A to-do task automatic extraction system based on meeting text intention analysis

Abstract

The application relates to the technical field of natural language processing, and discloses a to-be-done task automatic extraction system based on conference text intention analysis, which comprises the following steps: generating a semantic vector sequence; calculating an intention quantity and a saliency quantity; calculating a semantic bridging quantity; calculating a field migration quantity; screening a text segment; generating a time quantity; and merging the text segment, the time quantity and an evidence time period to generate a to-be-done record. Through a semantic vector generation method of fusing word-level coding and character-level coding and weighting according to confidence, the application can improve the robustness of semantic representation; through separate modeling of a to-be-done trigger intention and a field saliency degree, the application can filter background turns that do not contain a to-be-done intention; through combination of a topic change quantity and a semantic bridging quantity to constrain cross-turn field migration, the application can avoid field error association between different topics; and finally, the application generates a to-be-done record, which is convenient for subsequent conference minutes and task management.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to an automated task extraction system based on meeting text intent parsing. Background Technology

[0002] Meeting task extraction is an important application of natural language processing technology in enterprise office work. Its main goal is to automatically identify and extract task information to be performed from meeting transcripts, thereby improving meeting efficiency and the automation level of task management. Currently, the mainstream methods for meeting task extraction fall into two categories: one is based on keyword matching, which uses a pre-set dictionary of task trigger words and task-related entities to perform matching searches in the transcript; the other is based on single-sentence semantic classification, which uses a pre-trained language model to classify each sentence, determine whether it contains tasks, and extract the corresponding task fields.

[0003] While existing technologies perform well in routine meeting scenarios, their accuracy and completeness in cross-functional R&D design review meetings are significantly insufficient. These meetings typically involve technical discussions across multiple professional fields, and the expression of tasks is quite flexible, often involving omissions, coreferences, and cross-round references. Existing technologies are prone to misclassifying background discussions containing technical objects as tasks, and they also tend to overlook tasks with only brief commitments but lacking clear targets, resulting in numerous misclassifications and omissions in the extraction results. Furthermore, transcription errors and speaker separation errors generated by automatic speech recognition further degrade the extraction performance of existing technologies.

[0004] The root cause of these shortcomings lies in the fact that existing technologies all assume that all fields of a task are located within the same sentence or turn, failing to handle semantic associations and field migrations across turns. However, in cross-functional R&D design review meetings, the objects of tasks are typically discussed in detail before commitment, and the responsible parties and time constraints may be added by others after commitment, resulting in an asymmetrical distribution of task fields across turns. Existing technologies lack the ability to model this distribution and cannot effectively integrate task information scattered across multiple turns. Therefore, how to accurately and completely extract tasks from cross-functional R&D design review meetings and provide traceable supporting evidence has become a pressing technical problem in this field. Summary of the Invention

[0005] This invention provides an automated task extraction system based on meeting text intent parsing, which solves the technical problems mentioned in the background.

[0006] This invention provides an automated task extraction system based on meeting text intent parsing, comprising: The semantic vector sequence generation module collects meeting text, label sequence, start time sequence, end time sequence, and confidence sequence. It calculates the duration sequence from the start time sequence and end time sequence, and generates a semantic vector sequence by combining the meeting text, confidence sequence, and duration sequence. The feature calculation module calculates the intent quantity from the semantic vector sequence and calculates the saliency measure for the target category, which includes subject, action, object and time constraint. The semantic bridging quantity calculation module calculates the topic change quantity from adjacent semantic vectors, and then calculates the semantic bridging quantity based on the topic change quantity. The field migration calculation module calculates the offset and distribution width of the target category using saliency metrics and intent metrics, and combines intent metrics, saliency metrics, semantic bridging metrics, offset and distribution width values ​​to calculate the field migration. The text fragment filtering module extracts candidate text from the meeting text and tag sequence, calculates the compatibility value, and filters text fragments based on field migration amount and compatibility value; The time parameter generation module calculates the task combination quantity by combining the intent quantity, field migration quantity, and compatibility value, extracts the peak value of the task combination quantity as the trigger point sequence number, and generates time parameters from the text fragments corresponding to the time constraints. The to-do record generation module calculates the evidence weight value based on the field migration amount and compatibility value associated with the trigger point sequence number, and calculates the evidence time period by combining the evidence weight value, start time series and end time series, and merges text fragments, time parameters and evidence time periods to generate to-do records.

[0007] The beneficial effects of this invention are as follows: By integrating word-level encoding and character-level encoding and weighting semantic vectors by confidence, this invention improves the robustness of semantic representation and reduces the impact of transcription noise on subsequent analysis. By modeling the to-do trigger intent and field saliency separately, it effectively filters background turns that do not contain the to-do intent, while capturing to-do trigger sentences with only brief commitments, thereby reducing false positives and false negatives. By constraining cross-turn field migration with topic change and semantic bridging, it avoids erroneous field associations between different topics, thus improving the accuracy of field matching. By adaptively estimating field offset and distribution width, the system can adapt to the field distribution characteristics of different meetings, improving the system's versatility. By judging the complete to-do structure through task combination, it elevates to-do identification from single trigger word detection to complete structure judgment, further reducing the false positive rate. By weighting the evidence time period through field migration contribution, the to-do records have reviewable and replayable evidence, facilitating subsequent meeting minutes and task management. Attached Figure Description

[0008] Figure 1This is a schematic diagram of a module of an automated task extraction system based on meeting text intent parsing according to the present invention. Detailed Implementation

[0009] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, features described in some examples may be combined in other examples.

[0010] It should be noted that, unless otherwise defined, the technical or scientific terms used in one or more embodiments of the present invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in one or more embodiments of the present invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" indicate that the element or object preceding the term encompasses the elements or objects listed following the term and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0011] like Figure 1 As shown, an automated task extraction system based on meeting text intent parsing includes: The semantic vector sequence generation module is used to collect meeting text, label sequence, start time sequence, end time sequence and confidence sequence. It calculates the duration sequence from the start time sequence and end time sequence, and generates a semantic vector sequence by combining the meeting text, confidence sequence and duration sequence. The feature calculation module is used to calculate the intent quantity from the semantic vector sequence and to calculate the saliency measure for the target category, which includes subject, action, object and time constraint. The semantic bridging quantity calculation module is used to calculate the topic change quantity from adjacent semantic vectors and calculate the semantic bridging quantity based on the topic change quantity; The field migration calculation module is used to calculate the offset and distribution width values ​​of the target category from the saliency measure and the intent measure, and to calculate the field migration value by combining the intent measure, saliency measure, semantic bridging measure, offset and distribution width values. The text fragment filtering module is used to extract candidate texts from the conference text and tag sequence, calculate the compatibility value, and filter text fragments based on the field migration amount and compatibility value. The time parameter generation module is used to calculate the task combination quantity by combining the intent quantity, field migration quantity and compatibility value, extract the peak value of the task combination quantity as the trigger point number, and generate time parameters from the text fragments corresponding to the time constraints. The to-do record generation module is used to calculate the evidence weight value based on the field migration amount and compatibility value associated with the trigger point sequence number, and to calculate the evidence time period by combining the evidence weight value, start time series and end time series, and to merge text fragments, time parameters and evidence time periods to generate to-do records.

[0012] In one embodiment of the present invention, the specific calculation process of the semantic vector sequence generation module includes: Subtract the start time of the i-th turn from the end time of the i-th turn to obtain the duration of the i-th turn. Sum the durations of the N turns and divide by N to obtain the average turn duration during the conference. The formula for calculating the semantic vector of the i-th turn, based on the conference text, confidence level, and duration sequence, is as follows: in, This represents the semantic vector of the i-th turn. Indicates the turn number. This represents the confidence level for the i-th turn. This represents the text of the i-th turn. Let represent the word-level encoding vector of the i-th turn text. This represents the character-level encoded vector of the i-th turn text. This represents the duration of the i-th turn. Indicates the average duration of each turn in the meeting. Indicates the duration of the injection direction vector. This represents the log-normalized value of the duration of the i-th turn relative to the average turn duration during the conference.

[0013] It should be noted that the turn-taking text is a unit of text spoken by a speaker within a continuous period of time during a conference, which can be obtained by transcribing the conference recording using an automatic speech recognition system. The start time is the point in time when the turn begins, recorded synchronously by the automatic speech recognition system during transcription, in seconds. The end time is the point in time when the turn ends, also recorded synchronously by the automatic speech recognition system during transcription, in seconds. The confidence level represents the reliability of the transcribed text, output by the automatic speech recognition system, and ranges from 0 to 1; a higher value indicates a more reliable transcription result.

[0014] The duration is the turn duration calculated by subtracting the start time from the end time, representing the length of a single turn. The average turn duration within the meeting is the average of all turn durations, representing the average length of a turn in the current meeting. The word-level encoded vector is a vector representation of the turn text obtained by word-level encoding, used to characterize the word-level semantic information of the turn text. The character-level encoded vector is a vector representation of the turn text obtained by character-level encoding, used to characterize the character-level semantic information of the turn text.

[0015] The confidence complement is the value obtained by subtracting the confidence score from the unit value, used to characterize the unreliability of the transcription result. The duration injection direction vector is a vector used to embed turn-taking length features into the semantic space; it is a user-defined parameter, and its dimension is consistent with that of the word-level encoding vector and the character-level encoding vector. The semantic vector is a vector representation that integrates text content, transcription reliability, and turn-taking duration, used to comprehensively characterize the semantic features of a single turn. The semantic vector sequence is a set of semantic vectors arranged according to the turn number, used to characterize the semantic sequence features of the entire conference.

[0016] It should be noted that word-level encoding can be implemented using pre-trained language models such as BERT and RoBERTa. Specifically, the turn-taking text is input into the pre-trained language model, and the output of the first token of the last layer of the model is taken as the word-level encoding vector. For example, when using the BERT model, the input turn-taking text with a length of no more than 512 tokens will output a word-level encoding vector with a dimension of 768. For long turns-taking texts with a length of more than 512 tokens, a sliding window approach can be used for segmented encoding. The step size of the sliding window is preferably 256 tokens, that is, two adjacent windows overlap by 256 tokens to ensure the semantic continuity of long turns-taking texts. Then, the average value of the encoding results of each segment is taken to obtain the final word-level encoding vector.

[0017] It should be noted that character-level encoding can be implemented using models such as CNN (Convolutional Neural Network) and CharBERT (Character-level BERT pre-trained model). Specifically, when using a CNN model, each character in the turn-taking text is first mapped to a character embedding vector of fixed dimensions. Then, multiple convolutional kernels of different sizes are used to perform convolution operations on the character embedding sequence. Next, features are extracted through max pooling layers. Finally, the features extracted by different convolutional kernels are concatenated to obtain the character-level encoding vector. For example, using convolutional kernels of sizes 2, 3, and 4, with 256 kernels per kernel, and an input character embedding dimension of 128, the output character-level encoding vector has a dimension of 768. When using a CharBERT model, the turn-taking text is directly input into the pre-trained CharBERT model, and the output of the first token of the last layer of the model is taken as the character-level encoding vector.

[0018] It should be noted that the duration injection direction vector can be learned during training after random initialization. Specifically, during model training, the duration injection direction vector is initialized as a random vector following a normal distribution with a mean of 0 and a variance of 0.01. This value can effectively embed turn-taking length features without disrupting the original semantic vector distribution (for example, when the initial variance is 0.01, the magnitude of the duration injection direction vector is small, and its addition to the semantic vector will not significantly change the direction and magnitude of the original semantic vector, while still embedding the turn-taking length features into the semantic vector; if the initial variance is too large, such as 1, it will significantly change the distribution of the original semantic vector, leading to semantic representation distortion). Then, it is optimized together with other model parameters through the backpropagation algorithm. After training, the duration injection direction vector can effectively embed the turn-taking length features into the semantic vector.

[0019] It should be further explained that the above embodiments use both word-level encoding and character-level encoding, weighted by confidence level. The aim is to retain usable semantic information even when transcription errors exist, avoiding the direct discarding of low-confidence turns. In practical applications, automatic speech recognition systems frequently encounter transcription errors such as misspellings and omissions, especially in the presence of accents, background noise, or proper nouns. Word-level encoding has a strong ability to represent the semantics of complete words but is more sensitive to transcription errors. Character-level encoding is more robust to transcription errors, retaining some semantic information even when words are incorrectly transcribed. By weighting by confidence level, when the transcription confidence is high, more reliance is placed on word-level encoding; when the transcription confidence is low, more reliance is placed on character-level encoding, thereby improving the overall reliability of the semantic vector. For example, when "battery compartment buckle" in the turn text is transcribed as "battery compartment hook," word-level encoding may not accurately recognize its semantics, but character-level encoding can still capture the similarity between the two words, thus retaining relevant semantic information.

[0020] In one embodiment of the present invention, the specific calculation process of the feature quantity calculation module includes: Based on the semantic vector sequence, the formula for calculating the intent of the i-th turn is as follows: in, This represents the intent quantity for the i-th turn. Indicates the turn number. Represents the compression function. This represents the transpose of the intended projection vector. This represents the semantic vector of the i-th turn. Indicates the intended bias value; The formula for calculating the significance metric of the i-th turn in target category f, based on the semantic vector sequence and the target category, is as follows: in, This represents the significance measure of the i-th turn in the target category f. Indicates the turn number. Indicates the target category identifier. This represents a set of target categories, which includes the subject, action, object, and time constraint. This represents the transpose of the class projection vector corresponding to the target class f. This represents the semantic vector of the i-th turn. This represents the class bias value corresponding to the target class f.

[0021] It should be noted that the intent quantity represents the strength of the pending trigger intent carried in the turn-around text, with a value ranging from 0 to 1. A higher value indicates a greater likelihood that the turn-around contains a pending trigger intent. The intent projection vector is a vector used to project the semantic vector onto the pending trigger intent space; it is a user-defined parameter, and its dimension is the same as that of the semantic vector. The intent bias value is a bias parameter in intent calculation; it is a user-defined parameter and is a scalar value.

[0022] The target categories are four preset to-do fields: subject, action, object, and time constraint. The saliency metric measures the prominence of a target category field in the turn-through text, ranging from 0 to 1; a higher value indicates a greater likelihood that the turn-through contains the corresponding target category field. The category projection vector is a custom parameter used to project the semantic vector onto the corresponding target category space; its dimension matches that of the semantic vector, and each target category has an independent category projection vector. The category bias value is a custom parameter used in the calculation of the corresponding target category; it is a scalar value, and each target category has an independent category bias value. The compression function maps the linear score to a value between 0 and 1; specifically, the sigmoid function is used.

[0023] It should be noted that the intent projection vector, intent bias value, category projection vector, and category bias value were all trained using an annotated meeting task dataset. Specifically, an annotated dataset was first constructed by collecting over 1000 hours of transcribed text from cross-functional R&D design review meetings, and professional annotators annotated each turn. The annotation criteria were as follows: when a turn contains explicit request, commitment, instruction, or arrangement semantics, it is labeled as containing a task trigger intent; when a turn contains explicit person names, role names, or speaker tags related to task responsibilities, it is labeled as containing a subject field; when a turn contains explicit action verbs related to task execution, it is labeled as containing an action field; when a turn contains explicit technical objects, product components, or project content related to task processing, it is labeled as containing an object field; and when a turn contains explicit time expressions related to task deadlines, it is labeled as containing a time constraint field. Then, a supervised learning model is constructed, taking semantic vectors as input and intent quantity and four saliency metrics as output. The loss function is the arithmetic mean of the binary cross-entropy loss corresponding to the intent quantity and the four saliency metrics. The weight of each task is 1 to ensure that each task is optimized to the same extent during training. Finally, the model is trained using a labeled dataset, and parameters such as intent projection vector, intent bias value, class projection vector, and class bias value are optimized through the backpropagation algorithm until the model's performance on the validation set reaches the preset requirements (for example, the preset requirements could be an F1 score of 0.9 or higher for to-do trigger intent recognition and an average F1 score of 0.85 or higher for the four to-do field recognition on the validation set, which will not be elaborated here).

[0024] It should be further explained that the above embodiments model the pending task trigger intent and field saliency separately. This is to avoid misclassifying background talk turns containing task objects as pending tasks, and to prevent short sentences containing only promises but lacking objects from being directly output as incomplete tasks. In cross-functional R&D design review meetings, background talk turns often discuss various design objects, defects, and technical issues in detail, but these talk turns do not contain pending task trigger intents. If intent and fields are modeled together, the system can easily misclassify these background talk turns containing task objects as pending tasks. On the other hand, pending task trigger sentences often appear in very short forms, such as "I'll handle it" or "Add it to the next version." These talk turns contain clear pending task trigger intents but lack field information such as task objects. If intent and fields are modeled together, the system may ignore these pending task trigger sentences due to the lack of field information. By modeling them separately, the system can first identify all talk turns containing pending task trigger intents, and then extract relevant field information from these talk turns and the talk turns before and after them, thereby simultaneously solving the problems of misclassification and omission.

[0025] In one embodiment of the present invention, the specific calculation process of the semantic bridging quantity calculation module includes: The formula for calculating the topic change in the i-th turn based on adjacent semantic vectors is as follows: in, This represents the change in topic for the i-th turn relative to the (i-1)-th turn. Indicates the turn number. Let represent the inner product of the semantic vector of the i-th turn and the semantic vector of the (i-1)-th turn. Denotes the semantic vector norm of the i-th turn. Let represent the semantic vector norm of the (i-1)th turn; The formula for calculating the semantic bridging between candidate number k and source turn number i, based on the topic change amount, is as follows: in, This represents the semantic bridging quantity between candidate sequence number k and source turn sequence number i. Indicates the candidate sequence number. Indicates the source turn number. Indicates the smaller index between k and i. This indicates the larger index between k and i. This represents the talk round index from m+1 to n. This represents the change in topic from the q-th turn to the (q-1)-th turn.

[0026] It should be noted that the topic change factor represents the degree of semantic direction change between adjacent turns, ranging from 0 to 1. A higher value indicates a greater semantic difference between two adjacent turns, and a higher probability of a topic shift. The vector inner product is the result of the dot product of two semantic vectors, used to characterize the similarity between the two semantic vectors. The vector norm is the length of the semantic vector, used to characterize its magnitude. The adjacent similarity value is the cosine similarity between two adjacent semantic vectors, ranging from -1 to 1; a higher value indicates greater similarity between the two semantic vectors.

[0027] The candidate sequence number is the turn number of the candidate to-do trigger center, a preset parameter, and is an integer from 1 to the total number of turn times in the conference. The source turn number is the turn number from which the to-do field may originate, a preset parameter, and is an integer from 1 to the total number of turn times in the conference. The change sum is the sum of all topic changes between the candidate sequence number and the source turn number, used to characterize the cumulative degree of topic changes between the two turns. The semantic bridging quantity is the semantic strength between the candidate sequence number and the source turn number, which can still be mutually interpreted after crossing topic boundaries. The value ranges from 0 to 1, with higher values ​​indicating a stronger semantic connection between the two turns and a greater likelihood that they belong to the same discussion topic. Specifically, the vector norm is calculated using the L2 norm.

[0028] In the actual calculation process, starting from the second turn, the topic change between each turn and the previous turn is calculated sequentially. Then, for each pair of candidate sequence number and source turn number, the smaller and larger values ​​of the two numbers are determined. All topic changes within the range of the smaller value plus 1 to the larger value are added to obtain the change sum. Finally, the semantic bridging value is obtained by exponential operation on the negative of the change sum. For example, if the candidate sequence number is 10 and the source turn number is 5, then the smaller value is 5 and the larger value is 10. The topic changes from turn 6 to turn 10 need to be added to obtain the change sum before calculating the semantic bridging value.

[0029] It should be further explained that the above embodiments introduce topic change and semantic bridging to constrain cross-turn field migrations by semantic boundaries, avoiding the mechanical binding of distant background objects to the current commitment. In cross-functional R&D design review meetings, topics frequently change, and the content discussed between different topics is usually not directly related. If topic boundaries are not considered, the system may incorrectly bind objects discussed in previous topics to commitments in the current topic. For example, if the first half of the meeting discusses "motor design issues" and the second half discusses "battery design issues," the system may treat the motor as the task object of the battery-related commitment if topic boundaries are not considered. By introducing topic change and semantic bridging, when two turns cross multiple topic transitions, the semantic bridging will naturally decay, and the system will not forcibly bind the content of these two turns together, thereby improving the accuracy of field matching.

[0030] In one embodiment of the present invention, the specific calculation process of the field migration amount calculation module includes: Based on the saliency measure and the intent quantity, the formula for calculating the offset of the target category f is as follows: in, This represents the offset of the target category f. Indicates the target category identifier. Indicates the turn number. This represents the significance measure of the i-th turn in the target category f. This represents the intent quantity for the i-th turn. Indicates a significance level of zero. Indicates the intention to prevent zero values; Based on the significance metric, the formula for calculating the distribution width of the target class f is as follows: in, This represents the distribution width value of the target category f. Indicates the target category identifier. This represents the category center value of the target category f. Indicates the turn number. This represents the significance measure of the i-th turn in the target category f. Indicates a significance level of zero. Indicates the width to prevent zeroing; Based on intent, significance metric, semantic bridging, offset, and distribution width, the formula for calculating the field migration amount of target category f from source turn order number i to candidate order number k is as follows: in, This represents the field migration amount of the target category f from the source round number i to the candidate round number k. Indicates the candidate sequence number. Indicates the source turn number. Indicates the target category identifier. This indicates the intent quantity of the turn corresponding to candidate number k. This indicates the significance measure of the source turn number i under the target category f. This represents the semantic bridging quantity between candidate sequence number k and source turn sequence number i. This represents the offset of the target category f. This represents the distribution width value of the target category f.

[0031] It should be noted that the category center value is the saliency-weighted center of a target category in the conference turn-taking sequence, indicating the main distribution position of that target category field in the conference. The intent center value is the weighted center of the pending trigger intent in the conference turn-taking sequence, indicating the main distribution position of the pending trigger intent in the conference. The offset is the displacement of the target category center value relative to the intent center value; a negative value indicates that the target category field usually appears before the pending trigger intent, and a positive value indicates that the target category field usually appears after the pending trigger intent.

[0032] The saliency zero-prevention value is the smallest positive number that prevents the sum of field saliencies from being zero; it is a user-defined parameter. The intent zero-prevention value is the smallest positive number that prevents the sum of intent values ​​from being zero; it is a user-defined parameter. The distribution width value is the range of possible rounds for a target category field, representing the maximum number of rounds that the target category field may be distributed before and after the pending trigger intent. The width zero-prevention value is the smallest positive number that prevents the denominator from being zero when calculating the distribution width; it is a user-defined parameter. The field migration amount is the comprehensive weight of a target category field's migration from the source round to the candidate trigger round; a higher value indicates a greater contribution of the corresponding target category field in the source round to the current candidate trigger round.

[0033] Specifically, the preferred value ranges for saliency nullification, intention nullification, and width nullification are as follows: arrive These zero-prevention values ​​are used to prevent the denominator from being zero during calculations, thus ensuring the stability of the calculations. In practical applications, these zero-prevention values ​​are usually set to... This effectively prevents the denominator from being zero without significantly affecting the calculation result.

[0034] In the actual calculation process, for each target category, its category center value and offset are calculated; then, the distribution width value of each target category is calculated; for each pair of candidate sequence numbers and source turn sequence numbers, the field migration amount of each target category is calculated. For example, for the object category, its offset is usually negative, indicating that the object field usually appears before the pending trigger intent; the distribution width value is usually 5 to 10, indicating that the object field may appear within 5 to 10 turn cycles before the pending trigger intent.

[0035] It should be further explained that the above embodiments estimate field offsets and distribution widths from the current meeting's own data to avoid pre-setting manual rules and adaptively model the field distribution characteristics in different meetings. The distribution characteristics of pending fields can vary significantly between different types of meetings, and even between different meetings of the same type. For example, in cross-functional R&D design review meetings, task objects are usually discussed in detail before commitments, resulting in a large negative offset for the object field; while in routine work meetings, task objects and commitments often appear in the same sentence, resulting in an object field offset close to zero. If pre-set manual rules are used, such as stipulating that the object field can only appear within the first three rounds of conversation before a commitment, the system's performance will fluctuate greatly in different meetings. By estimating field offsets and distribution widths from the current meeting's own data, the system can adaptively adapt to the field distribution characteristics of different meetings, thereby improving the system's versatility and adaptability.

[0036] In one embodiment of the present invention, the specific calculation process of the text fragment filtering module includes: Candidate texts are extracted from the meeting text and tag sequence. The candidate texts are divided into subject candidate texts, action candidate texts, object candidate texts and time constraint candidate texts according to the target category. Based on the field migration amount and compatibility value, the formula for calculating the text fragments with candidate serial number k under target category f is as follows: in, This represents the text segment obtained by filtering candidate number k under target category f. Indicates the candidate sequence number. Indicates the target category identifier. Indicates candidate text, This represents the set of candidate texts corresponding to target category f in the i-th turn. This represents the union of the candidate text sets corresponding to target category f in each turn of the conversation. This represents the field migration amount of the target category f from the source round number i to the candidate round number k. This represents the compatibility value of candidate text s in the i-th turn as a text fragment of the target category f.

[0037] It should be noted that candidate text refers to text fragments extracted from meeting text and tag sequences that may serve as pending fields. Subject candidate text refers to text fragments that may represent responsible parties, including names, roles, and speaker tags. Action candidate text refers to text fragments that may represent actions to be performed, including verb phrases. Object candidate text refers to text fragments that may represent task objects, including noun phrases. Time constraint candidate text refers to text fragments that may represent time constraints, including time-related expressions.

[0038] The compatibility score is the degree of semantic matching between the candidate text and the target category text fragment. The value ranges from 0 to 1, with higher values ​​indicating a more suitable candidate text for the corresponding target category. The candidate text score is the overall score of the candidate text as a target category text fragment, weighted by the field migration amount and the compatibility score. The text fragment is the final text content corresponding to the target category obtained through filtering.

[0039] It should be noted that the compatibility value can be obtained by calculating the semantic similarity between the candidate text and the target category using a pre-trained language model. Specifically, the candidate text and the target category name are input into the pre-trained language model to obtain two semantic vectors, and then the cosine similarity between the two semantic vectors is calculated as the compatibility value. For example, for the candidate text "backend manager" and the target category "subject," their semantic similarity is calculated as the compatibility value of the candidate text as the subject field.

[0040] It should be noted that candidate text extraction can be achieved using natural language processing techniques such as named entity recognition and phrase segmentation. Specifically, name text and time-related text can be extracted using named entity recognition models, with commonly used models including BERT-NER (BERT-based named entity recognition) and RoBERTa-NER (RoBERTa-based named entity recognition). Role text and speaker tag text can be directly extracted from the speaker tags in the conference transcript. Verb phrase text and noun phrase text can be extracted using phrase segmentation models. The segmentation criteria for verb phrases are continuous text segments containing a core verb and its direct object or complement, such as "optimize interface performance" and "modify design document." The segmentation criteria for noun phrases are continuous text segments containing a core noun and its modifiers, such as "user login interface" and "battery compartment snap structure." Commonly used phrase segmentation models include phrase segmentation models based on dependency parsing and phrase segmentation models based on pre-trained language models.

[0041] In the actual screening process, for each candidate trigger turn and each target category, firstly, candidate texts for that target category are collected from all turns, then the score of each candidate text is calculated, and finally, the candidate text with the highest score is selected as the text fragment for that target category.

[0042] It should be further explained that the above embodiments filter text fragments based on field migration amount and compatibility value. The purpose is to transform abstract semantic weights into concrete and readable pending field content. Field migration amount represents the contribution of a certain target category field in a certain turn to the current candidate trigger turn, and cannot be directly used as the content of the pending record. Compatibility value represents the degree of semantic matching between a candidate text and the corresponding target category, reflecting the suitability of the candidate text as a pending field. By multiplying the field migration amount and compatibility value to obtain the candidate text score, both the contribution of the turn to the current pending task and the semantic matching degree of the candidate text itself are considered, thereby enabling the selection of the most accurate and suitable text fragments as the content of the pending field.

[0043] In one embodiment of the present invention, the specific calculation process of the time parameter generation module includes: Based on the intent quantity, field migration quantity, and compatibility value, the formula for calculating the task combination quantity corresponding to candidate number k is as follows: in, This represents the number of task combinations corresponding to candidate number k. Indicates the candidate sequence number. This indicates the intent quantity of the turn corresponding to candidate number k. Indicates the target category identifier. This represents a set of target categories, which includes the subject, action, object, and time constraint. Indicates the product to prevent it from becoming zero. This represents the field migration amount of the target category f from the source round number i to the candidate round number k. This represents the text segment in the i-th turn. As the compatibility value of the text fragment in the target category f; The formula for calculating the set of trigger point numbers based on the task combination quantity is as follows: in, Represents the set of trigger point numbers. Indicates the trigger point number. This indicates the number of task combinations corresponding to trigger point number k. This indicates the number of task combinations corresponding to the preceding sequence number adjacent to the trigger point number k. This indicates the number of task combinations corresponding to the next sequence number adjacent to the trigger point k; Based on the text fragment corresponding to the time constraint, the formula for calculating the time parameter is as follows: in, This indicates the time parameter corresponding to the trigger point number k. Indicates the trigger point number. This represents the time normalization function. This indicates the text segment obtained by filtering under the time constraint category based on trigger point number k. This indicates the start time of the first round of the meeting.

[0044] It should be noted that the category combination value is the contribution of a target category field to the formation of a complete to-do structure, and is obtained by weighting the field migration amount and compatibility value of that target category field. The product zero-prevention value is an extremely small positive number to prevent the field quality product from being zero, and is a user-defined parameter. The task combination quantity is a comprehensive measure of the formation of a complete to-do structure when a talk round acts as a to-do trigger center; the higher the value, the more likely the talk round is to become an effective to-do trigger center.

[0045] The trigger point number is the turn number corresponding to the peak of the task combination volume, i.e., the candidate pending trigger center. The time parameter is the normalized deadline or time constraint representation, which is a time representation in a unified format. The time normalization function is a function that maps the relative time expressions in the meeting to a unified time reference system. The start time of the first turn of the meeting is the start time of the first turn of the meeting, used as a time reference.

[0046] It should be noted that the preferred range of values ​​for the product to prevent it from becoming zero is as follows: arrive The purpose of the product zero-prevention value is to prevent the entire task combination from being zero when the contribution value of a certain target category field is zero. In practical applications, the product zero-prevention value is usually set to... This effectively prevents the product from being zero without significantly affecting the calculation results of the task combination quantity.

[0047] It should be noted that the time normalization function can be implemented by identifying keywords in time expressions and combining them with meeting calendar information. Specifically, a keyword dictionary for time expressions is first constructed, including keywords representing absolute time, such as year, month, day, hour, and minute, as well as keywords representing relative time, such as "today," "tomorrow," "next week," "before the next review," and "before the freeze." Then, keyword recognition is performed on the candidate texts for time constraints to extract time information. For absolute time expressions, they are directly converted to a unified date and time format; for relative time expressions, they are converted to absolute time by combining the meeting date and preset meeting schedule information. For example, if the meeting is held on May 16, 2026, and the time expression is "next week," it is converted to before May 23, 2026; if the time expression is "before the next review," and the preset review period is two weeks, it is converted to before May 30, 2026. When there is no preset meeting schedule information, for relative time expressions such as "before the next review" and "before the freeze," which cannot be converted to absolute time, the time parameter retains the original natural language expression and is not forcibly converted. If no explicit time information can be identified, the time parameter is set to "unspecified."

[0048] In the actual calculation process, the task combination quantity for each candidate trigger turn is first calculated. The minimum threshold for the task combination quantity is preferably set to 0.5. When the task combination quantity of a turn is lower than this threshold, even if the peak condition is met, it is not used as a trigger point number to filter out low-confidence candidates. Then, all turns are traversed to find the peak position of the task combination quantity as the trigger point number. The criterion for determining the peak position is that the task combination quantity of this turn is greater than the task combination quantity of the previous turn and greater than or equal to the task combination quantity of the next turn. Finally, for each trigger point number, its corresponding time parameter is calculated.

[0049] It should be further explained that the above embodiments determine the formation of a complete to-do structure by assessing the task combination quantity. The aim is to elevate to-do identification from single trigger word detection to complete structure assessment, avoiding misjudgments and omissions. Traditional to-do extraction methods are typically based on trigger word detection; any turn containing a to-do trigger word, such as "please," "need," or "should," is identified as a to-do task. This method has significant limitations. Many turn-by-turns containing trigger words do not form a complete to-do structure. For example, "We need to discuss this issue" contains the trigger word "need," but lacks a clear responsible party and specific task content, and should not be identified as a to-do task. On the other hand, many to-do tasks do not contain obvious trigger words. For example, "Then I'll handle it" lacks traditional trigger words, yet it is a clear to-do task. By determining the formation of a complete to-do structure through task combination quantity, the system no longer relies solely on trigger words but comprehensively considers the trigger intent and the contributions of the four types of fields. Only when these elements form a stable and complete structure can it be identified as a to-do task, thus effectively avoiding misjudgments and omissions.

[0050] In one embodiment of the present invention, the specific calculation process of the to-do record generation module includes: Based on the field migration amount and compatibility value associated with the trigger point sequence number, the formula for calculating the evidence weight value corresponding to the trigger point sequence number k and the source call sequence number i is as follows: in, This represents the evidence weight value corresponding to the trigger point sequence number k and the source turn sequence number i. Indicates the trigger point number. Indicates the source turn number. Indicates the target category identifier. This represents a set of target categories, which includes the subject, action, object, and time constraint. This represents the amount of field migration for target category f from source round number i to trigger point number k. This represents the text segment in the i-th turn. As the compatibility value of the text fragment in the target category f; Based on the evidence weight value, start time series, and end time series, the formula for calculating the evidence time period corresponding to the trigger point number k is as follows: in, This indicates the start time of the evidence corresponding to the trigger point number k. This indicates the end time of the evidence corresponding to the trigger point number k. Indicates the trigger point number. Indicates the source turn number. This represents the evidence weight value corresponding to the trigger point sequence number k and the source turn sequence number i. This indicates the start time of the i-th turn. This indicates the end time of the i-th turn. Indicates that the evidence is protected against zero values; Based on the text fragment, time parameter, and evidence time period, the calculation formula for generating the pending record corresponding to the trigger point sequence number k is as follows: in, This indicates the pending record corresponding to trigger point number k. Indicates the trigger point number. Represents the main field, This indicates the text fragment obtained by filtering under the main category based on the trigger point index k. Indicates the action field. This represents the text fragment obtained by filtering under the action category based on the trigger point index k. Represents an object field. This indicates the text fragment obtained by filtering under the object category when the trigger point number k is selected. This indicates a time constraint field. This indicates the time parameter corresponding to the trigger point number k. This indicates the time period field for evidence. Indicates the start time of evidence collection. Indicates the end time of evidence collection. This represents the set of trigger point numbers.

[0051] It should be noted that the evidence weight value is the comprehensive evidence contribution of a particular turn to the formation of a particular task record, obtained by weighting the field migration amount and compatibility values ​​of the four target category fields. The evidence start time is the start time of the evidence segment of the task record on the meeting timeline, in seconds. The evidence end time is the end time of the evidence segment of the task record on the meeting timeline, in seconds. The evidence zero-prevention value is an extremely small positive number to prevent the sum of evidence weights from being zero; it is a user-defined parameter.

[0052] The to-do list is a structured record that includes the subject, action, object, time constraint, and evidence time period. The evidence time period is the source segment of the to-do list on the meeting timeline, consisting of the evidence start time and the evidence end time.

[0053] Specifically, the preferred range for the evidence zero-value is as follows: arrive The purpose of the evidence zero-prevention value is to prevent the denominator from being zero when calculating the start and end times of evidence, thus ensuring the stability of the calculation. In practical applications, the evidence zero-prevention value is usually set to... .

[0054] In the actual calculation process, the evidence weight value is first calculated for each trigger point sequence number and each source turn sequence number. When a certain type of field is missing, such as when no explicit time constraint is identified, the migration amount and compatibility value corresponding to that field are not included in the calculation of the evidence weight value. The evidence weight value is only calculated by summing the migration amount and compatibility values ​​corresponding to the existing fields. Then, the evidence start time and evidence end time are calculated by weighting the evidence weight values. Specifically, the start time of each source turn is multiplied by the corresponding evidence weight value, summed, and then divided by the sum of all evidence weight values ​​and the evidence zero-prevention value to obtain the evidence start time. The end time of each source turn is multiplied by the corresponding evidence weight value, summed, and then divided by the sum of all evidence weight values ​​and the evidence zero-prevention value to obtain the evidence end time. Finally, the subject, action, object, time constraint, and evidence time period are combined into a structured to-do record. The to-do record is output in JSON format, with each field being a string type.

[0055] For example, if the evidence for a certain task record mainly comes from turns 10 to 15, with the evidence from turn 12 having the highest weight, then the start time of the evidence will be close to the start time of turn 10, and the end time of the evidence will be close to the end time of turn 15, thus accurately reflecting the main time range of the task discussion.

[0056] It should be noted that the generated to-do records include clearly defined evidence time periods, which facilitates review by meeting minutes, task systems, and manual review. Users can quickly find the corresponding meeting recording segments based on the evidence time periods to verify the accuracy of the to-do content. Meeting minutes takers can quickly organize the relevant meeting content based on the evidence time periods. The task management system can link the to-do records with the corresponding meeting segments for easy follow-up tracking and management.

[0057] Specifically, this invention can be deployed in an enterprise's internal meeting management system and integrated with an automatic speech recognition system and a task management system. During the meeting, the automatic speech recognition system transcribes the meeting recording in real time, generating transcribed text with speaker tags, timestamps, and confidence levels. After the meeting, the transcribed text is automatically input into the system of this invention for processing. The system executes the seven modules of this invention sequentially and finally outputs a structured list of to-do tasks.

[0058] The final output list of to-do tasks contains multiple records, each with five fields: subject, action, object, time constraint, and evidence time period. For example, a cross-functional R&D design review meeting is held on May 16, 2026, lasting 120 minutes, with a transcribed text containing 320 turns. After processing, the system generates 5 to-do records. One record has the subject "Backend Manager," the action "Optimize Interface Performance," the object "User Login Interface," the time constraint "Before May 23, 2026," and the evidence time period "from the 2720th second to the 2895th second after the meeting starts." Users can click on the evidence time period to directly jump to the corresponding position in the meeting recording for playback and verification, confirming the accuracy of the to-do content. Once confirmed, the to-do record can be automatically synchronized to the enterprise task management system, assigned to the corresponding manager, and a deadline set, achieving fully automated management of meeting tasks.

[0059] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain a formula that is closest to the real situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0060] The content of this embodiment has been described above, but this embodiment is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this embodiment, all of which are within the protection scope of this embodiment.

Claims

1. A to-do task automation extraction system based on meeting text intent parsing, characterized by, include: The semantic vector sequence generation module collects meeting text, label sequence, start time sequence, end time sequence, and confidence sequence. It calculates the duration sequence from the start time sequence and end time sequence, and generates a semantic vector sequence by combining the meeting text, confidence sequence, and duration sequence. The feature calculation module calculates the intent quantity from the semantic vector sequence and calculates the saliency measure for the target category, which includes subject, action, object and time constraint. The semantic bridging quantity calculation module calculates the topic change quantity from adjacent semantic vectors, and then calculates the semantic bridging quantity based on the topic change quantity. The field migration calculation module calculates the offset and distribution width of the target category using saliency metrics and intent metrics, and combines intent metrics, saliency metrics, semantic bridging metrics, offset and distribution width values ​​to calculate the field migration. The text fragment filtering module extracts candidate text from the meeting text and tag sequence, calculates the compatibility value, and filters text fragments based on field migration amount and compatibility value; The time parameter generation module calculates the task combination quantity by combining the intent quantity, field migration quantity, and compatibility value, extracts the peak value of the task combination quantity as the trigger point sequence number, and generates time parameters from the text fragments corresponding to the time constraints. The to-do record generation module calculates the evidence weight value based on the field migration amount and compatibility value associated with the trigger point sequence number, and calculates the evidence time period by combining the evidence weight value, start time series and end time series, and merges text fragments, time parameters and evidence time periods to generate to-do records. 2.The meeting text intent resolution based to-do task automatic extraction system of claim 1, wherein, The following steps are performed when generating a sequence of semantic vectors: Retrieve turn texts from the conference text by turn number, and obtain the start time, end time, and confidence level of each turn text from the start time sequence, end time sequence, and confidence level sequence; Subtract the matching start time from each end time to obtain a duration sequence. Sum the duration values ​​in the duration sequence and divide the sum by the number of turns to obtain the average turn duration in the conference. Each round of text is encoded at the word level to obtain a word-level encoding vector, and each round of text is encoded at the character level to obtain a character-level encoding vector. Multiply the confidence score by the word-level encoding vector, subtract the confidence score from the unit value to obtain the confidence score complement, multiply the confidence score complement by the character-level encoding vector, add one to the ratio of the duration value to the average turn duration in the conference and then perform a logarithmic operation, multiply the logarithmic value by the duration injection direction vector, and then sum the three products to obtain the semantic vector. Arrange the semantic vectors according to the turn number to obtain the semantic vector sequence. 3.The meeting text intent resolution based to-do task automatic extraction system of claim 1, wherein, When calculating intent and significance metrics, perform the following steps: Obtain each semantic vector from the sequence of semantic vectors; Perform an inner product operation on each semantic vector and the intent projection vector, and add the inner product value to the intent bias value before inputting it into the compression function to obtain the intent quantity corresponding to each turn text; The target categories are defined as the subject, action, object, and time constraint. For each target category, the semantic vector is multiplied by the category projection vector corresponding to the target category, and the product of the product and the category bias value corresponding to the target category is added to the compression function to obtain the saliency measure of each turn text under each target category.

4. The system of claim 1, wherein, The following steps are performed when calculating topic changes and semantic bridging: Calculate the dot product of two adjacent semantic vectors in the semantic vector sequence, and calculate the vector norm of each adjacent semantic vector. The ratio of the numerator to the denominator is calculated by taking the inner product of the vectors as the numerator and the product of the norms of the two vectors as the denominator. The adjacent similarity value is obtained by subtracting the adjacent similarity value from the unit value. The topic change is obtained by subtracting the adjacent similarity value from the unit value. For any candidate sequence number and any source call sequence number, determine the smaller and larger sequence numbers; Sum the changes in topics from the next position after the smaller sequence number to the larger sequence number to obtain the sum of changes; Perform an exponential operation on the opposite of the change and value to obtain the semantic bridging quantity between the candidate sequence number and the source turn sequence number.

5. The system of claim 1, wherein, When calculating field migration amounts, perform the following steps: For each target category, sum the products of each turn number and the corresponding significance metric, and divide the sum by the sum of the significance metric sum and the significance zero-value to obtain the category center value; Sum the products of each turn number and its corresponding intent quantity, and divide the sum by the sum of the intent quantity and the intent zero-prevention value to obtain the intent center value; Subtract the intent center value from the category center value to obtain the offset of the target category; Subtract the category center value from each round number, square the result, multiply the squared value by the corresponding significance metric, sum the products, divide the sum by the sum of the significance metric and the significance zero-prevention value, take the square root of the quotient and add the width zero-prevention value to obtain the distribution width value. The field migration amount is obtained by multiplying the intent quantity corresponding to the candidate sequence number, the saliency measure corresponding to the source talk round number, the semantic bridging quantity, and the exponential weight formed by the offset and distribution width value.

6. The automated task extraction system based on meeting text intent parsing according to claim 1, characterized in that, Perform the following steps when filtering text snippets: Extract candidate text from meeting text and tag sequences; Candidate texts are divided into subject candidate texts, action candidate texts, object candidate texts, and time constraint candidate texts. Subject candidate texts include name texts, role texts, and tag texts; action candidate texts include verb phrase texts; object candidate texts include noun phrase texts; and time constraint candidate texts include time expression texts. For each source turn number, each target category, and each candidate text, calculate the compatibility value of the candidate text as a text fragment of the target category; Multiply the field migration amount corresponding to each source turn number by the compatibility value, and sum the products of each source turn number to obtain the candidate text score; The candidate text with the highest score is taken as the text fragment.

7. The automated task extraction system based on meeting text intent parsing according to claim 1, characterized in that, When calculating task combination quantities, trigger point numbers, and time parameters, the following steps are performed: For each candidate sequence number and each target category, the compatibility value corresponding to the text fragment is multiplied by the field migration amount, and the product of each source turn number is summed. Then, the summation value is added to the product zero value to obtain the category combination value. Multiply the category combination values ​​of the four target categories—subject, action, object, and time constraint—and multiply the product result by the intent quantity corresponding to the candidate sequence number to obtain the task combination quantity; The candidate number whose task combination quantity is greater than the task combination quantity of the adjacent first number and not less than the task combination quantity of the adjacent second number is used as the trigger point number. Input the text fragment corresponding to the time constraint and the start time of the first round of the meeting into the time normalization function to obtain the time parameter.

8. The automated task extraction system based on meeting text intent parsing according to claim 1, characterized in that, When generating a to-do list, perform the following steps: For each trigger point number and each source turn number, the field migration amount under the four target categories of subject, action, object and time constraint is multiplied by the compatibility value of the corresponding text fragment, and the four products are added together to obtain the evidence weight value. Sum the products of each evidence weight value and its corresponding start time, and divide the sum by the sum of the evidence weight values ​​and the evidence zero-prevention value to obtain the evidence start time; Sum the products of each evidence weight value and its corresponding end time, and divide the sum by the sum of the evidence weight values ​​and the evidence zero-prevention value to obtain the evidence end time; The text fragments corresponding to the subject, actions, and objects, along with time parameters, evidence start time, and evidence end time, are merged to generate a to-do record.

Images