A leaderless group discussion system

Abstract

The application discloses a leaderless group discussion system, which converts captured speech into text, carries out word segmentation on the text, filters out stop words in the speech, retains words of specified parts of speech, obtains a filtered word segmentation set, uses the word segmentation in the filtered word segmentation set to construct a vertex set, constructs edges between any two points in the vertex set using a co-occurrence relationship, obtains a candidate keyword graph of the filtered word segmentation set, calculates the weight of each vertex set in the candidate keyword graph, sorts the weights of the candidate keyword graph from large to small, selects N vertexes with the largest weight values as keywords of the speech, and thus helps interviewers or examiners quickly understand whether the speech of an interviewee deviates from a theme by extracting keywords, thereby improving the interview ability of the interviewee or the interview efficiency of the interviewer.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of interview, in particular to a leaderless group discussion system. BACKGROUND

[0002] Leaderless group discussion is a kind of evaluation technology often used in evaluation center technology, which uses the way of scene simulation to interview examinees in groups. A leaderless group is formed by a certain number of examinees (6-9 people), and they discuss work-related problems for about an hour. During the discussion, no one is designated as the leader, and no one is designated to sit in a certain position. The examinees arrange the organization by themselves, and the evaluators observe the examinees' ability to organize and coordinate, the ability to express orally, the ability to persuade in debate, and other aspects of ability and quality to see whether they meet the requirements of the post to be appointed, and whether the individual characteristics such as self-confidence, enterprising spirit, emotional stability, and reaction flexibility are in line with the group atmosphere of the post to be appointed, so as to comprehensively evaluate the differences among the examinees.

[0003] When conducting a leaderless discussion, the interviewee often deviates from the topic without realizing it due to nervousness, which leads to failure in the interview. On the other hand, the examiner needs to carefully listen to each interviewee's speech to evaluate whether the interviewee meets the topic, which consumes a lot of time and energy of the examiner.

[0004] Therefore, how to help the interviewee or the examiner quickly understand whether the interviewee's speech deviates from the topic has become a technical problem that technicians in this field need to solve urgently. SUMMARY

[0005] The present application provides a leaderless group discussion system to solve the technical problem that the leaderless group discussion system cannot help the interviewee or the examiner quickly understand the interviewee's speech.

[0006] To solve the above technical problems, the technical solution provided by the present application is:

[0007] A leaderless group discussion system, comprising: a voice capturing module, an audio conversion module and a keyword extraction module, the voice capturing module, the audio conversion module and the keyword extraction module are connected in sequence;

[0008] The voice capturing module is used to capture the speech of the speaker in the form of voice;

[0009] The audio conversion module is used to convert the captured speech in the form of voice into text form;

[0010] The keyword extraction module is configured to: perform word segmentation on the speech in the text form, filter out stop words in the speech during the word segmentation, retain words of a specified part of speech, obtain a filtered word segmentation set, construct a vertex set using the word segmentation in the filtered word segmentation set, construct edges between any two points in the vertex set using a co-occurrence relationship, obtain a candidate keyword graph of the filtered word segmentation set, calculate a weight of each vertex set in the candidate keyword graph, sort the weights of the candidate keyword graph from large to small, and select N vertexes with the largest weights as keywords of the speech.

[0011] Preferably, the keyword extraction module comprises:

[0012] The cleaning unit is configured to identify special characters in the speech in the text form and delete the special characters in the speech in the text form.

[0013] The word segmentation unit is configured to load a dictionary, establish a Trie tree word segmentation module, perform DGA word segmentation, construct a word segmentation DGA graph, calculate a global probability of the word segmentation DGA graph, obtain a word frequency maximum segmentation combination based on a prefix dictionary, identify English and Chinese in the word segmentation DGA graph using Token for a non-registered word, identify English, numerical values and time forms in English and give corresponding labels, load a hidden Markov (HMM) probability model graph for Chinese words, and obtain word segmentation and labels using a Viterbi algorithm dynamic programming to obtain a filtered word segmentation set.

[0014] The extraction unit is configured to construct a vertex set using the word segmentation in the filtered word segmentation set, construct edges between any two points in the vertex set using a co-occurrence relationship, obtain a candidate keyword graph of the filtered word segmentation set, calculate a weight of each vertex set in the candidate keyword graph, sort the weights of the candidate keyword graph from large to small, and select N vertexes with the largest weights as keywords of the speech.

[0015] Preferably, the extraction unit calculates the weight of each vertex set in the candidate keyword graph by the following formula:

[0016]

[0017] In the formula, i and j are serial numbers of two different word segmentation vertexes, S(V i ) is the weight of the i-th word segmentation vertex, d is a damping coefficient, In(V i ) is a word segmentation vertex set pointing to the i-th word segmentation vertex, |Out(V j ) | is a word segmentation vertex set pointed to by the j-th word segmentation vertex, and S(V j ) is the weight of the j-th word segmentation vertex.

[0018] Preferably, the application further comprises a sensitive word extraction module connected with the audio conversion module, for extracting the speech in text form according to a pre-constructed sensitive word list.

[0019] Preferably, the sensitive word extraction module comprises:

[0020] a matching unit: constructing a Trie tree of the speech in text form, constructing a Fail pointer for a joint on the Trie tree, starting from the root node of the Trie tree, moving along the AC automatic machine downward according to the read-in character each time; when the read-in character does not exist in the branch, recursively walking the Fail pointer path; if the Fail pointer path is walked to the root node, skipping the character and processing the next character, after reading all the input text, finally recursively walking the failure path until reaching the root node, completing the matching;

[0021] a screening unit: screening the successfully matched word segmentation as the sensitive word according to the matching result of the matching unit.

[0022] Preferably, the screening manner of the screening unit adopts any one of the following:

[0023] a specific sensitive word list screening manner: extracting the existing sensitive word from the word segmentation set of the speech in text form through the specific sensitive word list;

[0024] a fuzzy word list screening manner: taking a preset word as the center, shielding the preset word and the words before and after it, so as to achieve the effect of sensitive word shielding and extraction.

[0025] Preferably, the matching unit is further used for dividing the speech in text form according to the punctuation marks to obtain a plurality of end texts before matching, and then matching the plurality of end texts.

[0026] Preferably, the application further comprises a feedback module for sending the extracted key word / sensitive word to the user.

[0027] The application has the following beneficial effects:

[0028] 1. The unmoderated group discussion system in the application, through converting the captured speech form speech into text form speech, carrying out word segmentation processing on the text form speech, and filtering out stop words in the speech and retaining words of specified parts of speech in the word segmentation processing process, a filtered word segmentation set is obtained; the word segmentation in the filtered word segmentation set is used to construct a vertex set, and an edge between any two points in the vertex set is constructed using a co-occurrence relationship, a candidate keyword graph of the filtered word segmentation set is obtained, the weight of each vertex set in the candidate keyword graph is calculated, and the weights of the candidate keyword graph are sorted from large to small, and the top N vertices in the weight value size are selected as keywords of the speech, so that the extraction of keywords is used to assist the interviewee or the examiner to quickly understand whether the speech of the interviewee deviates from the theme, thereby assisting to improve the interview ability of the interviewee or the interview efficiency of the examiner.

[0029] 2. In the preferred scheme, the application sends sensitive words that should not appear in the interview process to the interviewee or the examiner by extracting the sensitive words from the interview speech, so as to assist the interviewee or the examiner to quickly understand whether the speech of the interviewee is appropriate, thereby assisting to improve the interview ability of the interviewee or the interview efficiency of the examiner.

[0030] In addition to the purposes, features and advantages described above, the application has other purposes, features and advantages. The application will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0031] The accompanying drawings, which form a part of the present application, are included to provide a further understanding of the application, and are incorporated herein for purposes of illustrating the illustrative embodiments of the present application and the explanations provided herein. In the drawings:

[0032] Figure 1 A text speech schematic diagram provided for an embodiment of the application;

[0033] Figure 2 A keyword schematic diagram extracted from a text speech schematic diagram provided for an embodiment of the application;

[0034] Figure 3 jieba word segmentation algorithm flow provided by an embodiment of the application;

[0035] Figure 4 Structure diagram of the unmoderated group discussion system provided by an embodiment of the application. DETAILED DESCRIPTION

[0036] The embodiments of the application will be described in detail below with reference to the accompanying drawings, but the application can be implemented in various different ways limited and covered by the claims.

[0037] Embodiment one:

[0038] The embodiment discloses a non-leader group discussion system, comprising: a voice capturing module, an audio conversion module and a keyword extraction module, which are sequentially connected;

[0039] The voice capturing module is used for capturing speeches of speakers in the form of voice;

[0040] The audio conversion module is used for converting the captured speeches in the form of voice into speeches in the form of text;

[0041] The keyword extraction module is used for carrying out word segmentation processing on the speeches in the form of text, filtering out stop words in the speeches during the word segmentation processing, retaining words of specified parts of speech, obtaining a filtered word segmentation set, constructing a vertex set using the word segmentation in the filtered word segmentation set, constructing edges between any two points of the vertex set using a co-occurrence relationship, obtaining a candidate keyword graph of the filtered word segmentation set, calculating the weight of each vertex set in the candidate keyword graph, sorting the weights of the candidate keyword graph from large to small, and selecting N vertexes with the largest weight values as keywords of the speeches.

[0042] The non-leader group discussion system in the embodiment converts the captured speeches in the form of voice into speeches in the form of text, carries out word segmentation processing on the speeches in the form of text, filters out stop words in the speeches during the word segmentation processing, retains words of specified parts of speech, obtains a filtered word segmentation set, constructs a vertex set using the word segmentation in the filtered word segmentation set, constructs edges between any two points of the vertex set using a co-occurrence relationship, obtains a candidate keyword graph of the filtered word segmentation set, calculates the weight of each vertex set in the candidate keyword graph, sorts the weights of the candidate keyword graph from large to small, and selects N vertexes with the largest weight values as keywords of the speeches, so that the keywords are extracted to assist the interviewee or the examiner in quickly understanding whether the speeches of the interviewee deviate from the theme, thereby assisting in improving the interview ability of the interviewee or the interview efficiency of the examiner.

[0043] Embodiment two

[0044] Embodiment two is a preferred embodiment of embodiment one, and is different from embodiment one in that the specific structure of the non-leader group discussion system is expanded.

[0045] In this embodiment, as shown in Figure 4 A non-leader group discussion system is disclosed, comprising: a microphone, a camera, a master control end; wherein the master control end comprises an audio conversion module, a keyword extraction module and a sensitive word extraction module;

[0046] The student clicks the start speaking button to start the microphone to record, and clicks the end speaking button to transmit the audio and video file to the host end. The host end uses ffmpeg to separate the audio file, and then analyzes it.

[0047] The audio-to-text part uses the related interface of Tencent cloud voice recognition to directly convert the audio file into corresponding text, and then extracts the keywords in the keyword extraction module and extracts the sensitive words in the sensitive word extraction module.

[0048] The keyword extraction includes:

[0049] As shown in Figure 1 and Figure 2 In the non-leadership group discussion system, the keyword refers to the core vocabulary in the student's speech content, which can express the student's current discovered point of view.

[0050] The current mainstream keyword extraction method is to use deep learning related technology, input the text data into the neural network, and obtain the keywords through the network calculation. However, this method needs a large amount of text data for training, and the effect cannot be guaranteed. Therefore, in the implementation process, the TextRank algorithm is used for keyword extraction. First, jieba is used to divide the text, and stop words are filtered out, only keeping words with specified parts of speech, such as nouns, verbs, and adjectives, Figure 3 The specific process of text segmentation by jieba algorithm is shown.

[0051] According to Figure 1 After obtaining the segmentation result by the algorithm process, the vertex set is constructed for these words, and the candidate keyword graph G=(V,E) is constructed, where V is the node set composed of the segmentation result, and then the co-occurrence relationship is used to construct the edge between any two points. The edge exists between two nodes only when the corresponding words co-occur in a window of length K. K represents the window size, i.e. at most K words co-occur. Finally, the weight of each word is calculated according to the following formula;

[0052]

[0053] S(Vi) represents the word node extracted from the text information.

[0054] d is the damping coefficient, generally set to 0.85.

[0055] In(Vi) represents the node set pointing to the word node Vi.

[0056] Out(Vj) is the set of web pages pointed to by the links in web page j. |Out(Vj)| represents the node set pointed to by the word node Vj.

[0057] Finally, the top 5-10 keywords are selected based on their weight. The system records these keywords and provides real-time feedback to students and the control panel, which can be viewed on the page. Students can use the displayed keywords to determine if their remarks are off-topic.

[0058] The improvement of this invention over the prior art lies in:

[0059] 1. In the calculation process, to avoid certain parts of speech affecting the results, a syntax filter was added to filter out words of a certain type or several types as the vertex set, which can improve the accuracy of keyword extraction.

[0060] 2. A pause word list has been added to better construct the vertex set, which can further improve the accuracy of keyword extraction.

[0061] Sensitive word extraction:

[0062] In this project, "sensitive words" refers to words that should not appear during interviews. Sensitive word extraction, simply put, involves traversing the text based on a sensitive word list; if any words from the list appear, they are recorded. However, this simple implementation has a time complexity of O(n*m*k), where n is the text length, m is the number of keywords, and k is the average keyword length, which is impractical. Two optimization methods exist: First, calculate the intersection of the word segmentation set and the sensitive word list set to obtain sensitive words; this algorithm has a time complexity of O(m*k). Alternatively, use an Aho-Corasick automaton to perform rule matching on the word segmentation set; the time complexity remains O(m*k). Construct a Trie tree for the file, and create a Fail pointer for each node in the Trie tree. The Fail pointer points to a new node where matching begins if the current node fails to match. However, the new node must satisfy the condition that the prefix of its pattern string is a substring of the pattern string of the previous node (i.e., the successfully matched portion). The matching process is as follows:

[0063] 1) Starting from the root node of the Trie tree, move down the automaton each time according to the character read in.

[0064] 2) When a character read in does not exist in the branch, recursively follow the Fail pointer path. If the Fail pointer path reaches the root node, skip the character and process the next character.

[0065] 3) Since the Aho-Corasick automaton moves along the text, after reading all the input text, it recursively goes through the failed paths until it reaches the root node, which can detect all the patterns.

[0066] The screening mode has the following two modes:

[0067] 1. A specific sensitive word list, through which the existing sensitive words can be directly extracted from the word segmentation set, but the construction of this word list requires a large amount of manpower.

[0068] 2. A fuzzy word list, which screens the word and the words before and after the word as the center, so as to achieve the effect of sensitive word screening and extraction, but the error of this method is larger.

[0069] Improvement: Since the word set after word segmentation itself may have errors, the AC automatic machine is selected to realize the function, in order to avoid using too much memory in constructing the Trie tree and the Fail pointer when the translated text is divided into multiple end texts according to the punctuation marks, and the short texts are matched.

[0070] In the preferred scheme, the extracted keywords and question words are fed back to the user, and further comprising the following steps: calculating the correlation degree (similarity) between the extracted keywords and the no-leader group theme, based on the correlation degree and the number of sensitive words used, the speech of the speaker is comprehensively scored, and the comprehensive score is fed back to the user (such as the interviewer), wherein the comprehensive score is a weighted score.

[0071] In summary, the no-leader group discussion system in the application converts the captured speech in the form of speech into speech in the form of text, performs word segmentation processing on the speech in the form of text, filters out the stop words in the speech during the word segmentation processing, retains the words of the specified part of speech, obtains the filtered word segmentation set, uses the words in the filtered word segmentation set to construct the vertex set, uses the co-occurrence relationship to construct the edges between any two points in the vertex set, obtains the candidate keyword graph of the filtered word segmentation set, calculates the weight of each vertex set in the candidate keyword graph, sorts the weights of the candidate keyword graph from large to small, and selects the N vertexes with the largest weight values as the keywords of the speech, so that the interviewee or the examiner can quickly understand whether the speech of the interviewee deviates from the theme by extracting the keywords, thereby assisting to improve the interview ability of the interviewee or the interview efficiency of the examiner.

[0072] In the preferred scheme, the application extracts the sensitive words that should not appear in the interview process from the interview speech and sends them to the interviewee or the examiner, so as to assist the interviewee or the examiner to quickly understand whether the speech of the interviewee is appropriate, thereby assisting to improve the interview ability of the interviewee or the interview efficiency of the examiner.

[0073] The above merely provides the preferred embodiments of the present application, and is not used to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modifications, equivalent replacements, improvements, etc. made within the principles and technical scope of the present application shall fall into the scope of the present application.

Claims

1. A leaderless group discussion system, characterized by, It comprises a voice capturing module, an audio converting module and a keyword extracting module, which are connected in sequence. The voice capturing module is used for capturing the speech of a speaker in the form of voice. The audio converting module is used for converting the captured speech in the form of voice into the form of text. The keyword extracting module is used for carrying out word segmentation on the speech in the form of text, filtering out stop words in the speech during the word segmentation process, retaining words of specified parts of speech to obtain a filtered word segmentation set, constructing a vertex set using the word segmentation in the filtered word segmentation set, constructing edges between any two points of the vertex set using a co-occurrence relationship to obtain a candidate keyword graph of the filtered word segmentation set, calculating the weight of each vertex set in the candidate keyword graph, sorting the weights of the candidate keyword graph from large to small, and selecting the top N vertices according to the weight value as keywords of the speech. The correlation degree between the extracted keywords and the theme of the leaderless group is calculated, a comprehensive score of the speech of the speaker is calculated based on the correlation degree and the number of sensitive words used, and the comprehensive score is fed back to the user, wherein the comprehensive score is a weighted score. The keyword extracting module comprises: A word segmentation unit is used for loading a dictionary, establishing a Trie tree word segmentation module, carrying out DGA word segmentation, constructing a word segmentation DGA graph, calculating the global probability of the word segmentation DGA graph to obtain a word frequency maximum segmentation combination based on a prefix dictionary, identifying English and Chinese in the word segmentation DGA graph using Token for logged-in words recorded in the dictionary and for unlogged-in words, identifying English, numerical values and time forms in English and giving corresponding marks, loading a hidden Markov (HMM) probability model graph for Chinese words, and obtaining filtered word segmentation sets by using a Viterbi algorithm dynamic programming to obtain word segmentation and annotations. It also comprises a sensitive word extracting module connected with the audio converting module, which is used for extracting the speech in the form of text according to a pre-constructed sensitive word list. The sensitive word extracting module comprises: A matching unit is used for constructing a Trie tree of the speech in the form of text, constructing a Fail pointer for the junctions on the Trie tree, starting from the root node of the Trie tree, moving down along the AC automaton each time according to the read-in characters, recursively walking the Fail pointer path when the read-in characters do not exist in the branch, skipping the character and processing the next character if the Fail pointer path is walked to the root node, and finally recursively walking the failure path until the root node is reached to complete the matching after reading all the input text. A screening unit is used for screening the word segmentation that matches successfully as sensitive words according to the matching result of the matching unit. The matching unit matches short texts, which are obtained by dividing the translated text according to punctuation marks.

2. The leaderless discussion system of claim 1, wherein, The keyword extracting module further comprises: A cleaning unit is used for identifying special characters in the speech in the form of text and deleting the special characters in the speech in the form of text. The extraction unit is configured to construct a vertex set using the segmented words in the filtered segmented word set, and construct edges between any two points in the vertex set using a co-occurrence relationship, to obtain a candidate keyword graph of the filtered segmented word set, calculate the weight of each vertex set in the candidate keyword graph, sort the weights of the candidate keyword graph from large to small, and select the top N vertices with the largest weight values as keywords of the speech.

3. The leaderless discussion system of claim 1, wherein, The extraction unit calculates the weight of each vertex set in the candidate keyword graph by using the following formula: ； In the formula, i , j is the serial number of two different segmentation vertices, is the weight of the i th segmentation vertex, d is the damping coefficient; is the segmentation vertex set pointing to the i th segmentation vertex; is the segmentation vertex set pointed to by the j th segmentation vertex; is the weight of the j th segmentation vertex.

4. The leaderless discussion system of claim 1, wherein, The screening unit uses any one of the following screening methods: A specific sensitive word table screening method: extracts existing sensitive words from the segmented word set of the speech in text form through a specific sensitive word table; A fuzzy word table screening method: screens and extracts sensitive words by taking a preset word as the center and shielding the preset word and the words before and after the preset word.

5. The leaderless discussion system of claim 4, wherein, The matching unit is further configured to split the speech in text form according to punctuation marks to obtain a plurality of end texts, and then match the plurality of end texts.

6. The leaderless discussion system of claim 5, wherein, The system further includes a feedback module configured to send the extracted keywords / sensitive words to a user.

Images