A conference minutes generation method fusing text optimization and semantic relation resolution

Abstract

The application relates to a conference minutes generation method combining text optimization and semantic relationship analysis. First, the voice recognition component is used to transcribe the conference recording in real time, and the transcription result is processed again by combining a text optimization model to improve the consistency of the transcription text and the voice content. Then, the optimized text is subjected to deep semantic analysis to automatically identify key information such as conference theme, topic, decision and task content, and generate a structured conference minutes document under the constraint of a template engine to realize efficient arrangement and automatic archiving of conference content. The application has the advantages that the accuracy and professional adaptability of voice transcription are improved, the homonym confusion, proper noun recognition error and word order ambiguity problem are effectively solved, and high-quality, structured and semantically coherent conference minutes generation is realized.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and intelligent office, specifically to an intelligent meeting minutes generation method based on text optimization and semantic relationship parsing, which belongs to the cross-application technology of speech recognition, text generation and intelligent task management. Background Technology

[0002] In modern enterprise and organizational office settings, the number and complexity of meetings are constantly increasing, involving multiple aspects such as cross-departmental communication, task allocation, and decision-making. Traditional manual meeting minutes and summary compilation methods suffer from inefficiency, error-proneness, and difficulty in information tracking, failing to meet the demands of modern offices for efficient, accurate, and traceable information management. Especially in the task supervision stage, manual summary compilation often struggles to promptly translate decisions, responsible parties, and deadlines into actionable tasks, resulting in low implementation rates of meeting decisions and incomplete information loops. Therefore, developing an intelligent meeting minutes generation and supervision system to automate the entire process from meeting audio recording to structured minutes generation, task identification, and supervision has become a key approach to improving enterprise information flow efficiency and management level.

[0003] Although existing speech recognition-based automatic transcription systems exist, they primarily remain at the "text transcription" level, lacking a deep understanding of the semantic structure of meetings. Existing technologies have the following shortcomings: First, the accuracy of speech recognition is low. In scenarios with multiple speakers, noisy environments, or many industry-specific terms, the error rate is high, especially in terms of homonym confusion and industry terminology recognition.

[0004] Second, the semantic understanding capability is insufficient. Existing systems often can only generate plain text records and cannot automatically extract structured information such as meeting topics, key decisions, and task items.

[0005] Third, there is a lack of a task supervision mechanism. Action items in the meeting (such as "a certain person is responsible for completing a certain task") cannot be effectively transformed into task lists or supervision items, making it difficult to manage meeting decisions in a closed loop.

[0006] Therefore, a method is needed that combines speech recognition optimization, semantic parsing, and task objectification generation to achieve intelligent generation of meeting minutes and automated task supervision. Summary of the Invention

[0007] To address the aforementioned issues, the present invention aims to provide an intelligent meeting minutes generation method based on text optimization and semantic relationship parsing. This method constructs a text optimization model based on speech recognition results and an automated minutes generation method based on semantic relationships, thereby achieving intelligent processing of the entire process from meeting recording to structured meeting minutes.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for generating meeting minutes that integrates text optimization and semantic relation parsing includes the following steps: Step 1: Speech Recognition: The meeting recording is transcribed in real time using an open-source speech recognition component to obtain a preliminary meeting text sequence. ; Step 2, Text Optimization: Utilize a text optimization model to optimize the preliminary meeting text sequence. A first correction is performed, followed by a second correction based on industry lexicon and contextual semantics. This corrects homonym errors, proper noun deviations, and word order ambiguities, generating an optimized corrected text. ; Step 3, Semantic parsing and information extraction: Predicting and correcting text For each word element, the entity category and entity relation are obtained, and the output is a set of triples. Each triple includes: the original text of the sentence, the entity category of each word element in the sentence, and the entity relation label between the word elements and other words. Step 4: Aggregate sentences with the same entity relationship, generate a set of similar relationships, and remove duplicate sentences from the set; Step 5: Establish a semantic chain according to the requirements. The semantic chain contains at least the relationship between two entities. Guided by the semantic chain, perform secondary aggregation on the set of similar relationships. Then, based on the format document built by the template engine, map the secondary aggregated corpus to generate meeting minutes text.

[0009] The present invention has the following beneficial effects: 1. This invention uses a speech recognition component to transcribe meeting recordings in real time. It employs soft masking technology to reconstruct the contextual semantics of the initial transcribed text, predicts and corrects low-confidence words, and constructs a Trie tree based on an industry-specific lexicon and corresponding pinyin to achieve both precise and fuzzy matching. It also intelligently corrects proper nouns and industry terminology. This improves the accuracy and professional adaptability of the transcribed text, effectively resolving issues such as homonym confusion, errors in proper noun recognition, and word order ambiguity.

[0010] 2. This invention utilizes a pre-trained text model to perform deep semantic analysis on the optimized text, automatically identifying key information such as meeting topics, issues, decisions, and task content, extracting entity relationships, performing secondary aggregation and redundancy removal on entity relationships to ensure comprehensiveness and conciseness of information, and finally combining semantic chains and abstract generation models to form a unified format of meeting minutes under the specification of a template engine, achieving high-quality, structured, and semantically coherent meeting minutes generation.

[0011] 3. This invention can intelligently identify meeting tasks and form a closed-loop management of responsible person-task-deadline, supporting task assignment and supervision reminders, improving the efficiency of meeting decision-making and execution, significantly reducing manual processing costs, and has high practical value and potential for widespread application.

[0012] 4. This invention possesses high-precision transcription, semantic understanding, and automatic supervision capabilities. Through systematic and automated meeting record and minutes management, it can effectively reduce human intervention and information omissions, achieve efficient implementation of meeting decisions and a closed loop for task supervision, and provide reliable data support and operational basis for enterprise management. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a flowchart illustrating the secondary correction process performed by combining an industry-specific thesaurus in this invention. Detailed Implementation

[0014] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Please see Figure 1 This embodiment provides a method for generating meeting minutes that integrates text optimization and semantic relationship parsing, including the following steps: Step 1: Speech Recognition. The meeting recording is transcribed in real-time using an open-source speech recognition component to obtain a preliminary meeting text sequence. The original confidence level associated with each lexical unit ti This includes the timestamp and speaker ID. This step provides the raw corpus foundation for subsequent semantic understanding.

[0015] Step 2, Text Optimization. The preliminary meeting text sequence is then optimized using a text optimization model. A first correction is performed, followed by a second correction based on industry lexicon and contextual semantics. This corrects homonym errors, proper noun deviations, and word order ambiguities, generating an optimized corrected text. Correcting text Significant improvements were made in semantic consistency and professional accuracy, providing high-quality corpus for subsequent summary generation. The specific process is as follows: Step 21: Perform standardization preprocessing on the output text to reduce the interference of noise and non-standard symbols on the subsequent semantic model. Split the text sequence T into sentence sets according to semantic boundaries. Si represents the meeting text, where extra spaces are removed and the formats of numbers, times, dates, and symbols are standardized.

[0016] Step 22, Context Concatenation: Receive standardized conference text statement S i ={t1,…t m}, m < n, and splice its context window to form the input sequence: X = {C -n , S, C +n}, where C -n and C +n represent the previous and next n sentences of context respectively; Step 23, Predict the confidence of the correctness of the token : Input the sequence X into the BERT encoder to obtain the hidden layer representation sequence H containing the sum of the word vectors, position vectors, and segment vectors of each token, ; where is the hidden layer representation of the sum of the word vector, position vector, and segment vector of each token; input the hidden layer representation sequence H into the bidirectional gated recurrent unit, combined with the original confidence , and output the confidence of the correctness of each token , with a range of [0, 1]. The closer it is to 0, the greater the probability of misrecognition, and vice versa, the smaller the probability of misrecognition. The calculation formula of is as follows: ; where, is the sigmoid function, is the weight matrix of the fully connected layer, is the hidden state of the last layer of the bidirectional gated recurrent unit, represents feature splicing, is the bias term; Step 24, Generate the soft masking vector : Input the hidden layer representation corresponding to each token in the hidden layer representation sequence H into the soft masking layer. The soft masking layer weights and fuses the hidden layer representation with the learnable masking vector to mask the untrustworthy information for subsequent prediction of the masked information in combination with the context; the processing method of the soft masking layer is as follows: , where, is the trainable masking vector, and the soft masking layer outputs the soft masking vector corresponding to each token. Since at high confidence, that is, is large, ( ) is small, the soft masking vector is mainly composed of ; on the contrary, the soft masking vector is mainly replaced by . Therefore, through the processing of the soft masking layer, the high-confidence original text information can be retained, and replacement at low confidence is allowed.

[0017] Step 25, Semantic reconstruction: Use the context information to re-predict the correct value of the low-confidence tokens, so as to obtain a corrected text with consistent semantics. The signal is fed into the decoding layer, where contextual information is used to predict the masked words, as shown below: ;in, Here, D represents the weights of the fully connected layer, and D is the decoder with the same structure as BERT but with independent parameters. This represents the probability distribution of candidate words at the current position. The word with the highest probability is the corrected character, thus completing one verification.

[0018] The training dataset for the text optimization model includes ASR outputs and human-corrected transcripts from real conference transcriptions, covering various industry sectors, multi-person dialogues, and noisy scenarios. The main loss for model training is minimizing the cross-entropy loss between the corrected text and the human-annotated reference. ; The semantic label is the position i in the correctly annotated text. The auxiliary loss is the language model perplexity difference regularization, which encourages the fluency of generated text sentences. Where PLM() is the perplexity score calculated by the pre-trained language model. For the corrected text, This is the correct text with manual annotation. These are the weighting coefficients. The model's total loss objective function is the sum of these two.

[0019] The secondary correction is applied to the prediction confidence level. Below the threshold The revised lexical Use a vocabulary list to replace proper nouns. For details, please refer to [link / reference]. Figure 2 : Step 26: Construct a trie tree for the industry terminology and a trie tree for the corresponding pinyin, based on the industry terminology and its corresponding pinyin, for storage and fast matching. Includes industry-specific terminology, company name, and product model.

[0020] Step 27: Filter out Below the threshold For each word s, an exact match is performed using a Trie tree. If a match is found, the string is considered correct and no replacement is needed. If no exact match is found, a fuzzy match is performed in both the Trie tree and the Pinyin Trie tree. The overall match score is calculated. If the overall match score exceeds a threshold, the word is replaced; otherwise, no replacement is performed. The process of fuzzy matching of the vocabulary is as follows: Calculate the normalized score of character edit distance: ;in, Let be the minimum edit distance of word s in the industry vocabulary E, and let e be the candidate word in E with the minimum edit distance to s. There may be multiple candidate words with the minimum edit distance, such as the edit distance set {5, 4, 3, 2, 2}, in which two candidate words have the minimum edit distance of 2. Let the candidate word set be... ; The process of fuzzy matching of Pinyin is as follows: Convert 's' to Pinyin sequence Then calculate the pinyin sequence. Industry Glossary Pinyin The edit distance is obtained The candidate pinyin set is ; Three candidate sets are defined based on the overlap relationship of edit distance: characters belong to And its pinyin belongs to The set of words is the first-level candidate set. Pinyin belongs to The character corresponding to that pinyin does not belong to The set of pinyin is a secondary candidate set. The character belongs to C, but the pinyin corresponding to that character does not. The set is a three-level candidate set. The overall matching score is defined based on the candidate set level: The final overall matching score for each word s is: ; in, and For the corresponding candidate words and pinyin, These are the weighting coefficients. If a score is achieved Exceeding the threshold If the match is successful, the entry will be replaced; otherwise, the original text will be retained. This formula means that if both the pinyin and the character are successfully matched (i.e., both are in the candidate set), the error is very small, and the candidate should be prioritized for replacement. If only the pinyin is successfully matched, it may indicate a speech recognition error, and the pinyin should be prioritized. If only the character is successfully matched, it may indicate an encoding deviation. The final corrected text after replacement is... .

[0021] Step 3: Semantic parsing and information extraction. The corrected text will be... The BERT model is used to extract features, and a lightweight classifier is used to predict the entity category and entity relationship for each word. Entity categories include meeting topic, agenda, decision, task content, responsible person, and deadline. Entity relationships are the relationship labels between entity category words, including task-responsible person, task-deadline, and agenda-decision. The output is a set of triples, each triple including: the original text of a sentence, the entity category of each word in that sentence, and the relationship label between each word and other words. This can be represented as a triple: ,in This represents the j-th sentence. This indicates the entity category of the words contained in the sentence. A set representing the relationships between lexical units.

[0022] Step 4: In the extractive compression stage, key information segments are aggregated and redundancy is eliminated. This specifically includes: Step 41: Based on the entity relationship tags, aggregate sentences according to semantic relationships, grouping sentences with the same entity relationship into similar sets, including: grouping sentences containing issue-decision relationships into a set. Sentences containing task-responsibility relationships are grouped into a set. Sentences containing task-deadline relationships are grouped into a set. Generate preliminary structured semantic units ; Step 42: Perform semantic deduplication within each set of relations: When the entity categories and relations of a sentence are completely contained in another sentence, only the sentence with the wider coverage of entity categories and relations is retained, generating representative semantic units. This eliminates information redundancy and repetitive expression.

[0023] Step 5: Association Generation and Structured Minutes Output. Based on the deredundant corpus, semantic aggregation and formatted generation are performed. Semantic chains are established according to requirements, each containing at least two entity relationships. Using these semantic chains as guidance, a secondary aggregation of similar relationships is performed. Then, based on a formatted document constructed using a template engine, the secondary aggregated corpus is mapped to generate meeting minutes text. Details are as follows: Step 51: Perform secondary aggregation on various relationship sets to establish key semantic chains, including issue-decision semantic chains and responsible person-task-deadline semantic chains: for sentences of the "issue-decision" type. This involves merging multiple decisions under the same issue to form a clear "issue-multiple decision" relationship group; and establishing a task-responsibility human sentence set. And a collection of sentences related to tasks and deadlines. Multiple tasks assigned to the same person are integrated to generate a relationship group "person in charge - task - deadline", forming a task list structure.

[0024] Step 52: Construct a unified format document based on an XML template engine, including meeting topic, participants, meeting agenda, meeting decisions, and task list. Map the secondary aggregated corpus to template fields one by one to achieve automatic mapping of semantic content to structured placeholders. For the participant field, count the speaker IDs corresponding to the voice text tags; for the task list, populate the integrated task list "responsible person – task – deadline" for easy distribution to the OA system, realizing intelligent linkage and closed-loop management of meeting minutes and task supervision. For the meeting topic, meeting agenda, and meeting decision content, use an abstract generation model to automatically generate meeting minutes text. Specifically: use the secondary aggregated "agenda – multiple decisions" relationship group and the corresponding statement set Mad' as input prompts, call the text generation model for natural language generation. Based on the semantic structure and contextual dependencies of the input relationship group, the model completes content reorganization and language generation under template constraints, automatically integrates multiple decision information under the same agenda, and generates semantically coherent and hierarchically clear meeting minutes text. The general text generation model described can employ GPT series, Tongyi Qianwen, etc., but it does not rely entirely on extracting sentence information from the original meeting text. Instead, it generates new minutes under the guidance of structured semantic templates. The model receives aggregated and deredundant structured semantic units: "issue-multi-decision" relationship groups, corresponding statement sets such as Mad', etc. These inputs are structured, non-natural language information fragments, while the model outputs coherent and fluent meeting minutes text. For example, the structured semantic units are... "Topic: The delayed progress of the 220kV XX substation expansion project" decision making: 1. The construction unit will increase manpower to ensure that civil engineering and safety are completed before June 30.

[0025] 2. The materials department urgently allocated the main transformer, which is expected to arrive on May 20.

[0026] 3. The supervision unit should strengthen on-site monitoring and submit weekly progress reports. The complete narrative, reorganized using a text generation model, is "Regarding the delay in the expansion project of the 220kV XX substation, the meeting clarified the following improvement measures: The construction unit must immediately increase personnel to ensure the completion of the key milestone of civil engineering delivery and installation before June 30; the materials department has coordinated the emergency allocation of the main transformer, which is expected to arrive on May 20 to ensure the subsequent installation progress; the supervision unit must strengthen on-site supervision and submit progress reports every week to ensure that problems are discovered and closed in a timely manner." Step 53: Send the task list from the meeting minutes to the OA system to achieve intelligent linkage and closed-loop management between the meeting minutes and task supervision.

[0027] This invention provides a meeting minutes generation method that integrates text optimization and semantic relation parsing. It obtains preliminary transcribed text through speech recognition, improves semantic accuracy using a text optimization model, then utilizes a pre-trained text model for semantic parsing and entity relation extraction, and aggregates and removes redundancy from key information segments to form high-information-density semantic units. Finally, a template engine and an abstract generation model map the structured corpus into standardized meeting minutes. This invention is an intelligent meeting minutes generation solution with high-precision transcription, semantic understanding, and automatic follow-up capabilities.

[0028] The above description is merely a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Images