute relation and attribute extraction instructions. The fine-tuning model is input with clean text that has undergone preliminary entity alignment, and guided by preset prompts, it performs deep extraction based on thought chains and outputs the attributed relationships. The prompts are used to constrain the output of the fine-tuning model to include original text evidence fragments for verifying the existence of the attributed relationships.

[0010] In some embodiments, the attribute information specifically includes: the time when the relationship occurred, the location where the relationship occurred, the items or amounts involved in the relationship, and a textual description of the event process; By annotating the key citations on which the attributed relationships are based, as well as the position information of the key citations in the original conversation text, in the edge attributes of the relationship graph of the conversation content, the original evidence fragments can be associated.

[0011] In some embodiments, the method further includes: Obtain the updated conversation text, and extract the updated attributed relationships and their associated entity objects from the updated conversation text; Entity alignment is determined for the associated entity objects. If the entity object already exists in the current conversation content relationship graph, the updated attribute relationship is added as a new edge between the corresponding entity objects. If the updated attributed relationship is judged and it belongs to the same objective event as the target edge that already exists in the conversation content relationship graph, then the original evidence fragment associated with the updated attributed relationship is added as a new evidence source to the attribute list of the target edge.

[0012] Secondly, embodiments of this application provide a relationship graph construction system based on conversation content, the system comprising: The acquisition module is used to acquire and process conversation texts from specific professional scenarios, and generate clean texts to be analyzed. The processing includes preprocessing and standardizing the conversation texts. The alignment module is used to identify preliminary entities based on the clean text, calculate the comprehensive similarity score between the preliminary entities through a multi-dimensional feature fusion model, and align the preliminary entities in the clean text according to the comprehensive similarity score to determine globally unique entity objects. The extraction module is used to extract attributed relationships between entity objects from clean text that has undergone preliminary entity alignment using a large language model that has been fine-tuned by instructions, and to associate the attributed relationships with the original evidence fragments in the clean text. The attributed relationships include a preset relationship type and attribute information associated with the attributed relationship. A construction module is used to construct a conversation content relationship graph based on the entity objects and the attributed relationships.

[0013] Thirdly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method described in the first aspect above.

[0014] Fourthly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect above.

[0015] Compared to related technologies, the relationship graph construction method based on conversation content provided in this application solves common problems in spoken text such as homophones and multiple referents through multi-dimensional feature fusion entity alignment, ensuring the accuracy and uniqueness of graph nodes. Secondly, it uses a large language model with fine-tuning instructions for deep relationship extraction, upgrading the traditional single "person-relationship-person" triple in the graph into dynamic relationships with rich attributes such as time, place, event, and evidence source, and forcibly binding relationships to original evidence, solving the problems of information traceability and interpretability, realizing the construction of a rich attribute and traceable deep relationship graph, improving the efficiency and depth of conversation content analysis, and providing reliable technical support for the construction of evidence chains. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a relationship graph construction method based on conversation content according to an embodiment of this application; Figure 2 This is a structural block diagram of a relationship graph construction system based on conversation content, according to an embodiment of this application; Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application is described and illustrated below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.

[0018] Obviously, the accompanying drawings described below are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar scenarios based on these drawings without any inventive effort. Furthermore, it is understood that although the efforts made in this development process may be complex and lengthy, for those skilled in the art related to the content disclosed in this application, any changes to design, manufacturing, or production based on the technical content disclosed in this application are merely conventional technical means and should not be construed as insufficient disclosure of the content of this application.

[0019] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.

[0020] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms “a,” “an,” “an,” “the,” and similar words used in this application do not indicate quantity limitation and may indicate singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or units, but may also include steps or units not listed, or may include other steps or units inherent to these processes, methods, products, or devices. The terms “connected,” “linked,” “coupled,” and similar words used in this application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. The character " / " generally indicates that the preceding and following objects are in an "or" relationship. The terms "first," "second," and "third" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects.

[0021] Embodiments of the present application provide a method, system, computer device and computer-readable storage medium for constructing a knowledge graph based on conversation content. The purpose is to construct a deep knowledge graph with rich attributes, traceable evidence and precisely aligned entities from unstructured conversation texts through an automated processing workflow, so as to meet the requirements of digging deep into hidden associations and restoring complex interaction scenarios in specific professional scenarios.

[0022] Figure 1 is a flowchart of a method for constructing a knowledge graph based on conversation content according to an embodiment of the present application, as shown in Figure 1 , the process includes the following steps: S101: Acquire and process conversation texts of a specific professional scenario, and generate clean texts to be analyzed.

[0023] In this step, one or more conversation record texts are used as input. These original texts come directly from manual recording, or are converted from audio through Automatic Speech Recognition (ASR) technology, so they may contain errors, colloquial expressions and unstructured features. To lay a foundation for subsequent precise entity and relation extraction, preprocessing and standardization of the original texts are required.

[0024] In a specific embodiment, the execution process of this step can be decomposed into: S1: Perform text error correction and normalization. For the text transcribed from original speech, a language model or preset error correction rules are used for correction. For example, for a speech transcribed text "Wang Wei went there yesterday... uh, that is to say...", the homophonic error "Wang Wei" can be corrected to "Wang Wei" according to a preset proper noun dictionary; at the same time, the colloquial expression "uh, that is to say..." is normalized to "that is...".

[0025] This step S1 significantly improves the quality and regularity of the text by correcting text errors and removing colloquial interferences, and provides a reliable data foundation for subsequent precise semantic analysis and information extraction.

[0026] S2: Perform chapter structured segmentation. The long conversation text after normalization processing is segmented according to dialogue turns or by using a topic clustering algorithm. For example, a conversation record with a length of tens of thousands of words can be segmented into multiple independent semantic units, such as "discussion on Project A", "recollection of the meeting with Person B", etc.

[0027] This step S2 decomposes the lengthy text into context fragments focused on specific events, effectively reduces the complexity of subsequent model processing, and enables entity and relation extraction to be carried out in a more focused context.

[0028] In this embodiment, step S101 cleans, corrects, and structures the original, noisy conversation text to solve the technical problems of low source data quality and inconsistent format, and outputs high-quality, machine-readable clean text for subsequent core modules such as entity alignment and relation extraction.

[0029] S102 performs entity recognition based on clean text and aligns entities using a multi-dimensional feature fusion model to determine globally unique entity objects.

[0030] This step aims to address the core issue of unclear entity references in conversational texts, ensuring that each node in the graph uniquely represents a real-world entity.

[0031] Specifically, the execution process of this step may include: S1, Preliminary Entity Recognition and Intra-Document Referential Resolution. Using a model finely tuned on data from a specific professional domain, Named Entity Recognition (NER) is performed on clean text to extract entities such as person names and organization names. Simultaneously, the pronoun referential problem is resolved within a single document; for example, in the sentence "Wang Wei went to the meeting room, where he met Li Mou," the pronoun "he" is accurately linked to the entity "Wang Wei."

[0032] Step S1 completes entity annotation and relationship linking within a single document, providing a basic, preliminarily disambiguated set of candidate entities for cross-document entity alignment.

[0033] S2, Cross-Document Entity Alignment and Fusion. It determines whether entities appearing in different texts refer to the same entity by calculating a preliminary "comprehensive similarity score." For example, "Director Wang" appears in conversation A, "Old Wang" in conversation B, and "Wang Wei" in conversation C. The system will analyze from multiple dimensions: Literal and phonetic features: The literal editing distance between "Teacher Wang", "Old Wang", and "Wang Wei" is large, and their similarity is low.

[0034] Contextual semantic features: Extracting the surrounding context of the three records, we found that the context of "Teacher Wang" in record A is "responsible for approving project A", the context of "Old Wang" in record B is "speaking at the project A meeting", and the context of "Wang Wei" in record C is "signing the final documents of project A". After converting these contexts into vectors using models such as BERT, we calculated that their cosine similarity is very high.

[0035] Entity attribute characteristics: The job title of "Wang Wei" extracted from external knowledge bases or text is "teacher," and his nickname may include "Old Wang." These attribute information are highly consistent.

[0036] Model decision: When analyzing, multi-dimensional feature fusion models (such as gradient boosting trees) will give higher weights to contextual semantic features and entity attribute features with high similarity, while reducing the weight of literal features with similarity. The final calculated comprehensive similarity score will exceed the preset threshold, thus determining that "Teacher Wang", "Old Wang", and "Wang Wei" all refer to the same entity object.

[0037] Step S2 combines multi-dimensional features to make a comprehensive judgment, which solves the problem of ambiguous entity reference caused by aliases, abbreviations, homophones, etc., and avoids the appearance of multiple isolated nodes representing the same person in the graph.

[0038] Step S102 ensures the uniqueness and accuracy of graph nodes through intra-document disambiguation and cross-document alignment. This overcomes the problem of fragmented entity recognition in traditional techniques when processing colloquial and unstructured text.

[0039] S103 extracts attributed relationships between entity objects from clean text that has been aligned with entities by using a large language model that has been fine-tuned by instructions, and associates them with original evidence fragments.

[0040] In this embodiment, this step extracts rich information about the "edges" that make up the graph from the text, realizing the transformation from simple relationships to depth-traceable relationships. Specifically, it includes the following sub-steps: S1. Define the relationship and attribute system and train the model. A "relationship label system" that meets business needs is predefined, such as "stakeholder interests" and "joint participation," and attribute fields such as time, location, and amount are defined for each relationship. Subsequently, using an instruction dataset containing a large number of samples (input text, extracted instructions, and structured relationships output), a basic large language model is fine-tuned to become a professional "relationship and attribute extraction" model.

[0041] Step S1 provides a framework tool for the relation extraction task, enabling the model to extract targeted information according to a pre-defined structure that meets professional requirements.

[0042] S2 performs deep relation extraction based on thought chains. Clean text is input into the trained model, and pre-designed prompts guide its extraction process. For example, given the input text fragment: "Person A said: 'Around last year before the Spring Festival, I gave Teacher Wang a technical document at that teahouse in the south of the city...'", the model might receive the instruction: "Please analyze the text, find the relationship between 'Person A' and 'Teacher Wang,' and output the time, location, amount, and original evidence in JSON format." The model will follow the thought chain, first locating key information, and then providing a structured output.

[0043] Step S2 utilizes the understanding and reasoning capabilities of the large model to transform unstructured natural language descriptions into structured relational data containing rich attributes, and forces them to be bound to original text evidence, thus solving the problems of incomplete information extraction and black-box operations.

[0044] This step S103 addresses the technical pain points of traditional graph information being single-dimensional and untraceable by extracting "edges" with rich attributes and direct evidence, making the graph an analyzable and verifiable knowledge network.

[0045] Step S104: Based on entity objects and attributed relationships, construct and incrementally merge the conversation content relationship graph.

[0046] This step is responsible for the dynamic construction and continuous updating of the knowledge map. It integrates information extracted from texts from different times and sources into a unified, dynamically evolving knowledge network, including the following steps: S1, Initialization and Incremental Integration. After processing the first conversation text, an initial graph is constructed based on the extracted entities and relationships. When a new conversation text is input, the system repeats the aforementioned steps and integrates the newly extracted information with the existing graph.

[0047] For example: Adding a new relationship: The graph already contains two nodes, "Zhang San" and "Li Si," and a relationship "(Zhang San) - [Gives money / gifts] -> (Li Si)" extracted from record A. When processing record B, a new relationship "(Zhang San) - [Requests a favor] -> (Li Si)" is extracted. Since this is a new event, a new relationship edge will be added between the two nodes.

[0048] Suppose that when processing C's conversation record later, the matter of "Zhang San" giving money to "Li Si" is also mentioned. If this event is determined to be the same objective event described in record A, it will not create a new relation edge. Instead, it will add evidence (citations, sources, etc.) from record C to the attribute list of the existing "[Giving Money]" edge, thereby achieving cross-verification of multiple pieces of evidence.

[0049] Step S104 integrates the isolated information extraction results into a dynamic and unified knowledge network, realizing the accumulation and reuse of knowledge. This allows the knowledge map to be continuously enriched and improved with the addition of new data, ultimately presenting a more comprehensive and reliable picture of interpersonal relationships and events, providing a platform for in-depth analysis.

[0050] This application provides a method for constructing a relationship graph based on conversation content. Through entity alignment using multi-dimensional feature fusion, it addresses common issues in spoken text such as homophones and multiple referents, ensuring the accuracy and uniqueness of graph nodes. Secondly, it utilizes a large language model with fine-tuned instructions for deep relationship extraction, upgrading the traditional single "person-relationship-person" triples in the graph into dynamic relationships with rich attributes such as time, location, event, and evidence source. Furthermore, it forcibly binds relationships to original evidence, resolving issues of information traceability and interpretability. This enables the construction of a rich, traceable deep relationship graph, improving the efficiency and depth of conversation content analysis and providing reliable technical support for the construction of evidence chains.

[0051] Secondly, embodiments of this application also provide a method for constructing a relationship graph based on conversation content. Figure 2 This is a structural block diagram of a relationship graph construction system based on conversation content, according to an embodiment of this application. Figure 2 As shown, the system includes: an acquisition module 20, an alignment module 21, an extraction module 22, and a construction module 23, wherein: The acquisition module 20 is used to acquire and process conversation texts from specific professional scenarios, generate clean texts to be analyzed, and the processing includes preprocessing and standardization of the conversation texts. Alignment module 21 is used to identify preliminary entities based on clean text. It calculates the comprehensive similarity score between preliminary entities through a multi-dimensional feature fusion model, and aligns the preliminary entities in the clean text according to the comprehensive similarity score to determine globally unique entity objects. Extraction module 22 is used to extract attributed relationships between entity objects from clean text that has undergone preliminary entity alignment through a large language model that has been fine-tuned by instructions, and to associate the original evidence fragments of attributed relationships in the clean text. The attributed relationships include preset relationship types and attribute information associated with the attributed relationships. Module 23 is used to construct a conversation content relationship graph based on entity objects and attributed relationships.

[0052] The system described above utilizes a large language model with fine-tuned instructions to extract deep relationships. This transforms the traditional single "person-relationship-person" triples in the graph into dynamic relationships with rich attributes such as time, location, event, and source of evidence. It also forcibly binds relationships to original evidence, solving the problems of information traceability and interpretability. This enables the construction of attribute-rich and traceable deep relationship graphs, improving the efficiency and depth of conversation content analysis and providing reliable technical support for the construction of evidence chains.

[0053] In one embodiment, Figure 3 This is a schematic diagram of the internal structure of an electronic device according to an embodiment of this application, such as... Figure 3 As shown, an electronic device is provided, which can be a server, and its internal structure diagram can be as follows. Figure 3 As shown, the electronic device includes a processor, a network interface, internal memory, and non-volatile memory connected via an internal bus. The non-volatile memory stores an operating system, computer programs, and a database. The processor provides computing and control capabilities, the network interface communicates with external terminals via a network connection, the internal memory provides an environment for the operating system to run, the computer programs are executed by the processor to implement a conversation-based relational graph construction method, and the database stores data.

[0054] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the electronic device to which the present application is applied. The specific electronic device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0055] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by hardware related to computer program instructions. This computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0056] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0057] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for constructing a relationship map based on conversation content, characterized by, The method includes: Acquire and process conversation texts from specific professional scenarios to generate clean texts to be analyzed. The processing includes preprocessing and standardizing the conversation texts. Entity recognition is performed based on the clean text to determine preliminary entities. A multi-dimensional feature fusion model is used to calculate the comprehensive similarity score between the preliminary entities. The preliminary entities in the clean text are then aligned according to the comprehensive similarity score to determine globally unique entity objects. Using a large language model finely tuned by instructions, attributed relationships between entity objects are extracted from the clean text after the initial entity alignment, and the original evidence fragments of the attributed relationships in the clean text are associated with them. The attributed relationships include a preset relationship type and attribute information associated with the attributed relationships. Based on the attributed relationships between the entity objects and the associated original evidence fragments, a conversation content relationship graph is constructed.

2. The method of claim 1, wherein, Preprocessing the conversation text includes: Using a language model or pre-defined error correction rules, the homophonic errors of proper nouns in the conversation text are initially corrected, and text normalization processing is performed. The text normalization processing includes: removal of colloquial words, merging of repetitive expressions, and regularization of sentence logic. Furthermore, the normalized conversation text is segmented into semantic units centered on specific events and with limited context, according to dialogue rounds, and / or time stamps, and / or topic clustering, to obtain the clean text.

3. The method of claim 1, wherein, The comprehensive similarity score between the preliminary entities is calculated using a multi-dimensional feature fusion model, including: Obtain the multidimensional features of the preliminary entity, wherein the multidimensional features include: literal features, contextual semantic features, and entity attribute features; By introducing a negative sample learning mechanism into the machine learning model, the comprehensive similarity score is obtained by performing a weighted calculation based on the multidimensional features and preset weights. The literal features are obtained by calculating the literal edit distance of the entity name and / or the similarity of the pinyin sequence, and the contextual semantic features are obtained by converting the context fragments around the initial entity into vector representations and calculating the cosine similarity between the vector representations.

4. The method of claim 3, wherein, The entity attribute features are obtained by extracting attribute information associated with the entity object from the clean text of the initial entity alignment and comparing the similarity of the attribute information. The attribute information includes job information, affiliated unit information, place of origin information or known associated person information.

5. The method according to claim 1, characterized in that, Attributed relations are extracted from a large language model that has been fine-tuned by instructions, including: By using instruction fine-tuning technology, a basic large language model is fine-tuned and trained based on a predefined structured instruction dataset that meets the needs of specific professional scenarios, resulting in a fine-tuned model that can execute relation and attribute extraction instructions. The fine-tuning model is input with clean text that has undergone preliminary entity alignment, and guided by preset prompts, it performs deep extraction based on thought chains and outputs the attributed relationships. The prompts are used to constrain the output of the fine-tuning model to include original text evidence fragments for verifying the existence of the attributed relationships.

6. The method according to claim 1, characterized in that, The attribute information specifically includes: the time when the relationship occurred, the location where the relationship occurred, the items or amounts involved in the relationship, and a textual description of the event process; By annotating the key citations on which the attributed relationships are based, as well as the position information of the key citations in the original conversation text, in the edge attributes of the relationship graph of the conversation content, the original evidence fragments can be associated.

7. The method according to claim 1, characterized in that, The method further includes: Obtain the updated conversation text, and extract the updated attributed relationships and their associated entity objects from the updated conversation text; Entity alignment is determined for the associated entity objects. If the entity object already exists in the current conversation content relationship graph, the updated attribute relationship is added as a new edge between the corresponding entity objects. If the updated attributed relationship is judged and it belongs to the same objective event as the target edge that already exists in the conversation content relationship graph, then the original evidence fragment associated with the updated attributed relationship is added as a new evidence source to the attribute list of the target edge.

8. A relationship graph construction system based on conversation content, characterized in that, The system includes: The acquisition module is used to acquire and process conversation texts from specific professional scenarios, and generate clean texts to be analyzed. The processing includes preprocessing and standardizing the conversation texts. The alignment module is used to perform entity recognition based on the clean text to determine preliminary entities, calculate the comprehensive similarity score between the preliminary entities through a multi-dimensional feature fusion model, and align the preliminary entities in the clean text according to the comprehensive similarity score to determine globally unique entity objects. The extraction module is used to extract attributed relationships between entity objects from the clean text after the initial entity alignment using a large language model that has been fine-tuned by instructions, and to associate the attributed relationships with the original evidence fragments in the clean text. The attributed relationships include a preset relationship type and attribute information associated with the attributed relationship. A construction module is used to construct a conversation content relationship graph based on the attributed relationships between the entity objects and the associated original evidence fragments.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images