CN-122021596-A - Conference content structured recording method based on knowledge graph

CN122021596ACN 122021596 ACN122021596 ACN 122021596ACN-122021596-A

Abstract

The invention discloses a conference content structured recording method based on a knowledge graph, which comprises the steps of collecting conference multisource data, executing cleaning standardization processing to generate a standardized dataset, building a conference fact behavior graph, building a hierarchical structure and association relation, building a self-growing standardization graph generator, extracting a behavior pattern, evolving to generate a standardization knowledge graph, executing bidirectional consistency reasoning to generate a desired link, calculating the link consistency, inputting the graph and the structure to a Graphormer network, executing global coding to generate a conference hidden sequence vector, generating a structured record based on a hidden sequence and the consistency, executing filling prompt, and updating a model and the graph. According to the invention, intelligent structured recording and standard consistency check of the whole process content of the conference are realized by constructing a conference fact behavior pattern, a self-growing standard knowledge pattern and an improved Graphormer network.

Inventors

LI JIEKE
CHEN XI
LI MINGTING
WANG MENGLAN
YE LIANG
WANG MIAO
GUO JIA
LIU FUPENG
DU BIYU
YUAN YONGQING
BAI XUE
WU YIXI
ZHANG XINGXIA

Assignees

四川大学华西医院
成都布鲁奥森信息科技有限责任公司

Dates

Publication Date: 20260512
Application Date: 20260127

Claims (8)

1. The structured recording method for the conference content based on the knowledge graph is characterized by comprising the following steps of: collecting conference multisource business data in conference activities, and preprocessing the conference multisource business data to obtain a normalized conference data set; establishing a conference fact behavior map based on a normalized conference data set, identifying conference link entities and behavior micro-unit entities, establishing a behavior hierarchical structure, and establishing a sequence relationship, a dependency relationship and a reference relationship among behavior nodes according to a syntactic dependency relationship and an event association relationship; constructing a self-growth canonical spectrum generator based on the conference fact behavior spectrum, and executing self-growth processing on the conference behavior pattern to generate a canonical knowledge spectrum; Based on the standard knowledge graph and the conference fact behavior graph, performing bidirectional logic consistency reasoning, generating an expected conference behavior chain through downward standard driving reasoning, identifying missing behaviors and link anomalies in the conference behavior chain through upward fact driving back-pushing, and calculating link consistency degree; Inputting the conference fact behavior pattern, the standard knowledge pattern and the behavior hierarchy structure into an improved Graphormer network, and executing global structure coding processing to generate a conference hidden sequence vector; And generating a structured conference record based on the conference hidden sequence vector and the link consistency, executing conference basic information filling, conference process chain generation and missing element prompt processing, and updating the standard knowledge graph and Graphormer network.
2. The method for structured recording of conference content based on knowledge graph according to claim 1, wherein the conference multisource business data comprises conference notification text, conference summary text, check-in table data, topic content data, learning material summary data, resolution item data and implementation record data.
3. The method according to claim 1, wherein preprocessing the conference multisource business data includes performing a sentence processing, a word segmentation processing, a clause division processing, and a standardization processing of personnel names, organization names, and time information on the conference notification text, the conference summary text, and the form data, and performing a field mapping processing on the form data.
4. The structured recording method of conference content based on knowledge graph according to claim 1, wherein the building of the conference fact behavior graph based on the normalized conference data set comprises: Carrying out speech segment division and time axis alignment processing on the normalized conference data set, dividing conference notification texts, conference summary texts and form data into speaking round segments according to time sequence, and marking speaker identification, time stamp and data source type; Executing entity and event extraction processing on the speaking round section, identifying a conference link entity and a behavior micro-unit entity, and constructing a behavior hierarchical structure with the conference link entity as a father node and the behavior micro-unit entity as a child node based on extracted links and micro-units; generating a relation edge aiming at the behavior hierarchical structure, establishing a sequential relation edge according to the speaking turn and the time stamp, establishing a dependency relation edge according to the relation between the trigger word and the argument, establishing a reference relation edge according to the data reference and the issue instruction, establishing a participation relation edge between the interviewee and the behavior micro-unit, and establishing an evidence tracing edge pointing to the original language segment and the field position for each node and the relation edge; Performing cross-source fusion and conflict resolution processing on co-fingered nodes and synonymous relations derived from meeting notification text, meeting summary text and form data, generating a unique topic main chain for each topic based on time sequence and dependency constraint, and setting occupation identification on missing necessary nodes to keep links coherent; outputting a conference fact behavior map comprising a behavior hierarchical structure, a sequence relation side, a dependency relation side, a reference relation side, a participation relation side and an evidence tracing side.
5. The method for structured recording of conference content based on knowledge-graph according to claim 1, wherein said generating a canonical knowledge-graph comprises: receiving a conference fact behavior pattern, initializing a self-growing canonical pattern generator, creating a processing task queue facing the boundary and the time axis of the topic, and registering a behavior pattern acquisition layer, a pattern structure pattern abstract layer and a canonical pattern evolution and stabilization layer; Performing segment sliding window scanning processing on the conference fact behavior map by a behavior pattern acquisition layer, extracting sequential chain segments, dependent chain segments, reference chain segments and participation chain segments, reserving or eliminating each segment by adopting a double-threshold confidence degree discrimination strategy, and merging repeated segments by using a cross-conference time window aggregation strategy to form a behavior pattern segment set; Executing multidimensional consistency alignment processing on the behavior pattern fragment set by the graph structure pattern abstract layer, carrying out unified normalization on node types, relationship types and entity and fingers, quickly removing duplication in a structural fingerprint hash mode, generating a canonical candidate structure set, and setting space occupying nodes for the missing necessary nodes by a cross-layer space occupying identification mechanism; And executing the adaptive evolution processing of the standard structure on the standard candidate structure set by the standard spectrum evolution and stabilization layer, executing the growth processing according to the occurrence frequency of the fragments and the consistency score, executing the attenuation processing according to the continuous low frequency and conflict detection, executing the dependency reinforcement processing according to the link closure contribution degree, judging whether the variation of the conference fact behavior spectrum reaches the stability threshold value or not by the rolling steady state evaluation mechanism, and outputting the standard knowledge spectrum.
6. The structured recording method of conference content based on knowledge-graph according to claim 1, wherein said calculating link consistency comprises: according to a canonical dependency chain corresponding to the target conference type and the organization level in the canonical knowledge graph, determining conference behavior nodes which the target conference should contain, canonical sequence relations and canonical closed-loop relations between the nodes, and arranging the behavior nodes and the canonical relations into an expected conference behavior chain; In the conference fact behavior map, taking the topic boundary and the time sequence as constraints, extracting actual sequence relations and actual closed-loop relations among actually-generated conference behavior nodes, and arranging the behavior nodes and the actual relations into an actual conference behavior chain; Performing a behavior node-by-behavior node and relation-by-relation matching process on an expected conference behavior chain and an actual conference behavior chain, marking expected behavior nodes covered by the actual conference behavior, adjacent node pairs with consistent sequences and behavior links forming a closed loop according to specification requirements, and generating an uncovered expected behavior node set and a specification link set not forming a closed loop; Based on the total number of behavior nodes, the total number of sequence relations and the total number of standard closed loop links in an expected conference behavior chain, the number of covered behavior nodes, the number of relations consistent in sequence and the number of links actually forming a closed loop are combined, the coverage condition, the sequence maintenance condition and the closed loop implementation condition are calculated respectively, and the three conditions are integrated according to a preset weight rule to obtain the link consistency.
7. The method for structured recording of conference content based on knowledge-graph according to claim 1, wherein the generating the conference hidden sequence vector comprises: merging the conference fact behavior pattern, the standard knowledge pattern and the behavior hierarchical structure to form a joint input pattern serving as Graphormer network input features; An improved Graphormer network is constructed, the improved Graphormer network sequentially comprises an input feature embedding layer, eight graph transform coding blocks, two canonical-fact dual-channel attention modules, a hierarchical convergence gating unit and an output head, wherein: The input feature embedding layer performs unified embedding on the nodes and the edge marks and adds organization topological codes; Inserting a specification-fact dual-channel attention module after the third and fifth graph Transformer coding blocks for intersecting the specification nodes and the fact nodes within the same layer; each coding block is provided with residual connection and normalization processing, the hidden dimension is set to be five hundred twelve, the attention head number is set to be eight, and the discarding rate is set to be ten percent; the hierarchical aggregation gate control unit performs step-by-step aggregation in three stages according to the sequence of the behavior micro unit, the conference link and the conference stage, and controls the information flow by using the organization topology priori mask for each stage of output; Performing joint training on an improved Graphormer network, wherein training input is joint input graphs and labels, the labels comprise missing behavior marks, link abnormality marks and the consistency of the category and the link of the next walking, the training adopts a mode of combining supervised learning and self-supervised tasks, a supervising part uses multi-label classification loss, a self-supervising part uses edge shielding recovery tasks, and the learning rate and regularization strength are dynamically adjusted in the training process; In the reasoning stage, the joint input diagram is input into a modified Graphormer network, and the conference hidden sequence vector is output through multi-layer coding, specification-fact interaction attention and hierarchical convergence processing.
8. The method for structured recording of conference content based on knowledge-graph according to claim 1, wherein the generating the structured conference recording based on the conference hidden sequence vector and the link consistency comprises: Receiving conference hidden sequence vectors and link consistency, loading a record generation rule set according to a behavior hierarchy structure, and completing record template selection, paragraph layout determination and field enabling configuration; According to a record template, importance sorting and field mapping are carried out on conference links and behavior micro units indicated by the conference hidden sequence vector, corresponding information in a conference fact behavior map is written into an enabling field, checking is carried out on the process integrity level reflected by the link consistency, filling checking and consistency checking are started or tightened, and occupation and supplementary prompt are set for the necessary fields which are not covered yet; generating a conference process chain item based on a behavior sequence and an association relation given by a conference hidden sequence vector, writing the link consistency, coverage condition and abnormal prompt into the conference process chain item as record head indexes, and marking a source language segment and a position index of a map evidence in each item to form a structured conference record draft; And executing consistency review and formatting output on the structured meeting record draft, generating a structured meeting record file which can be backfilled to a management platform, and storing the corresponding relation of the field-node-relation.

Description

Conference content structured recording method based on knowledge graph Technical Field The invention relates to the technical field of data processing and knowledge graph construction, in particular to a conference content structured recording method based on knowledge graphs. Background With the development of informatization office work and digital management, various conferences have become important carriers for organization decision, business collaboration and management execution. Current meeting content typically exists in the form of meeting notices, meeting notes, check-in lists, issue materials, and implementation records, which are mostly stored in a text or form-and-list fashion. The existing conference management system mainly focuses on conference flow management and document archiving, the processing of conference contents is still mainly performed by manual arrangement and static field filling, the whole conference process is difficult to be structurally expressed, and conference links, behavior sequences and internal logic relations thereof cannot be accurately described from conference texts. In recent years, natural language processing and knowledge graph technology are gradually applied to the field of conference text analysis, and some technical schemes attempt to carry out structuring processing on conference contents through entity extraction, keyword recognition or template matching. However, such techniques typically focus only on surface information in the meeting content, lacking the ability to model the system of the chain of meeting behavior, the process of subject evolution, and decision and implementation relationships. In the prior art, when facing multi-source conference data, the common-finger information and semantic conflicts in different sources are difficult to effectively fuse, and a sustainable evolution conference standard structure cannot be formed, so that a structural result is unstable and the universality is insufficient. The conventional conference content analysis method generally lacks quantitative evaluation capability for consistency between actual behaviors and expected flows of a conference, and cannot automatically identify the problems of missing links, abnormal sequence or incomplete closed loop in the conference process. In terms of depth modeling, it is difficult for a generic text model or shallow graph model to express conference fact structure and canonical logic at the same time, and to generate a global representation that can be used directly to automatically generate structured conference recordings. Therefore, how to provide a structured recording method for conference content based on knowledge graph is a problem that needs to be solved by those skilled in the art. Disclosure of Invention The invention aims to provide a conference content structured recording method based on a knowledge graph, which fully utilizes knowledge graph construction technology, graph structure mode abstract technology and an improved graph transform network, and realizes automatic structured expression and normalization check of conference content by constructing a conference fact behavior graph, generating a standard knowledge graph capable of adaptively evolving, executing bidirectional logic consistency reasoning and outputting conference hidden sequence vectors. The method can completely characterize the conference flow, identify missing elements and quantify the flow consistency, automatically generate the structured conference record, and has the advantages of strong standardization, high calculability, high intelligent degree and sustainable improvement of recording quality. According to the embodiment of the invention, the conference content structured recording method based on the knowledge graph comprises the following steps: collecting conference multisource business data in conference activities, and preprocessing the conference multisource business data to obtain a normalized conference data set; establishing a conference fact behavior map based on a normalized conference data set, identifying conference link entities and behavior micro-unit entities, establishing a behavior hierarchical structure, and establishing a sequence relationship, a dependency relationship and a reference relationship among behavior nodes according to a syntactic dependency relationship and an event association relationship; constructing a self-growth canonical spectrum generator based on the conference fact behavior spectrum, and executing self-growth processing on the conference behavior pattern to generate a canonical knowledge spectrum; Based on the standard knowledge graph and the conference fact behavior graph, performing bidirectional logic consistency reasoning, generating an expected conference behavior chain through downward standard driving reasoning, identifying missing behaviors and link anomalies in the conference behavior chain through upward fact driving back-