Method, device and storage medium for extracting keywords in video conference

By combining automatic word segmentation, keyword extraction, cosine similarity and semantic distance in video conferences, we can identify and play the voice of the interjector, and solve the problem of robbing and not answering questions in a timely manner, and achieve the maintenance of meeting order and time saving , improves the efficiency of live courses.

CN113505597BActive Publication Date: 2024-11-12SUIRUI TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110848123.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-07-27
Publication Date
2024-11-12
Estimated Expiration
2041-07-27

AI Technical Summary

Technical Problem

In video conferencing and live courses, snatching calls and failing to answer students' questions in a timely manner leads to confusion and waste of time, affecting the effectiveness of the meeting.

Method used

Using the method of combining automatic word segmentation, keyword extraction, cosine similarity and semantic distance, we identify the voice of the intercomer and the host, judge the voice of the intercomer, scientifically play the voice of the intercomer, and ensure the order of the meeting and effective Q&A.

Benefits of technology

Effectively prevent robbery, reasonably play the voice of the interjector, maintain meeting order, save teachers and students' time, and improve the efficiency and learning effect of live courses.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113505597B_ABST
    Figure CN113505597B_ABST
Patent Text Reader

Abstract

The present invention discloses a method, device and storage medium for extracting keywords in a video conference, belonging to the technical field of video multimedia communication, the method comprising: S1: acquiring the voice of the intervening speaker and the host, identifying the text in the voice, and segmenting the text; S2: extracting keywords from the segmented text; S3: extracting keywords from the voice of the intervening speaker and the host respectively, using the cosine similarity algorithm to calculate the similarity of the keywords, if the similarity is not less than a threshold, then proceeding to step S5; S4: extracting keywords from the voice of the intervening speaker and the host respectively, calculating the semantic distance of the extracted keywords, if the semantic distance is not less than a threshold, then proceeding to step S5; wherein steps S3 and S4 are performed simultaneously; S5: displaying the voice content of the intervening speaker. The present invention can standardize conference discipline and refine the classroom questioning process based on the combination of automatic segmentation, keyword extraction, semantic distance and cosine similarity.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of video multimedia communication, and in particular relates to a method, device and storage medium for extracting keywords in a video conference. Background Art

[0002] At present, in video conferences, due to disagreements and disputes, when there are people interrupting and talking at the same time, it will cause chaos and disrupt the order of the meeting. Another phenomenon is that during the teacher's live broadcast, when students have questions that they don't understand, the teacher will call on them or look at the questions one by one in the Q&A area at will. In this way, there is no time in the class to listen to all the questions and classify and answer them scientifically and reasonably. As a result, students who don't understand this knowledge point will also not understand the next knowledge point, causing the whole course to fall behind.

[0003] Currently, in video conferences, live courses mainly have the following problems:

[0004] During a meeting, it is not uncommon for other participants to interrupt the speaker before he or she has finished speaking. This not only causes chaos in the meeting, interrupts the participants' thinking, but also seriously affects the effectiveness of the meeting.

[0005] During the live broadcast, it is impossible to pay attention to the puzzled expressions of classmates online. This problem can be solved by collecting data through cameras. In addition, due to limited class time, the teacher cannot listen to or read all the questions before answering them. The following technology can judge the similarity and relevance of questions through cosine similarity and semantic distance to solve this problem encountered in live courses.

[0006] In view of this, the present invention is proposed. Summary of the invention

[0007] The purpose of the present invention is to provide a method, device and storage medium for extracting keywords in a video conference, which can regulate conference discipline and refine the classroom questioning process based on a combination of automatic word segmentation, keyword extraction, semantic distance and cosine similarity.

[0008] In order to achieve the above object, the present invention provides a method for extracting keywords in a video conference, comprising the following steps:

[0009] S1: Obtain the voice of the interrupter and the host, recognize the text in the voice, and segment the text;

[0010] S2: Extract keywords from the segmented text;

[0011] S3: extract keywords from the speech of the interrupter and the host respectively, use the cosine similarity algorithm to calculate the similarity of the keywords, and determine whether the similarity is less than the similarity threshold. If the similarity is not less than the similarity threshold, proceed to step S5; if the similarity is less than the similarity threshold, terminate this step;

[0012] S4: extract keywords from the speech of the interjector and the host respectively, calculate the semantic distance of the extracted keywords, and determine whether the semantic distance is less than the semantic distance threshold. If the semantic distance is not less than the semantic distance threshold, proceed to step S5. If the semantic distance is less than the semantic distance threshold, terminate this step. Steps S3 and S4 are performed simultaneously.

[0013] Among them, steps S3 and S4 are performed simultaneously

[0014] S5: Display the voice content of the interjector.

[0015] Furthermore, in step S1, the host's voice is selected as the voice of the interjector 30 seconds before the interjection node.

[0016] Furthermore, the step S1 also includes: the camera recognizes the facial expressions of the participants in the meeting to determine whether there is a change in facial expression, if so, proceed to the next step S2, otherwise, repeat step S1.

[0017] Furthermore, in step S2, part-of-speech tagging and TF-IDF weighting method are combined to extract keywords.

[0018] Furthermore, the step S2 comprises:

[0019] S201: Remove function words through part-of-speech tagging;

[0020] S202: Calculate the weight of the keyword using the TF-IDF weight method;

[0021] S203: Obtain the speech vectors of the interrupter and the host respectively.

[0022] Furthermore, the step S3 includes: extracting speech from the speech of the interrupter and the host respectively, obtaining vectors of their speech to calculate cosine similarity.

[0023] Furthermore, step S3 also includes: determining whether the keywords in the voices of the interrupter and the host contain synonyms or antonyms. If the two voices contain synonyms or antonyms, the frequency of the corresponding keywords in the frequency vector generated by the voices of the interrupter and the host is increased by 1.

[0024] Furthermore, the step S4 includes: arranging the keywords in the speech of the interjector and the host by parts of speech, comparing the semantic distances of the keywords with the same parts of speech, and determining whether the semantic distance is greater than a threshold;

[0025] The threshold is set according to the number of keywords in the speech.

[0026] The present invention also provides a device for extracting keywords in a video conference, which is connected to a conference terminal and includes: a storage module, a recognition module and a microprocessor;

[0027] The storage module is used to store the voices of the interjector and the host;

[0028] The recognition module is used to recognize the voice of the interjector, and when the voice of the interjector is recognized, it is sent to the microprocessor for processing;

[0029] The microprocessor is used to process the voices of the interrupter and the host using the above-mentioned method of extracting keywords in the video conference, and send the processing results to the conference terminal.

[0030] The present invention also provides a storage medium on which a computer program is stored, characterized in that when the computer program is executed by a processor, the steps of the method for extracting keywords in a video conference are implemented.

[0031] The present invention provides a method, device and storage medium for extracting keywords in a video conference, which have the following beneficial effects:

[0032] 1. Based on the trigger condition of two voices speaking at the same time during the meeting, the speaker's speech before this node for 30 seconds is converted into text, and the cosine similarity is compared. If the similarity is not less than 0.6 or the semantic distance is within 3, it is used as a judgment condition to scientifically and intelligently play the interrupter's voice during the pause.

[0033] 2. In live classes, students’ questions are screened and classified through cosine similarity and semantic distance, and classroom questions are simplified and refined, allowing teachers to answer questions selectively, saving time for teachers and students.

[0034] 3. You can configure your own processing plans for different usage scenarios, industries, etc., which is more practical.

[0035] 4. Combine automatic word segmentation, keyword extraction, cosine similarity and semantic distance, and use cosine similarity and semantic distance to analyze and determine whether the conditions for playing the interrupter's voice are met; when applied to live classrooms, it can track in real time and summarize and screen doubtful knowledge points in live courses, saving time for teachers and students. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 This is a flowchart of a method for extracting keywords in a video conference in this specific implementation.

[0037] Figure 2 It is a schematic diagram of a device for extracting keywords in a video conference in this specific implementation mode. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the scheme of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods.

[0039] like Figure 1 As shown, an implementation method of the present invention provides a method for extracting keywords in a video conference, which is triggered when the interrupter and the host speak at the same time, recognizes the interrupter's voice, and compares and identifies it with the host's voice to obtain the similarity and relevance of the content of the interruption. If the similarity and relevance match, the interrupter's voice content is played when the host pauses.

[0040] This method can not only prevent the interrupter from interrupting the speaker in time, but also play the interrupter's voice reasonably, maintain the meeting order, and provide intelligent guarantee for the smooth progress of the meeting. This technology is also applicable to dual-teacher classrooms. During the teacher's live broadcast, the camera collects the puzzled expression. After reminding the teacher, all questions can be collected by voice or text, and then the threshold is set by semantic distance to identify whether each question is closely related to this knowledge point and whether it is the content that needs to be mastered for this knowledge point. At the same time, cosine similarity comparison is performed according to different questions, so that the same or similar questions can be uniformly processed, so that the teacher can select the same or similar questions according to semantic distance and cosine similarity and classify them. This processing will omit repeated questions and arrange them according to the relevance of the questions, saving time for teachers and students.

[0041] The method comprises the following steps:

[0042] S1: Obtain the voice of the interrupter and the host, recognize the text in the voice, and automatically segment the text.

[0043] The host's voice is selected from the voice 30 seconds before the interruption node of the interrupter.

[0044] For the automatic word segmentation of the text in the meeting, the words in the sentence are segmented through the word list. The maximum number of words in the word list is usually set to 8. An example of automatic word segmentation is as follows: This equity incentive policy will take effect at the end of the month. The FMM segmentation result is: This / the / equity incentive / policy / will / take effect / at / the / end of the month / .

[0045] In addition, when this step S1 is applied to meeting scenarios such as online teaching, the following steps may also be included: perform facial expression recognition on the participants in the meeting through a camera, and determine whether there is a facial expression change (for example, a confused expression appears during online teaching). If so, proceed to the next step S2; otherwise, repeat this step.

[0046] Among them, for facial expression recognition, methods such as image acquisition, facial recognition, preprocessing, micro-expression discovery, feature extraction, and micro-expression classification are used to extract confused expressions.

[0047] S2: Extract keywords from the segmented text.

[0048] Keywords are extracted by combining part-of-speech tagging and the TF-IDF weighting method, which can extract keywords more accurately.

[0049] The specific steps of this step S2 are as follows:

[0050] S201: Remove function words through part-of-speech tagging to obtain effective feature vectors as candidate words. Among them, function words include stop words such as "de, shi, zai, le..." and functional words.

[0051] S202: Use the TF-IDF weighting method to calculate the weights of keywords. The calculation method is as follows:

[0052] The term frequency TF (Term Frequency) represents the probability that the word t appears in the speech d. The calculation formula is: TF = (the number of times the word t appears here) / (the total number of words in the content);

[0053] The calculation formula for the inverse document frequency index IDF (Inverse Document Frequency) is: IDF = log (the total number of speeches in the corpus / (the number of speeches containing the word t + 1)), which can be simplified to the formula: IDF = log (D / Dt), where D is the total number of speeches in the corpus and Dt is the number of speeches containing the word t plus 1;

[0054] The calculation formula for the weight value Weight of TF-IDF is: Weight = TF * IDF.

[0055] Among them, the corpus is a corpus of a considerable scale. There are many speech documents in this corpus. Some speech documents contain candidate words, and some speech documents do not contain candidate words. Therefore, calculate the total number of documents containing this candidate word, and then substitute "the total number of speech documents in the corpus" and "the total number of speech documents containing this candidate word" into the above formula to obtain the weight value Weight of TF-IDF.

[0056] S203: Get the vectors of the interrupter's and the host's voices respectively; after the above calculations, each keyword will correspond to a coordinate axis, and the value of the voice on each coordinate axis is the weight of the corresponding keyword. Therefore, the vector corresponding to the voice d is v(d) = {weight1, weight2, ... weightn}; usually set to sort by weight, the number of keywords with the largest weight is at most 5, that is, n is at most 5.

[0057] In addition, it can be applied to keyword generation for recorded courses or meeting minutes, that is, to accurately extract keywords by combining the word span algorithm, the distance between the first and last appearance of a word, and the principle that the greater the distance, the more important it is.

[0058] S3: Extract keywords from the speech of the interrupter and the host respectively, use the cosine similarity algorithm to calculate the similarity of the keywords, and determine whether the similarity is less than the similarity threshold. If the similarity is not less than the similarity threshold, proceed to step S5; if the similarity is less than the similarity threshold, terminate this step.

[0059] Through this step, the cosine similarity of the contents of the two voices can be identified, and the vector features can be identified by the cosine similarity, which specifically includes the following steps:

[0060] Extract voice a and b from the voice of the interrupter and the host respectively, and obtain their voice vectors, v(a) = {N1, N2, ...Nn}, v(b) = {N1, N2, ...Nn}, where N is the frequency (number of occurrences) of the corresponding keyword, and n is the number of keywords with the largest weight selected in step S2; then merge the two keywords into a collection, generate their respective frequency vectors, and calculate the cosine similarity cos(v(a), v (b)). The closer the value (cos(v(a), v(b))) is to 1, the higher the similarity. Judge according to the set similarity threshold. If the similarity exceeds the threshold, proceed to step S5 and play the interrupter's voice; if the similarity does not exceed the threshold, terminate this step.

[0061] In the present invention, the threshold of similarity is set to 0.6, that is, if the similarity of the keywords in the speech of the interrupter and the host is not less than 0.6, the keywords in the speech of the interrupter are retained, and if the similarity is less than 0.6, the step is abandoned.

[0062] At the same time, in this step, the synonym lookup table and the antonym lookup table can also be combined to make a judgment, that is, to judge whether the keywords in the voices of the interjector and the host contain synonyms or antonyms; if the two voices contain synonyms or antonyms, the word frequency of the keywords in the voices of the interjector and the host is increased by 1, that is, the word frequency of the corresponding keywords in the generated word frequency vector is increased by 1, otherwise, no operation is performed. The purpose of the above steps is not to judge whether the content or viewpoints of the two voices are completely consistent or similar, but to focus on whether the two voices are on the same topic.

[0063] An example of this step S3 is as follows:

[0064] The keyword frequency vector in speech 1 is: change (1) / contract (1) / re-sign (1) / supplement (0) / agreement (1); the keyword frequency vector in speech 2 is: change (1) / contract (1) / re-sign (0) / supplement (1) / agreement (1);

[0065] The conversion of the speech vector is: v(1) = {1, 1, 1, 0, 1}; v(2) = {1, 1, 0, 1, 1};

[0066] Afterwards, the cosine similarity cos(v(1), v(2))=0.828298 is calculated. Since the cosine similarity is not less than 0.6, it means that the voice of the interrupter is similar to the voice of the host, and step S5 can be performed to play the voice of the interrupter.

[0067] S4: extract keywords from the speech of the interrupter and the host respectively, calculate the semantic distance of the extracted keywords, and determine whether the semantic distance is less than the semantic distance threshold. If the semantic distance is not less than the semantic distance threshold, proceed to step S5; if the semantic distance is less than the semantic distance threshold, terminate this step; wherein, steps S3 and S4 are performed simultaneously.

[0068] For example, the similarity between "doctor" and "disease" is very low, but the correlation is very high; the correlation between cars and gasoline is very high, but the similarity is low, which shows that correlation and similarity are not the same. Word similarity reflects the aggregation characteristics between words, while word correlation reflects the combination characteristics between words.

[0069] Specifically, a semantic distance calculation method based on ontology concepts is adopted, which mainly applies the hyponymy relationship in the ontology. The ontology is an existing framework of semantic association. In the ontology, we can directly obtain the semantic distance between certain concepts, for example, the semantic distance between "animal" and "plant" is 2, and the semantic distance between "mammal" and "reptile" is 2.

[0070] This step is as follows: arrange the parts of speech according to the keywords in the voice of the interjector and the host, and compare the semantic distance of the keywords with the same part of speech, that is, compare the semantic distance of the two sentences by the semantic distance formula, and judge whether the semantic distance Dis(d1, d2) (where d1 and d2 are the keywords with the same part of speech in the voice of the interjector and the host, respectively) is greater than the threshold. If it is not greater than the threshold, proceed to step S5. If it is greater than the threshold, terminate this step. Different thresholds can be set here according to the number of keywords. For example, when the number of keywords with the largest weight is selected as 5, the threshold is set to 3.

[0071] The semantic distance formula is a prior art, which is directly obtained based on the shortest path of words in the semantic association hierarchy.

[0072] Among them, steps S3 and S4 are performed synchronously.

[0073] S5: Display the voice content of the interjector.

[0074] Specifically, after the judgment of steps S3 and S4, if any one of the judgments is successful, step S5 needs to be performed, which means that the voice of the interrupter is similar or related to the voice of the host. The voice of the interrupter can be played when the host pauses, otherwise, the voice of the interrupter will not be played.

[0075] An operation example of a method for extracting keywords in a video conference of the present invention is as follows:

[0076] Example 1:

[0077] In a meeting on equity distribution for core staff, the department manager said, "The equity incentive policy this time is that core staff who have worked in the company for more than five years can purchase shares at the original price of RMB 1 per share, and the shares can be traded on the market five years after the company goes public." When talking about market transactions, Wang, who was attending the meeting, asked, "What if the company cannot go public after five years? How to calculate the principal and interest?" At this time, the words of the interrupter Wang were recorded in the voice backup, and the cosine similarity of the speaker's voice 30 seconds before this time point was compared with the interrupter's voice. Since the similarity was less than 0.6, it did not meet the judgment conditions. At the same time, the semantic distance of the keywords was compared. The semantic distance was <= 3, which met the judgment conditions. At this time, the interrupter's voice was played when the department manager paused.

[0078] Example 2:

[0079] In the live course, the teacher's lecture content is detected by voice detection and keywords are extracted based on the voice content. The semantic distance is calculated based on the keywords. When the camera detects that 20% of the students are confused, the alarm will be triggered to remind the teacher that they do not understand this knowledge point. This time node is used as the trigger condition. The content 30 seconds before the detection is "The project manager is recruiting the required team members as planned, and suddenly received a notice from the company's human resources department that the company has just signed a 2-year labor-management agreement with Company A. From now on, all project personnel must be imported from this company. After the interview, the project manager found that Company A did not have personnel that fully met the requirements. The project manager should What should we do? Choose C, find available resources from the company and conduct training. We cannot not use Company A or recruit again. None of the other options are suitable. ", which proves that many students do not understand this question. The system alarms and reminds the teacher that 20% of the students do not understand this knowledge point. At this time, the teacher asks the students what they do not understand according to the prompt. At this time, the system can collect the students' questions through text or voice. Zhang San: "Can't Company A recruit employees who meet the requirements according to the original plan?"; Wang Er: "Can't the contract be changed?"; Li Si: "Can the supplementary agreement be re-signed?"; Zhao Wu: "Can't we recruit again according to the original requirements?"...

[0080] At this time, the content of the first 30 seconds of the puzzled expression is intercepted for automatic word segmentation, keywords are extracted, function words and stop words are removed, and the 5 values ​​with the largest weights are selected from large to small. The semantic distance with different questions is calculated, and the semantic distance is ranked as follows: Zhang San (2), Zhao Wu (2), Wang Er (4), Li Si (5). The questions are compared by cosine similarity, and the similarity between Zhang San and Zhao Wu is 0.8. The system then extracts and selects the simplest one between the two for display, that is, if it is voice playback, the one with shorter speaking time is selected, and if it is text display, the one with shorter text is selected, and then the questions of Wang Er and Li Si are listed. The teacher can selectively answer according to the relevance of the question, with the first choice being the one with a large proportion of repeated questions, and then the arrangement is based on the principle of semantic distance (that is, when there are repeated questions, they are sorted by proportion, and when there are no repeated questions, they are sorted by semantic distance). In this way, the relevance and similarity are used to extract and simplify the students' questions, saving time for teachers and students, and making the live classroom have better learning effects.

[0081] Based on the same inventive concept, Figure 2 As shown, in one embodiment of the present invention, a device for extracting keywords in a video conference is provided, which is used to be connected to a conference terminal and includes: a storage module 1, an identification module 2 and a microprocessor 3.

[0082] Storage module 1 is used to store the voices of the interjector and the host;

[0083] The recognition module 2 is used to recognize the voice of the interjector, and when the voice of the interjector is recognized, it is sent to the microprocessor 3 for processing;

[0084] The microprocessor 3 is used to process the voices of the interrupter and the host by using steps S1-S5 of the above-mentioned method for extracting keywords in a video conference, and send the processing results to the conference terminal.

[0085] Among them, the device for extracting keywords in a video conference of the present invention also includes a camera 4, which is used to recognize the expressions of participants in the meeting and determine whether a puzzled expression appears. If so, a signal is sent to the microprocessor 3 to adopt steps S1-S5 of the above-mentioned method for extracting keywords in a video conference to process the voices of the interrupter and the host.

[0086] Based on the same inventive concept, an embodiment further provides a non-transitory computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of a method for extracting keywords in a video conference as described in the above embodiment are implemented.

[0087] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0088] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0089] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0090] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0091] The invention concept is described in detail using specific examples herein, and the description of the above embodiments is only used to help understand the core idea of ​​the invention. It should be pointed out that any obvious modification, equivalent substitution or other improvement made by a person of ordinary skill in the art without departing from the invention concept should be included in the protection scope of the present invention.

Claims

1. A method for extracting keywords in a video conference, characterized in that: The following steps are involved: S1: Obtain the voice of the interrupter and the host, recognize the text in the voice, and segment the text; S2: Extract keywords from the segmented text; S3: extract keywords from the speech of the interrupter and the host respectively, use the cosine similarity algorithm to calculate the similarity of the keywords, and determine whether the similarity is less than the similarity threshold. If the similarity is not less than the similarity threshold, proceed to step S5; If the similarity is less than the similarity threshold, terminate this step; S4: extract keywords from the speech of the interrupter and the host respectively, calculate the semantic distance of the extracted keywords, and determine whether the semantic distance is less than the semantic distance threshold. If the semantic distance is not less than the semantic distance threshold, proceed to step S5. If the semantic distance is less than the semantic distance threshold, terminate this step. Wherein, steps S3 and S4 are performed simultaneously; S5: Display the voice content of the interjector; Wherein, the step S3 further includes: determining whether the keywords in the voices of the interjector and the host contain synonyms or antonyms, if the two voices contain synonyms or antonyms, then the word frequency of the corresponding keywords in the generated word frequency vector in the voices of the interjector and the host is increased by 1, otherwise, no operation is performed; The step S4 includes: arranging the keywords in the speech of the interrupter and the host according to their parts of speech, comparing the semantic distances of the keywords with the same parts of speech, and judging whether the semantic distance is greater than a threshold; wherein the semantic distance is obtained according to the shortest path obtained by the words in the semantic association hierarchical structure; and the threshold is set according to the number of keywords in the speech.

2. The method for extracting keywords in a video conference according to claim 1, characterized in that: In step S1, the host's voice is selected as the voice 30 seconds before the interruption node of the interrupter.

3. The method for extracting keywords in a video conference according to claim 1, characterized in that: The step S1 also includes: the camera recognizes the facial expressions of the participants in the meeting to determine whether there is a change in facial expression. If so, proceed to the next step S2; otherwise, repeat step S1.

4. The method for extracting keywords in a video conference according to claim 1, characterized in that: In step S2, part-of-speech tagging and TF-IDF weighting method are combined to extract keywords.

5. The method for extracting keywords in a video conference according to claim 4, characterized in that: The step S2 comprises: S201: Remove function words through part-of-speech tagging; S202: Calculate the weight of the keyword using the TF-IDF weight method; S203: Obtain the speech vectors of the interrupter and the host respectively.

6. The method for extracting keywords in a video conference according to claim 5, characterized in that: The step S3 comprises: extracting speech from the speech of the interrupter and the host respectively, obtaining the vectors of their speech, and calculating the cosine similarity.

7. A device for extracting keywords in a video conference, connected to a conference terminal, characterized in that: include: A storage module, an identification module and a microprocessor; The storage module is used to store the voices of the interjector and the host; The recognition module is used to recognize the voice of the interjector, and when the voice of the interjector is recognized, it is sent to the microprocessor for processing; The microprocessor is used to process the voices of the interrupter and the host using the method for extracting keywords in a video conference described in any one of claims 1 to 6, and send the processing results to the conference terminal.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for extracting keywords in a video conference according to any one of claims 1 to 6 are implemented.