Intelligent construction method and system for semantic path structure of conversation record
By constructing a five-element semantic path structure for conversation records, the problems of the dispersion of key semantic elements in conversation records and the adaptation to general models are solved, achieving efficient semantic association and logical conflict detection, and providing unified data structure support.
Patent Information
- Application Number
- CN202512056089.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-14
AI Technical Summary
Existing technologies struggle to efficiently and accurately extract and organize core facts from unstructured conversation transcripts, and lack the ability to detect semantic associations and conflicts across texts. General NLP models are also ill-suited to the specific semantics of the discipline inspection field.
Using a semantic analysis model and a domain behavior ontology library, a chain-like semantic path is constructed through five-element semantic units, and path vectors are generated by combining neural networks, which are then stored and visualized.
It achieves efficient structuring of conversation records, supports semantic association and logical conflict detection across sentences and documents, provides a solid data foundation, and supports consistency analysis and evidence chain generation.
Smart Images
Figure CN121859924A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing technology, specifically to a method and system for intelligently constructing the semantic path structure of conversation records. Background Technology
[0002] In disciplinary inspection, auditing, and accountability investigations, interview records are crucial written materials for recording facts, securing evidence, and reflecting the case. These records are typically written in natural language and contain a wealth of narrative information about people, actions, events, times, and places. The fundamental challenge in efficiently and accurately extracting and organizing core facts from these unstructured texts is to support subsequent review, comparison, and inferential analysis.
[0003] Currently, the main technical challenges in processing this type of text are as follows:
[0004] Key semantic elements are difficult to represent in a structured way: Core information in conversation transcripts (such as the people involved, specific actions, event content, time points, and locations) is usually scattered across different sentences and paragraphs, presenting an unstructured or semi-structured state. Existing methods struggle to automatically and systematically extract and organize this information into a unified, computable, and standardized form, hindering efficient automated analysis and information integration.
[0005] Semantic relationships and conflicts across texts are difficult to identify automatically: descriptions of the same event may form complementary, corroborative, or contradictory relationships between different conversation records or different parts of the same record (e.g., one person states "received payment," while another claims "never paid"). Existing natural language processing (NLP) technologies, such as named entity recognition (NER) and keyword matching, can usually only identify specific entities or words at the isolated sentence level. They lack the ability to deeply understand and model the semantic level (such as the affirmation or negation of the nature of behavior, and the association and continuation of events), and therefore cannot automatically construct semantic correspondence chains across sentences and documents, making it even more difficult to identify potential logical conflicts.
[0006] General-purpose NLP models struggle to adapt to domain-specific semantics: the behavioral patterns (such as admission, denial, defense, evasion, and accusation) and role types (such as parties involved, witnesses, and handlers) in the discipline inspection and supervision field possess distinct professional characteristics and specific expression systems. General-purpose NLP models and event extraction methods do not incorporate domain-specific knowledge ontology, resulting in limited accuracy in identifying and classifying these specific semantics and insufficient usability of the extraction results.
[0007] The lack of ability to construct a structured "semantic chain" from elements: Existing technical solutions mostly focus on the identification of discrete entities or local events, failing to achieve a complete construction from "entity" to "relationship" and then to "structured event chain". Specifically, there is a lack of effective methods to organically link the five core semantic elements of "who", "what did / said", "about what", "when", and "where", forming a coherent "semantic path" that can represent complete factual fragments. Therefore, it is impossible to provide a directly computable data structure for advanced applications such as consistency checks based on logic and constraints and automatic generation of evidence chains.
[0008] Therefore, there is an urgent need for a deep semantic structuring method specifically for disciplinary inspection interview texts. This method needs to be able to deeply understand the domain semantics of the text, automatically extract key elements, and construct these elements into a unified and coherent chain structure according to their inherent logical relationships, providing a solid data foundation for the intelligent analysis, review, and comparison of interview records. Summary of the Invention
[0009] To address this, embodiments of the present invention provide a method and system for intelligently constructing the semantic path structure of conversation records, thereby solving the technical problem that existing technologies can only extract isolated entities or events, making it difficult to support cross-sentence and cross-document semantic association and logical conflict detection.
[0010] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:
[0011] According to a first aspect of the present invention, a method for intelligently constructing the semantic path structure of a conversation record is provided, the method comprising:
[0012] The conversation transcript is preprocessed and segmented to identify the speaker and their role in the conversation, resulting in a segmented sequence of sentences.
[0013] Based on semantic analysis models and domain behavior ontology libraries, each statement is extracted with person entity, behavior type, event content, time information, and location information to form a five-element semantic unit that represents the core semantics of the statement.
[0014] Based on semantic dependency relationships and contextual associations, the elements in the five semantic units are linked in a preset logical order to generate a chain-like semantic path;
[0015] Based on the vector representation of the five semantic units and their contextual information, a path vector representing the entire semantic path is generated through a neural network model.
[0016] The generated chain semantic paths and their corresponding path vectors are stored and / or visualized.
[0017] Furthermore, the role recognition includes distinguishing the participants in the conversation based on text format marking, referential resolution, or voiceprint features, and classifying their roles as at least one of the following: party, witness, handler, or informant.
[0018] Furthermore, the behavior type identification relies on a preset domain behavior ontology library, which maps the extracted behavior descriptions to at least one standardized behavior category among admission, denial, defense, evasion, accusation, recollection, or explanation.
[0019] Furthermore, the event content recognition includes: extracting core phrases describing the event and normalizing different event phrases that express the same semantics.
[0020] Furthermore, the time information recognition includes parsing the fuzzy or relative time expressed in natural language and converting it into a standardized time representation or time interval.
[0021] Furthermore, the path vector is generated by taking into account the input of a unit representation vector obtained by fusing the element vectors of the five semantic units.
[0022] Furthermore, the generation of the path vector includes, as input, a context encoding vector of semantic path information before and after the current statement.
[0023] Furthermore, the construction of the context encoding direction is selected from any one or more combinations of the following: semantic paths of adjacent statements, continuous behavioral sequences belonging to the same person, and cross-statement continuation relationships describing the same event.
[0024] Furthermore, the generated chain semantic paths and their corresponding path vectors are stored, including:
[0025] The chain-like semantic path is stored in a graph database as a relationship between nodes and edges, and the path vector is stored in a vector database to support path retrieval and association analysis based on semantic similarity.
[0026] According to a second aspect of the present invention, a semantic path structure intelligent construction system for conversation recordings is provided, the system comprising:
[0027] The conversation text parsing and role recognition module is used to preprocess and segment the conversation transcript text, identify the speaker and their role type in the conversation, and obtain the sentence sequence after segmentation.
[0028] The key semantic element extraction module is used to extract the person entity, behavior type, event content, time information and location information from each statement based on the semantic analysis model and the domain behavior ontology library, forming a five-element semantic unit that represents the core semantics of the statement.
[0029] The semantic path construction module is used to link the elements in the five semantic units in a preset logical order according to semantic dependency relationships and context associations to generate a chain semantic path.
[0030] The path vectorization representation module is used to generate a path vector representing the entire semantic path through a neural network model based on the vector representation of the five semantic units and their context information.
[0031] The structured output module is used to store and / or visualize the generated chain semantic paths and their corresponding path vectors.
[0032] The embodiments of the present invention have the following advantages:
[0033] This invention aims to address the problems of scattered and difficult-to-structure association of key semantic elements in conversation transcripts, and the difficulty of adapting general models to domain-specific semantics. The method includes: parsing and role identification of the conversation text; extracting five categories of elements—person, behavior, event, time, and location—from each statement based on a domain behavior ontology library to form five-element semantic units; linking elements within units according to semantic dependencies and contextual relationships to generate chain-like semantic paths; further integrating contextual information to generate vectorized representations of the paths; and finally outputting a storable and visualizeable semantic path structure. This invention transforms unstructured text into a unified and computable chain-like semantic expression, providing a fundamental data structure for consistency analysis, conflict detection, and evidence chain generation in conversation transcripts. Attached Figure Description
[0034] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.
[0035] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.
[0036] Figure 1 A schematic diagram of the logical structure of a semantic path structure intelligent construction system for conversation recording provided in an embodiment of the present invention;
[0037] Figure 2 A flowchart illustrating a method for intelligently constructing the semantic path structure of a conversation record, as provided in an embodiment of the present invention;
[0038] Figure 3 A schematic diagram of the key semantic element extraction module in a semantic path structure intelligent construction method for conversation records provided in an embodiment of the present invention;
[0039] Figure 4 This is a schematic diagram of the five-element fusion vector construction structure in a semantic path structure intelligent construction method for conversation records provided in an embodiment of the present invention;
[0040] Figure 5 This is a schematic diagram of semantic path generation in a method for intelligently constructing the semantic path structure of a conversation record, provided in an embodiment of the present invention.
[0041] Figure 6 This is a schematic diagram of the path vector construction structure in a semantic path structure intelligent construction method for conversation records provided in an embodiment of the present invention. Detailed Implementation
[0042] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] To address the technical problem that existing technologies can only extract isolated entities or events, making it difficult to support cross-sentence and cross-document semantic association and logical conflict detection.
[0044] refer to Figure 1 This invention discloses an intelligent system for constructing the semantic path structure of conversation records. The system includes: a conversation text parsing and role recognition module 1; a key semantic element extraction module 2; a semantic path construction module 3; a path vectorization representation module 4; and a structured output module 5.
[0045] Corresponding to the aforementioned intelligent construction system for the semantic path structure of conversation records, this invention also discloses an intelligent construction method for the semantic path structure of conversation records. The following details the intelligent construction method for the semantic path structure of conversation records disclosed in this invention, in conjunction with the aforementioned intelligent construction system for the semantic path structure of conversation records.
[0046] refer to Figures 2 to 6This invention discloses an intelligent method for constructing the semantic path structure of a conversation record, comprising: preprocessing and segmenting the conversation record text, identifying the speaker and their role type in the conversation, and obtaining a sequence of sentences after segmentation; extracting the person entity, behavior type, event content, time information, and location information from each sentence based on a semantic analysis model and a domain behavior ontology library to form a five-element semantic unit representing the core semantics of the sentence; linking the elements in the five-element semantic unit in a preset logical order according to semantic dependency relationships and contextual associations to generate a chain-like semantic path; generating a path vector representing the entire semantic path through a neural network model based on the vector representation of the five-element semantic unit and its contextual information; and storing and / or visualizing the generated chain-like semantic path and its corresponding path vector.
[0047] Example 1
[0048] Suppose we have a transcript of a disciplinary inspection interview, containing multiple rounds of dialogue. The implementation steps of this method are as follows:
[0049] S101: Conversation Text Parsing and Role Recognition
[0050] First, the input raw conversation text is preprocessed, including removing meaningless interjections, repetitive expressions, and other noise. Next, based on periods, question marks, semicolons, and semantic integrity boundaries, the text is segmented into independent semantic sentences. Then, the system identifies the speaker's identity for each sentence according to the inherent conversation transcript format within the text. Furthermore, using a pre-trained role classification model and considering contextual referential relationships, the system categorizes the speaker's role as a party involved, witness, or handler, etc. The final output is a sequence of sentences labeled with the speaker and their role.
[0051] S102: Key Semantic Element Extraction
[0052] For each statement obtained in step S101, five types of key information are extracted using a combination model. Specifically:
[0053] Person entity recognition: Using a named entity recognition model based on domain dictionary and context fine-tuning, the system identifies all persons' names, titles or pseudonyms (such as "he", "Director Li") appearing in the sentence and normalizes them into standard person identifiers.
[0054] Behavior type recognition: Relying on a pre-built ontology of disciplinary inspection behaviors, which defines behavioral categories such as admission, denial, justification, and shirking responsibility, along with their common expression patterns, the system uses semantic matching and classification models to determine which category the core behavior of the current statement belongs to.
[0055] Event content recognition: Through event trigger word detection and phrase extraction technology, phrases describing core facts (such as receiving subsidies, modifying contracts) are extracted and normalized using an event thesaurus.
[0056] Time information recognition: Using time parsing tools, identify absolute time (such as May 1, 2023), relative time (such as last week, after the meeting), and fuzzy time (such as the end of the year), and normalize them into standard timestamps or calculable time intervals.
[0057] Location information recognition: Identify specific location names (such as Beijing) and complete and standardize abstract locations such as meeting rooms and offices based on the context of the conversation (such as company names).
[0058] The above five elements together constitute a five-element semantic unit U corresponding to this statement.
[0059] Step S103, semantic path construction.
[0060] For each five-element semantic unit U, the elements are linked according to the logical order of person, action, event, time, and place to form a basic chain-like semantic path. For example, for the statement "Zhang denies receiving a subsidy in his office last week," the constructed path is: Zhang → denies → received subsidy → last week → office. Simultaneously, by analyzing the syntactic dependencies between statements, behavioral logic (such as defense usually targeting a specific accusation), and contextual reference, the system connects the paths of different statements at nodes such as person and event, initially forming a rudimentary network structure.
[0061] The format of a quinary semantic unit is as follows: .
[0062] Step S104: Path Vectorization Representation
[0063] To facilitate deep analysis and comparison by computers, semantic paths need to be transformed into numerical vectors. First, each element in each five-element semantic unit is converted into a vector using a pre-trained language model, and then fused through a multilayer perceptron to obtain the unit representation vector S. Second, a context encoding vector Context is constructed. This vector aggregates information from adjacent paths before and after the current statement, the behavioral sequence information of the same character in the dialogue, and the continuation information of the same event in different statements through an attention mechanism. Finally, the unit representation vector S, the context encoding vector Context, and the specific encoding vectors for character, time, and location are concatenated and input into another multilayer perceptron model to generate the final path vector Z. This path vector comprehensively represents all the semantics and contextual information of the semantic path.
[0064] Semantic units are encoded into vector representations: Person vector: incorporating name, job title, and organizational structure; Behavior vector: combining behavior ontology embedding; Event vector: utilizing event phrase embedding; Time vector: normalized timestamp + natural language time encoding; Location vector: location representation + geographic encoding, jointly input into MLP / Transformer to generate fused vectors. .
[0065] By analyzing grammatical dependency, semantic relationships between sentences, and behavioral logic, a chain-like semantic path is generated:
[0066]
[0067] Simultaneously, we construct path hierarchy relationships, path context adjacency relationships, similarity and associated edges between multiple paths, and calculate path vectors (Path-Embedding):
[0068]
[0069] It includes: the relationship between actions and events in adjacent sentences, the chain of continuous actions of the same character, and the chain of cross-sentence continuation of the same event.
[0070] Step S105: Structured Output
[0071] All chained semantic paths generated in step S103, and the corresponding path vectors calculated in step S104, are stored in a graph database (such as Neo4j) and a vector database (such as Milvus). In the graph database, elements such as people, behaviors, and events are used as nodes, and the order and association between nodes are used as edges, forming a visual graph. In the vector database, path vectors support fast retrieval based on similarity. The system can also provide various visualizations such as timeline views and character behavior flow diagrams to help reviewers intuitively understand the internal structure of the conversation.
[0072] Example 2
[0073] Implementation of cross-sentence semantic fusion
[0074] This embodiment is a further optimization of Embodiment 1, specifically for situations where event descriptions are scattered across multiple consecutive statements.
[0075] After completing the element extraction in step S102, a cross-sentence semantic fusion step is performed. The system will detect whether the current sentence and the sentences before and after it belong to the same speaker and describe the same core event. For example, two consecutive sentences, "I was organizing materials in the meeting room" and "Zhang came over afterwards," describe two successive actions, but belong to the same scene.
[0076] The system merges the five-element semantic units extracted from the two statements: the person element is the union (person A, Zhang), the time element is integrated into a sequential relationship based on the description (T1, after), the location element is the meeting room, and the behavior and event elements are integrated into organizing materials and Zhang's arrival. This fusion generates a new, more complete five-element semantic unit U' to represent this micro-scene. Subsequent path construction and vectorization will be based on U', thereby generating a semantic path that better reflects contextual coherence and effectively solves the problem of information fragmentation.
[0077] The embodiments of the present invention achieve a high degree of structuring, which can transform natural language conversation records into a refined five-element semantic path structure, thereby completing the systematic extraction of key elements in the text.
[0078] Adapting to the semantic system of the discipline inspection field, the system significantly improves the accuracy and professionalism of semantic extraction by introducing discipline inspection behavior ontology and role semantic knowledge. The constructed path structure is naturally suitable for cross-document comparison and calculation based on multi-dimensional constraints, laying an ideal data foundation for subsequent conflict detection and consistency analysis.
[0079] This structure possesses excellent interpretability, visually displaying the semantic logic of conversations through chain-like graphs and other formats, assisting reviewers in quickly understanding core facts. The embodiments of this invention exhibit strong scalability; its technical approach and framework can be extended to various similar scenarios requiring in-depth semantic analysis and structuring of conversation records, such as auditing, compliance investigations, and judicial inquiries.
[0080] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.
Claims
1. A method for intelligently constructing the semantic path structure of a conversation record, characterized in that, The method includes: The conversation transcript is preprocessed and segmented to identify the speaker and their role in the conversation, resulting in a segmented sequence of sentences. Based on semantic analysis models and domain behavior ontology libraries, each statement is extracted with person entity, behavior type, event content, time information, and location information to form a five-element semantic unit that represents the core semantics of the statement. Based on semantic dependency relationships and contextual associations, the elements in the five semantic units are linked in a preset logical order to generate a chain-like semantic path; Based on the vector representation of the five semantic units and their contextual information, a path vector representing the entire semantic path is generated through a neural network model. The generated chain semantic paths and their corresponding path vectors are stored and / or visualized.
2. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The role recognition includes distinguishing the participants in the conversation based on text format marking, referential resolution, or voiceprint features, and classifying their roles as at least one of the following: party, witness, handler, or informant.
3. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The behavior type identification relies on a preset domain behavior ontology library, which maps the extracted behavior descriptions to at least one standardized behavior category among admission, denial, defense, evasion, accusation, recollection, or explanation.
4. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The event content recognition includes: extracting core phrases describing the event and normalizing different event phrases that express the same semantics.
5. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The time information recognition includes parsing fuzzy or relative time expressions in natural language and converting them into standardized time representations or time intervals.
6. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The path vector is generated by inputting a unit representation vector obtained by fusing the element vectors of the five semantic units.
7. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The path vector is generated by taking into account the context encoding vector of the semantic path information before and after the current statement.
8. A method for intelligently constructing the semantic path structure of a conversation record as described in claim 6 or 7, characterized in that, The construction of the context encoding direction is selected from any one or more combinations of the following: semantic paths of adjacent statements, continuous action sequences belonging to the same person, and cross-statement continuation relationships describing the same event.
9. The method for intelligently constructing the semantic path structure of a conversation record as described in claim 1, characterized in that, The generated chain semantic paths and their corresponding path vectors are stored, including: The chain-like semantic path is stored in a graph database as a relationship between nodes and edges, and the path vector is stored in a vector database to support path retrieval and association analysis based on semantic similarity.
10. A semantic path structure intelligent construction system for conversation recordings, characterized in that, The system includes: The conversation text parsing and role recognition module is used to preprocess and segment the conversation transcript text, identify the speaker and their role type in the conversation, and obtain the sentence sequence after segmentation. The key semantic element extraction module is used to extract the person entity, behavior type, event content, time information and location information from each statement based on the semantic analysis model and the domain behavior ontology library, forming a five-element semantic unit that represents the core semantics of the statement. The semantic path construction module is used to link the elements in the five semantic units in a preset logical order according to semantic dependency relationships and context associations to generate a chain semantic path. The path vectorization representation module is used to generate a path vector representing the entire semantic path through a neural network model based on the vector representation of the five semantic units and their context information. The structured output module is used to store and / or visualize the generated chain semantic paths and their corresponding path vectors.