Conference content context-aware question and answer retrieval method and system for continuous conference

Abstract

The present application relates to the technical field of electric data processing, more particularly, the present application relates to a conference content context-aware question and answer retrieval method and system for continuous conferences, comprising: acquiring multi-session conference text data of continuous conferences and user input retrieval requirements; performing vectorization processing on the conference text data and the retrieval requirements to obtain conference content vectors and retrieval vectors; and calculating semantic similarity of each session conference and the retrieval requirements.The present application constructs theme consistency drift index and time-effect correlation saturation, accurately quantifies evolution and turning point of conference theme and actual value of historical information, and on this basis, uses context correction coefficient to improve the gravitational search algorithm, so that the search process can dynamically adjust the gravity according to the context correlation strength, thereby effectively excluding obsolete information and achieving accurate and time-effective knowledge retrieval in long-span and complex logic continuous projects.

Description

Technical Field

[0001] This invention relates to the field of electrical data processing technology. More specifically, this invention relates to a context-aware question-answering retrieval method and system for continuous meetings. Background Technology

[0002] With the increasing prevalence of digital transformation and remote collaboration in enterprises, meeting information has become a crucial component of corporate knowledge assets. To efficiently utilize this information, question-and-answer retrieval technology for meeting content has emerged. This technology typically uses Automatic Speech Recognition (ASR) to convert meeting audio into text, and then combines it with Natural Language Processing (NLP) and Information Retrieval (IR) algorithms. This allows users to quickly locate the information they need from massive amounts of meeting records by asking questions in natural language, greatly improving the efficiency of information retrieval.

[0003] However, in many real-world work scenarios, meetings are not isolated events but rather continuous activities centered around a specific project or task, such as weekly project meetings or product iteration meetings. In these series of meetings spanning weeks or even months, the topics discussed evolve, deepen, and even shift as the project progresses. Early discussions may focus on conceptual design and planning, while later discussions focus on specific implementation details and problem reviews. Existing meeting question-and-answer retrieval systems, when processing such continuous meeting data, often treat all historical meeting records as a flat pool of information or simply use a time decay model to assess the importance of information. This approach ignores the dynamic evolution of meeting topics, leading to a problem in practical applications: when users ask questions about issues at the current stage, the system may retrieve outdated or subsequently overturned information due to the presence of numerous seemingly related keywords from earlier meetings. This could mislead users and even cause project decision-makers to make inappropriate judgments based on incorrect historical information. Summary of the Invention

[0004] This invention provides a context-aware question-and-answer retrieval method and system for continuous meetings, aiming to solve the problem in related technologies where, when a user asks a question about the current stage, the system may retrieve outdated information that has been overturned by subsequent decisions due to the presence of many seemingly related keywords in earlier meetings. This could mislead the user and even lead to project decision-makers making inappropriate judgments based on incorrect historical information.

[0005] In a first aspect, the present invention provides a context-aware question-answering retrieval method for continuous meetings, comprising: acquiring text data of multiple sessions of a continuous meeting and user-inputted retrieval requests; vectorizing the meeting text data and retrieval requests to obtain meeting content vectors and retrieval vectors; calculating the semantic similarity between each session and the retrieval requests, constructing a topic consistency drift index for each session to characterize the degree of shift in the topic of the session; constructing a time-related saturation for each session, wherein the time-related saturation reflects the degree of relevance between historical sessions and the current retrieval request, and also reflects the degree of shift in the topic of the session; constructing a context-aware gravity correction coefficient for each session, wherein the product of the context-aware gravity correction coefficient and the topic consistency drift index of the session is positively correlated with the time-related saturation of the session; using the context-aware gravity correction coefficient to correct the gravity constant of the gravity search algorithm, and using the corrected gravity search algorithm, with the retrieval vector as the search agent, retrieving target text fragments from the meeting content vector and generating answers. By constructing a topic consistency drift index and a timeliness-related saturation, we can accurately perceive the dynamic changes in conference topics and the actual support strength of historical conferences for current retrieval needs. Furthermore, by utilizing a context-aware gravity correction coefficient to improve the gravity search algorithm, the retrieval process no longer relies solely on keyword matching or simple time decay, but can dynamically adjust the convergence speed and direction of the search according to the tightness of the context. This effectively avoids retrieving outdated or refuted information, significantly improving the accuracy and timeliness of conference content retrieval in long-term projects or continuous task scenarios.

[0006] Furthermore, this includes defining a sliding window on a continuous sequence of meetings, obtaining the cosine similarity between the meeting content vector and the retrieval vector within the sliding window to obtain a semantic similarity sequence, and calculating the standard deviation of the semantic similarity sequence. By introducing a sliding window mechanism on a continuous sequence of meetings and calculating the standard deviation of the semantic similarity sequence, the fluctuations in the focus of meeting discussions within a local time window can be keenly captured, quantifying the discreteness of the semantic environment. This provides reliable data support for accurately determining whether the current situation is a turning point in the dramatic evolution of the topic, overcoming the problem of contextual fragmentation caused by viewing a single meeting in isolation.

[0007] Furthermore, a topic consistency drift index is constructed for each meeting session, including: the topic consistency drift index is positively correlated with the standard deviation of the semantic similarity sequence, and positively correlated with the difference between the maximum semantic similarity in the semantic similarity sequence and the semantic similarity of the meeting session. This method can identify potential logical jumps or topic shifts at nodes where the retrieval request has low semantic similarity to a particular meeting but the surrounding context fluctuates drastically. This allows for accurate differentiation between simply irrelevant information and key topic turning points, helping the system locate crucial historical moments that signify changes in project direction within a complex network of consecutive meetings.

[0008] Furthermore, a timeliness-related saturation level for each meeting is constructed, including calculating the time interval between the current meeting and any historical meeting. The timeliness-related saturation level is positively correlated with the product of the semantic similarity between historical meetings and the search query, and the topic consistency drift index of the corresponding meeting, and negatively correlated with the logarithm of the time interval between the current meeting and the corresponding historical meeting. A timeliness assessment method based on enhanced association is proposed, ensuring that high-value historical meetings, even those from a long time ago, containing key decisions or serving as topic turning points, retain high weight. This prevents crucial early decision information from being misjudged as low-value information over time, achieving a deep understanding of the project background spanning long periods.

[0009] Furthermore, the gravitational constant in the gravity search algorithm is modified, and the modified formula is as follows: In the formula, For the current algorithm, the first The improved gravitational constant after the second iteration; This is the preset total number of iterations; For the first Context-aware gravity correction coefficients for each meeting session; For the current iteration number At that time, the original The gravitational constant in the algorithm. This endows the gravitational search algorithm with dynamic adaptive capabilities. By introducing a sinusoidal periodic function and a context correction coefficient, the gravitational constant can be nonlinearly enhanced when the system perceives strong contextual correlations (such as continuous and consistent discussion intentions). This prompts the search agent to converge quickly and lock onto relevant clusters of meeting fragments, preventing the algorithm from prematurely converging to local optima and avoiding ineffective wandering in the sea of ​​weakly correlated historical information, thus improving retrieval efficiency.

[0010] Furthermore, acquiring text data from multiple sessions of a continuous meeting includes: collecting audio signals using a multi-microphone array in the conference room; converting the audio signals into timestamped meeting text data using automatic speech recognition technology; and obtaining precise start and end timestamps for each meeting through interaction with an office collaboration system. This combination of hardware acquisition and office software interaction ensures the alignment of the meeting audio data with the precise time dimension, guaranteeing the integrity and accuracy of the temporal sequence of the continuous meetings. This provides a high-quality, noise-free data foundation for subsequent time-series-based contextual analysis and avoids logical errors caused by inconsistent timestamps.

[0011] Furthermore, the retrieval vector is obtained, including: based on the user's input retrieval needs, extracting the core keyword set through a keyword extraction algorithm and converting it into a retrieval vector.

[0012] Furthermore, the meeting content vector is obtained by using the BERT model to vectorize the meeting text data.

[0013] Furthermore, the process of generating an answer includes applying automatic text summarization technology to the retrieved target text fragments to generate a logically coherent summary text as the final answer. This transforms the fragmented meeting notes into logically coherent and concise summaries, directly answering the user's question and avoiding the burden of reading large amounts of original meeting transcripts.

[0014] In a second aspect, a context-aware question-and-answer retrieval system for continuous meetings is also provided, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the context-aware question-and-answer retrieval method for continuous meetings as described in any of the above embodiments.

[0015] Beneficial effects: By constructing a topic consistency drift index and a timeliness relevance saturation, the evolution and turning points of conference topics and the actual value of historical information can be accurately quantified. Based on this, the gravity search algorithm is improved by using a context correction coefficient, which enables the search process to dynamically adjust gravity according to the strength of contextual relevance, thereby effectively eliminating outdated information and achieving accurate and timely knowledge retrieval in continuous projects with long spans and complex logic. Attached Figure Description

[0016] Figure 1 This is a schematic diagram illustrating a question-and-answer retrieval flowchart according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the perceived comparison of historical background association strength according to an embodiment of the present invention; Figure 3This is a schematic diagram illustrating the dynamic change of the topic consistency drift index of the t-th meeting according to an embodiment of the present invention. Detailed Implementation

[0017] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0018] like Figure 1 As shown, S101: Data acquisition and preprocessing.

[0019] In this embodiment, the raw audio signal of a continuous meeting is first acquired in real time using a multi-microphone array pickup device deployed in the conference room and the audio stream interface of the remote conferencing system. Then, automatic speech recognition (…) is used… The technology converts the acquired audio stream into raw text data with timestamps. Simultaneously, by interacting with an interface with an office collaboration system, it can obtain precise start and end timestamps for each meeting, ensuring the linear sequence integrity of consecutive meetings in the time domain.

[0020] The collected raw text data needs to undergo in-depth preprocessing. Specifically, firstly, word segmentation tools from natural language processing are used to segment the text, and a pre-defined stop word dictionary is applied to filter out colloquial and meaningless words in the meeting text, such as "um" and "that." Secondly, vector models are used, for example... or The model maps the text content of each meeting into a high-dimensional vector space, thus obtaining a meeting content vector that can represent its semantic information. Finally, the meeting data of all meetings are sorted according to the chronological order of the meetings, and a data matrix containing "meeting number - meeting content vector - timestamp" information is constructed, laying the data foundation for subsequent cross-meeting contextual association analysis and indicator construction.

[0021] S102: Construct the topic consistency drift index for each session.

[0022] In continuous meeting scenarios, topic drift inevitably occurs as projects progress or the focus of discussion shifts. According to the dynamic evolution theory in information theory, a long-term, continuous meeting sequence often follows the lifecycle of a specific task, leading to non-stationary fluctuations in the distribution of core keywords between different sessions. If the traditional gravity search algorithm is directly used... Its fixed gravitational decay model will be unable to perceive this shift in the core topic, causing the search subject to fall into an outdated historical context, affecting retrieval accuracy. Therefore, this embodiment constructs a topic consistency drift index to quantify the degree of logical offset between the current query requirement and the historical meeting sequence (e.g., Figure 3(Graph showing the dynamic change of the topic consistency drift index for the t-th meeting).

[0023] To construct this metric, we first use the user's current search query as a benchmark, and then use keyword extraction algorithms (such as...) Extract its core keyword set and convert it into a retrieval vector. Then, define a fixed-size sliding window on the continuous meeting sequence, calculate the cosine similarity between the meeting content vector and the retrieval vector of each meeting within the window, forming a semantic similarity sequence. The fixed-size sliding window is 5~7, indicating that the sliding window contains 5~7 consecutive meetings.

[0024] To quantitatively describe the first This embodiment constructs a topic consistency drift index to measure the degree of topic consistency drift across meetings. This indicator is constructed based on the fact that when the topic of a meeting changes drastically, the dispersion of the semantic similarity sequence increases significantly, leading to a larger standard deviation. This indicator can sensitively capture sudden shifts in context. Its calculation formula is as follows: In the formula, Indicates the first Thematic consistency drift index of meetings; As a preferred embodiment, the total number of meeting sessions within the sliding window is used. The reference value is Those skilled in the art can flexibly adjust this value according to the frequency of changes in the meeting topic; Indicates that in the first The standard deviation of the semantic similarity sequence of each session within a sliding window centered on the session; Indicates the first The semantic similarity between a meeting and the retrieval request can be calculated by the inner product of the meeting content vector and the retrieval vector. This represents the maximum value of the semantic similarity sequence within the sliding window; It is an exponential function of the natural constant.

[0025] As can be seen from the above formula, the more drastic the evolution of the meeting's topic and the greater the logical leaps, the more pronounced the fluctuations in the semantic similarity sequence within the window become, leading to an increase in the standard deviation. Increase. Also, if the current session similarity Relative to the peak value within the window If the exponent is lower, then the exponent term The value of will also increase. These two factors work together to increase the calculated topic consistency drift index. This indicates that the ... This meeting is likely to be a turning point in the theme, as this indicator provides a key basis for the subsequent dynamic correction of the gravitational constant.

[0026] S103: Construct the timeliness-related saturation of each meeting session.

[0027] Based on the analysis of topic drift, consecutive meeting data also exhibits complex characteristics of both time-related decay and enhanced association. According to the association evolution theory of knowledge graphs, although the weight of meetings older in time should naturally decrease, if a preceding meeting contains key decision-making nodes, such as determining the technical roadmap or budget allocation, its influence on all subsequent meetings will have a continuous guiding effect across sessions (e.g., Figure 2 (As shown). This characteristic means that the importance of conference data does not simply decay linearly in the time domain, but rather exhibits localized bursts of energy at key nodes.

[0028] Therefore, this embodiment utilizes the semantic similarity calculated in step S102. By calculating the weighted cumulative effect of the topic consistency drift index, a timeliness-related saturation is constructed to reflect the overall support strength of historical background for the current question and answer.

[0029] To construct this metric, the first step needs to be pre-calculated. This meeting and any previous meeting Time interval between meetings . No. Timeliness-related saturation of each meeting The construction is based on the fact that the contribution of a historical meeting is determined by its similarity to the query and its importance as a thematic turning point, and is expressed by a logarithmic function over time intervals. A smooth attenuation is achieved. The calculation formula is as follows: In the formula, Indicates the first The timeliness-related saturation of each meeting; For the first The semantic similarity between the meeting session and the retrieval request; For the first Thematic consistency drift index of meetings; For the first The meeting and the first The difference in timestamps between meetings.

[0030] This formula shows that when a certain meeting in history... It is highly similar to the current search requirements, that is When it is relatively large; and at a turning point where the theme changes, that is... The larger the value, the greater its contribution to the current context will be. As accumulation proceeds, the timeliness-related saturation can effectively accumulate the energy of all historical key nodes, thereby achieving a deep background perception across conference sessions.

[0031] S104: Construct context-aware gravity correction coefficients for each session.

[0032] It should be noted that there is also a context-locking phenomenon in continuous conference retrieval. That is, when a search intent is frequently and consistently reflected in multiple consecutive conferences, the relevant data will show extremely high semantic cohesion. This means that the algorithm should quickly converge and lock into a specific cluster composed of these conferences, rather than wandering in irrelevant historical information.

[0033] Based on the time-related saturation obtained in step S103, a context-aware gravity correction coefficient was further constructed, which is used to reflect the pressure of search space contraction.

[0034] Specifically, no. Context-aware gravity correction coefficients for each meeting The basis for its construction is that when the context is extremely strong, that is... Approaching historical maximum value When the theme evolves smoothly, that is When the value is relatively small, the suppression term in the denominator decreases, causing the context-aware gravity correction coefficient to increase non-linearly, thus enhancing gravity. Its calculation formula is as follows: In the formula, Indicates the first Context-aware gravity correction coefficients for each meeting session; For the first The timeliness-related saturation of each meeting; For the first Thematic consistency drift index of meetings; It represents the maximum value of time-related saturation that has occurred in all historical events up to the current moment, and its function is to serve as a dimensionless benchmark.

[0035] This formula reflects specific scenario characteristics: when the time-related saturation is close to its historical peak. And the current meeting Theme Consistency Drift Index When the value is small, the value of the context-aware gravity correction coefficient will increase sharply, thereby guiding the algorithm to avoid interference from irrelevant historical information and precisely control the search trajectory by directly adjusting the core gravity parameters.

[0036] S105: Improved gravity search algorithm and question answering retrieval implementation.

[0037] The first step calculated according to step S104 Context-aware gravity correction coefficients for each meeting Compared with traditional gravity search algorithms ( Core improvements will be made. In this embodiment, the original... The gravitational constant in the algorithm is reconstructed as a function that dynamically adapts to the iteration process and context.

[0038] Improved gravitational constant The basis for its construction is the introduction of a system composed of... The modulated sinusoidal periodic function allows the system to dynamically enhance its gravitational pull when it senses strong contextual correlation, preventing the algorithm from missing crucial information nodes due to premature convergence or insufficient momentum. Its calculation formula is as follows: In the formula, For the current algorithm, the first The improved gravitational constant after the second iteration; This is the preset total number of iterations; For the first Context-aware gravity correction coefficients for each meeting session; For the current iteration number At that time, the original The gravitational constant in the algorithm.

[0039] By introducing correction terms When the system determines that the current context is highly relevant, that is... The value is very large. The value will be significantly amplified during the iteration process. This dynamic adjustment mechanism can effectively enhance the attractiveness of the search subject, ensuring that it can be accurately and quickly attracted to relevant content nodes across multiple sessions.

[0040] In the final question-answering retrieval implementation stage, the meeting content vector model constructed in the preprocessing stage is used to map the user-input natural language questions into a target vector in a multi-dimensional feature space. This vector constitutes the search agent or search subject in the gravity search algorithm. During the algorithm iteration process, the improved gravitational constant is utilized. Calculate the attraction of each meeting segment (text block) to the search agent. Driven by a powerful gravitational pull, the search agent quickly converges on the cluster of meeting fragments that semantically best fit the current context. After the algorithm converges, it finally extracts the set of text fragments with the optimal position of the search agent and uses automatic text summarization technology to generate a logically coherent, accurate, and context-integrated answer, which is then presented to the user.

[0041] This invention also provides a context-aware question-and-answer retrieval system for continuous meetings. The system includes a processor and a memory, the memory storing computer program instructions. When the processor executes the computer program instructions, it implements the context-aware question-and-answer retrieval method for continuous meetings according to the first aspect of this invention.

[0042] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and therefore will not be described in detail here.

[0043] In this invention, the aforementioned memory can be any tangible medium containing or storing a program that can be used or combined with an instruction execution system, apparatus, or device. For example, a computer-readable storage medium can be any suitable magnetic or magneto-optical storage medium, such as Resistive Random Access Memory (RRAM), Dynamic Random Access Memory (DRAM), Static Random Access Memory (SRAM), Enhanced Dynamic Random Access Memory (EDRAM), High-Bandwidth Memory (HBM), Hybrid Memory Cube (HMC), etc., or any other medium that can be used to store desired information and can be accessed by an application, module, or both. Any such computer storage medium can be part of a device or accessible to or connected to a device. Any application or module described in this invention can be implemented using computer-readable / executable instructions stored or otherwise maintained on such a computer-readable medium.

[0044] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A context-aware question-answering retrieval method for meeting content in continuous meetings, characterized in that, include: Acquire text data from multiple sessions of a continuous conference and user-inputted search requests; vectorize the conference text data and search requests to obtain conference content vectors and search vectors; calculate the semantic similarity between each session and the search requests, and construct a topic consistency drift index for each session to characterize the degree of shift in the topic of each session; The timeliness relevance saturation of each meeting session is constructed. The timeliness relevance saturation reflects the degree of relevance of the current search needs in the historical meetings session, and also reflects the degree of change in the theme of the meeting session. Context-aware gravity correction coefficients are constructed for each session. The product of the context-aware gravity correction coefficients and the topic consistency drift index of the session is positively correlated with the timeliness relevance saturation of the session. The context-aware gravity correction coefficient is used to correct the gravity constant of the gravity search algorithm. Using the corrected gravity search algorithm, the target text fragment is retrieved from the meeting content vector using the retrieval vector as the search agent, and a response is generated.

2. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, Also includes: Define a sliding window on a continuous sequence of meetings, obtain the cosine similarity between the meeting content vector and the retrieval vector within the sliding window, obtain a semantic similarity sequence, and calculate the standard deviation of the semantic similarity sequence.

3. The context-aware question-answering retrieval method for continuous meetings according to claim 2, characterized in that, Construct a topic consistency drift index for each session, including: The topic consistency drift index is positively correlated with the standard deviation of the semantic similarity sequence and with the difference between the maximum semantic similarity in the semantic similarity sequence and the semantic similarity of the meeting.

4. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, Construct the timeliness-related saturation of each meeting session, including: The time interval between the current meeting and any meeting in history is calculated. The time-related saturation and the product of the semantic similarity between each meeting in history and the search query and the topic consistency drift index of the corresponding meeting are positively correlated and negatively correlated with the logarithm of the time interval between the current meeting and the corresponding historical meeting.

5. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, The gravitational constant of the gravity search algorithm is corrected using the following formula: ； In the formula, For the current algorithm, the first The improved gravitational constant after the second iteration; This is the preset total number of iterations; For the first Context-aware gravity correction coefficients for each meeting session; For the current iteration number At that time, the original The gravitational constant in the algorithm.

6. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, Obtain text data from multiple sessions of a continuous conference, including: Voice signals are collected using a multi-microphone array in the conference room, and automatic speech recognition technology is used to convert the voice signals into time-stamped meeting text data. By interacting with the office collaboration system, the precise start and end timestamp information of each meeting is obtained.

7. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, The retrieval vector is obtained, including: Based on the user's input search requirements, the core keyword set is extracted using a keyword extraction algorithm and then converted into a search vector.

8. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, The meeting content vector is obtained, including: The BERT model is used to vectorize the meeting text data to obtain content vectors.

9. The context-aware question-answering retrieval method for continuous meetings according to claim 1, characterized in that, Generate answers, including: Automatic text summarization technology is applied to the retrieved target text fragments to generate a logically coherent summary text as the final answer.

10. A context-aware question-answering retrieval system for continuous meetings, comprising a processor and a memory, characterized in that, The memory stores a computer program, and the processor executes the computer program to implement the context-aware question-answering retrieval method for continuous meetings as described in any one of claims 1-9.

Images