CN109918650B - Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method - Google Patents
Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method Download PDFInfo
- Publication number
- CN109918650B CN109918650B CN201910109002.7A CN201910109002A CN109918650B CN 109918650 B CN109918650 B CN 109918650B CN 201910109002 A CN201910109002 A CN 201910109002A CN 109918650 B CN109918650 B CN 109918650B
- Authority
- CN
- China
- Prior art keywords
- interview
- user
- module
- answer
- question
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Abstract
The invention discloses an intelligent interview robot device for automatically generating an interview draft and an intelligent interview method for automatically generating the interview draft, wherein the intelligent interview robot device comprises an interview template, a knowledge base, an answer discrimination module, a keyword extraction module, an answer pursuit module, a voice synthesis module, a voice recognition module, a character input module, an interview draft one-key generation module, an emotion discrimination and adjustment module, a speech speed selection module and a retrieval module; adopting a constructed interview template in combination with a retrieval mode, interviewing according to different logic branches according to the condition that a user answers questions, extracting keyword information from the user answers, and retrieving corresponding contents for follow-up; and automatically generating the interview draft by extracting the keywords and matching the interview template. The invention can solve the problems of few chatting turns, unclear chatting logic structure, inaccurate voice recognition and the like in the interview process.
Description
Technical Field
The invention belongs to the field of language character and voice processing, relates to an intelligent interview technology, and particularly relates to an interview robot device and an intelligent interview method for automatically generating an interview draft.
Background
A chat bot (chatterbot) is a program used to simulate a human conversation or chat and may also be referred to as a voice assistant, chat assistant, conversation bot, and the like. It attempts to set up a program that, at least temporarily, lets a real human think they are chatting with another.
Various types of chat robots are available in the market, such as jinmi customer service robot in kyotong, children education robot, mini-ice entertainment chat robot, Alexa home control, vehicle-mounted control robot, Viv all-directional service type robot, and the like. This is a division of the chat robot from the application side.
According to the application purpose, the chat robot is divided into: target-driven and non-target-driven chat robots.
The target-driven chat robot refers to a chat robot with a definite service target or service object, such as a customer service robot, a childhood education robot, a service robot providing weather/ticket/meal ordering services like Viv, and the like.
By targetless driven chat robots is meant chat robots that are not developed for domain specific service purposes, such as pure chat or virtual character chat robots for entertainment chat purposes as well as in computer games. Such a chat robot without a clear task target may also be referred to as an open-domain chat robot.
According to the technical means, the chat robot can be divided into: a search-type robot, a generative chat robot and a robot based on artificial templates.
The search-type chat robot walks along a route similar to a search engine, a developer stores a conversation library in advance and establishes an index, and a chat system extracts response contents in a search matching mode in the conversation library after receiving a sentence input by a user. Obviously, this method has high requirements on the dialog library, and needs the dialog library to be large enough to match with the question of the user as much as possible, otherwise, the situation that the user cannot find the appropriate answer content often occurs (because what the user says is possible in the real scene), but it has the advantage of high answer quality because the content in the dialog library is the real dialog data and the expression is natural.
The manual template-based technique is implemented by manually setting dialog scenarios and writing a targeted dialog template for each scenario, the template describing possible questions and corresponding answers for the user. The technical route has the advantages that the technical route is accurate, various dialogue robots similar to Siri on the market use a large amount of manual template technology, the accuracy of the manual template technology is incomparable with other methods, and the expandability is poor.
The generating type chatting robot adopts different technical ideas, a sentence is automatically generated as a response by adopting a certain technical means after a sentence input by a user is received, the route robot has the advantages that the user question of any topic can be covered, most of the technologies are improved under the Encoder-Decoder (or Sequence to Sequence) deep learning technical framework, and the generating type chatting robot has the defect that the quality of the generated response sentence is likely to have problems, such as sentence incompleteness, syntax error and the like which look low-level errors.
The target-driven robot is mainly realized by searching and a mode based on an artificial template, and the non-target-driven robot increasingly uses a deep learning method to generate answers in recent years. No matter which kind of chat robot, there are few turns of chatting at present, the logical architecture of chatting is unclear scheduling problem, and do not possess the function of automatic information extraction from user's speech, generation manuscript.
Disclosure of Invention
In order to overcome the defects of the prior art, the invention aims to provide an intelligent robot device for automatically generating an interview draft and an intelligent interview method for automatically generating the interview draft, which are used for solving the problems of few chatting turns, unclear chatting logic architecture, inaccurate voice recognition and the like during interview.
The technical scheme provided by the invention is as follows:
an interview intelligent robot device capable of automatically generating interview manuscripts comprises an interview template, a knowledge base, an answer discrimination module, a keyword extraction module, an answer pursuit module, a voice synthesis module, a voice recognition module, a character input module, an interview manuscripts one-key generation module, an emotion discrimination and adjustment module, a speech speed selection module and a retrieval module;
the interview template comprises a plurality of sub-modules of various information of interview objects, including sub-modules of character basic information, education background, working condition, marriage and love condition, future prospect and the like; the basic information submodule of the person is used for questioning the basic information of the person, including name, sex, native place, date of birth and the like, and replying questions according to the corresponding information of the user; the education background submodule is used for interviewing the education backgrounds of people, such as the school life of junior high school and college school, and after the ages of the people are obtained from the people information module, the campus life which the people are likely to experience is interviewed, and then the problem is answered according to the corresponding information of the user; judging whether the person possibly works or not according to the age of the person and the education background information, if the person possibly works, entering a working condition submodule, interviewing the working experience and personal experience of the person, answering the user according to the corresponding reply of the user, and forming interaction with the user; the marriage condition submodule is mainly used for gathering the emotional experience of the user, firstly gathering the love experience of the user, then judging whether the user is probably married by combining the information obtained by the character basic information module, and if married, then gathering the marriage condition of the user, including children and the like, and replying the user according to the reply of the user; the future prospect submodule acquires the prospect of the user for the future after knowing the basic information of the user through the modules, and intelligently replies the user according to the answer of the user.
The knowledge base module mainly stores knowledge which may be used by the machine in the interview process, and comprises sub-modules of geography, year, constellation, celebrity birth date, career table and the like; the geographic submodule stores the place names of cities, counties and the like contained in each province and is used when the person information module interviews the native places of the users; the year submodule stores big events which occur in 1900-2018 every year, is used for interviewing the birth date of the characters in the character basic information submodule and replying the big events which occur in the year of birth of the user according to the answer of the user so as to create a harmonious and interesting interview atmosphere; the constellation submodule stores the birth date corresponding to each constellation and the characteristics corresponding to each constellation, and the character basic information submodule replies the characteristics of the corresponding constellation of the user after obtaining the birth month date of the user so as to promote the interestingness of the interview; the celebrity birth date submodule stores celebrities born each day in one year, and randomly replies three celebrities and the careers of the user, which are the same as the celebrities on other birthdays, after the person basic information module obtains the birth date of the user; the career table sub-module is used for storing each career and relevant characteristics of the career, and the working condition module in the interview module can acquire career information of the user and reply corresponding conversations according to the career of the user to give the user a careful experience;
the answer judging module filters out the language word which is frequently used by people under consciousness, such as: 'chess', 'woolen', and the like, and analyzes the rest contents of the user answer words to judge whether the user answers the current question. If the user age is interviewed, the answer judging module checks whether the answer words of the user have reasonable numbers (between 3 and 110), if the answer words of the user do not contain the numbers or contain the unreasonable numbers, the answer judging module judges that the user does not answer the question and repeatedly asks the question once, and if the answer words of the user still do not contain expected answers, the answer judging module asks the question: "is it inconvenient to answer the question? ", if the user answers the question inconveniently, then answer: "good, that is, it is not hard to get, we enter the next question bar-", if the user answers the question conveniently, reply: "just not clearly heard, whether to select keyboard input? If the user selects keyboard input, the user uses the keyboard configured by the intelligent device to manually input answers, and if the user selects not to use keyboard input, the user asks questions again and continues to analyze the information of the user answer.
And the keyword extraction module is used for extracting answers according to all possible reply sentence patterns corresponding to the question after the answer judgment module judges that the current reply of the user contains the content required by the interview, and comprises the following steps: "what name you call? ", possible replies are: "I call XXX", "I name is XXX", "I surname X (X), name X (X)", etc., and correspondingly extracted: the keyword (name) can be obtained by following the words such as "call" or "yes" or "last name" and "first name".
The answer follow-up module is used for adding related questions to follow up the question according to the answer of the user in order to improve the interactive experience of the user in the interview process, for example, when the interest and hobbies of the user are asked, the user answers the question of watching the movie, and then the user likes which movie is the favorite of the user, which movie star is the favorite of the user, and the like; for the situation that the related information of the keywords is continuously updated in an iterative manner, the intelligent robot device searches related problems on line through the search module to pursue.
And the voice synthesis module is used for converting the selected intelligent answer (or interview words) from the text into voice and speaking the voice to the user to listen, wherein the whole interview process is based on voice interaction.
The voice recognition module is used for interacting the interview process based on voice, but the process of analyzing the answer of the user and extracting the keywords is directly operated on the text, and the answer of the user is converted into the text from the voice by the voice recognition module so as to be analyzed by the intelligent robot.
A text input module, in the answer judging module, if a user feels convenient to answer and already answers a certain question, but the voice recognition module does not recognize the content of the user answer or, for the question that the voice recognition is easy to make mistakes, such as: "ask for your name is? And the like, the problem of wrongly recognizing characters is easy to be recognized by voice recognition, and the user inputs answers by selecting a keyboard, so that the user answers directly on the keyboard through a character input module instead of the voice recognition module.
The interview draft one-key generation module is used for converting the reply of a user into characters through the voice recognition module after each interview question is proposed, inputting the characters into the keyword extraction module after the answer discrimination module confirms that the question is answered, extracting information required by the interview draft and filling the information into the interview draft or finding description corresponding to the keywords and filling the description into the interview draft; after the interview is finished, all interview manuscripts about the interview can be generated by clicking a one-key generation button of the intelligent robot device.
And the emotion judging module and the adjusting module are used for detecting the returned moods of the user corresponding to a certain problem, and if the moods of the user are low, retrieving the jokes related to the current topics as the moods of the user to be adjusted in response. For certain topics, such as: the emotional experience is that the user may recall the matters of the heart, so for the sensitive topics, the answer words of the user need to pass through the emotion judging module and the adjusting module, the trained neural network model judges the emotion of the current user, if the user is heart-impaired, the mood is adjusted by laughing, and if the user is happy or neutral, the interview is continued.
The speech rate selection module, because this intelligent robot's human-computer interaction is based on pronunciation realization almost, both sides speech rate is comparatively close can let the dialogue carry on more harmonious, harmonious. Therefore, when the user answers the first three questions, the corresponding speech synthesis speech rate is selected according to the average speech rate of the reply content, and the difference between the speech rates of the human-machine party and the human-machine party is ensured to be small.
The search module and the keyword extraction module obtain the keywords in the user answers, and for some keywords with update conditions at any time, the question-hunting database does not store corresponding questions but searches online to obtain related information and then asks for the corresponding questions.
The invention also provides an intelligent interview method for automatically generating the interview draft, which utilizes the intelligent robot device for automatically generating the interview draft to automatically generate the interview draft, adopts an interview template which is manually constructed in combination with a retrieval mode, carries out interview according to different modules of the interview template or a certain logic branch line of the modules according to the condition of answering questions by a user, extracts keyword information from the answers of the user, and retrieves corresponding contents for follow-up. And automatically generating the interview draft by extracting the keywords and matching the interview template. The method comprises the following steps:
1. constructing an interview template:
in specific implementation, a set of complete interview templates is constructed manually according to interview requirements and cautions. The interview template mainly relates to modules of character basic information, education background, working condition, marriage and love condition, future prospect and the like, and different modules or different logic branches in the same module are selected for different people to ask questions. Some questions in the same module have multiple answers, and different answers of the user will result in a large difference in the contents of the next questions. Therefore, the logic branch design is considered during template construction, and the interview logic branch is constructed according to each sub-module; each logical leg corresponds to a sub-template.
Such as: "do you talk about a love? "if the user answers yes, then go to the logical branch associated with the love, such as by" what is the first meet of you and your other half? "," the most specific way you express love is? "," what is you most important to keep a sense of emotion? "and so on, interview the past emotional experience and feelings of the user; if the user answers that the love is not mentioned, the user enters a logic branch line of the love thinking, and interviews the user's conception of the love, for example, by "do your partner ideal type? Why? "," is a desire to develop a piece of love? Is the attitude of your love? "etc., interview the user's prospect for love, etc. When designing an interview template, different conditions of users are considered in each part, a plurality of logic branches are designed, and branch points of the logic branches mainly comprise: whether to go to university, whether to work, whether to talk about love, whether to marry, whether to have children, etc.
In the basic information module of the person, the name, the sex, the native place, the birth date and the like of the person are mainly asked. And replying the question according to the corresponding information of the user. If the user answers: 'I is a black-roughneck finish man', the system can randomly answer 'Xinjiang roughneck chicken with lattice flowing water' or 'lattice feels that Xinjiang is beautiful like blood mixing' or 'Xinjiang has a caray oilfield', and the like. According to different years and months of birth of the user, the invention can also reply the answer words with corresponding year characteristics and constellations.
In the specific implementation of the invention, after the basic information of the character is obtained, 3 logic branches are divided according to the age:
the logic branch is a path for questions at the time of interview, which means that when a question has different answers, a series of different subproblems for subsequent questions is determined according to the answer (i.e., the logic branch is represented as a plurality of subprograms). For users under 22 years old, their childhood memories, learning experiences, love experiences and hobbies are mainly interviewed. And combined with the user's answer, enter the corresponding logical branch. For example, whether the user talks about a love or not, if the user talks about the love, the user enters a logic branch line for talking about the love, and past emotional experiences and feelings of the user are interviewed; and if the user does not talk about the love, entering a branch line without talking about the love, interviewing the prospect of the user for the love and the like. After the content of the logic branch line is acquired, the logic branch line returns to the main logic line, then the interest and hobbies of the user are acquired, and the follow-up is carried out according to the answer of the user, for example, when the interest and hobbies of the user are inquired, if the user likes music, the favorite music, singers and the like of the user are inquired; if the user likes to watch a movie, the user is asked about his favorite actors, the type of movie, and the like.
For users 22-30 years old, the study experience, the work experience, the marriage and love situation, the hobbies and the future prospect and the like of the users are mainly interviewed. And selects corresponding logic branch line questions according to different conditions of users. Such as asking the user if he is at school too much, and if he is at university, entering a logical branch of the university's life by: "why do you decide on college? "," which specialty you chose? Why? "," do you have a social organization in place? How do you feel about the experience? "question of college experience and feeling of interviewing users; if the user has not visited the university, the main logical leg is pursued, interviews whether the user is working, and if the user is already working, the working logical leg is entered by: "what work is done now? "so far, do you jump over a slot several times? "ask the user's job status, if the user is not already working, then follow the main logic branch to ask the user's love and marriage status. If the user is not married, entering a logic branch line of the unmarried state, and interviewing the opinion of the user on love and future prospect; if the user has married, by "what do you feel the largest change in life before and after marrying? "," will you choose a wedding if you are given you a chance again? "ask questions and so on to interview the user's opinion of the marriage. Then enter logical node, ask user whether have children, if user have children, through "you have several children to? "what do you feel most important about education of children? "interview the situation of the user and the child, and educate the child's law of growth; if the user does not have children, by: "do you like children? Why? "ask the user's opinion of the child.
For users over 30 years old, the system mainly adopts the work experience, the hobbies and interests, the marriage and love conditions, the future prospect and the like. And for the answers of different users, corresponding logic branches are entered for further questioning. The branching points of the main logic legs are still: whether to go to university, whether to work, whether to talk about love, whether to marry, whether to have children, etc.
2. In order to keep the atmosphere in the interview process comfortable and pleasant and have high flexibility and improve the experience of a user, a knowledge base is built, and information is extracted from the knowledge base and filled into a template to interact with the user.
The main knowledge base is constructed as follows:
(1) province city and county city lists of China;
when the system is used specifically, as long as a user speaks a place name, the interview robot replies an answer with regional characteristics. The relation between the place names is stored in a dictionary form, each province or direct prefecture city is used as a key, and the corresponding city or county is used as a value; the feature description corresponding to province or direct prefecture city is also stored in a dictionary mode, the place name is key, and the corresponding description is value.
(2) A year knowledge base which records the important events of nearly one hundred years and the years of occurrence thereof; expressing in a dictionary mode;
after the user answers the birth year, the invention replies the information related to the corresponding year. The year knowledge base is represented by a dictionary, the year is key, the important event is value, and the corresponding value is found as an answer according to the year of the answer of the user.
(3) The constellation knowledge base stores constellations and corresponding descriptions thereof in a dictionary mode;
and according to the birth date of the user, the robot replies a message containing corresponding constellation information to the user. The dictionary indicates that the constellation is key and the description corresponding to the constellation is value.
(4) A celebrity date of birth table storing celebrities born daily in the month of history;
after the interview robot obtains the information of the birth date of the user, the interview robot tells the user which celebrities and users have the same birth date in history. The month and day are keys, the celebrity corresponding to the birthday and the occupation are values, one key may correspond to a plurality of values, and three values are randomly selected each time to serve as answers of the replying user.
(5) A career table in which careers and their related features are stored;
randomly extracting a certain characteristic of the occupation of the user to answer according to the reply of the user, and if the user is a programmer, saying that: "listen and speak programmers to overtime, wonder" or "programmers can all be rigorously" etc. Key is a career, value is a candidate description corresponding to the corresponding career, and one of a plurality of values is randomly selected for replying each time. (6) The problem-pursuing database stores a plurality of keywords which may appear in the interviewing process and related pursuits thereof. Such as: the question corresponding to the keyword "movie" is "which movie you like most is like? "," who the favorite shadow star is ", etc. Key is a movie, value is a relevant question, and a question is randomly selected from the value to be asked each time.
3. Extracting keywords, judging and asking for questions;
for answers of users in the interview process, responses with more special answer types, such as ages, birthdays and the like, contain numbers, the numbers in the responses of the users are searched for through regular matching, if the numbers contain reasonable numbers, the responses are considered to be answered, and if the numbers do not contain reasonable numbers, questions are asked repeatedly; for the reply with flexible and changeable answer types, the number of coincident words of the user answer and the question is calculated, the number of coincident words of the user answer and the answer template is evaluated, whether the user answers the question or not is judged, and if the user does not answer the question, the question is asked repeatedly; if the user has answered, a question is asked based on the content of the user's answer.
For the follow-up, according to the answer of the user, if the answer of the user is short (after the language words such as 'hiccup' and 'en' are removed, the answer is less than 2 words) or the number of follow-up times reaches two times, the round of conversation is ended, and the follow-up visit is carried out in the next stage. Otherwise, extracting keywords in the answer of the user, searching based on the extracted keywords in the existing question-following database, and if corresponding questions are found, performing question-following. And if no corresponding item exists in the question-chasing question library, performing online retrieval, selecting the item with the highest similarity to the current question and the answer of the user for question-chasing according to the retrieved item and on the basis of top20 of the retrieval result in combination with the question and the answer of the user (word vectors are trained in advance, and the similarity between sentences is calculated by using a word2vec model). For example, when the user answers that he likes to watch the heddles at ordinary times, the user searches for videos with heddles labels, randomly selects some information and interacts with the user, such as: "is you looking at a very hot woollen cloth recently detected by star big? "and the like.
4. Speech synthesis and speech recognition with user interaction
When the method is specifically implemented, the voice recognition and voice synthesis interface of science news is combined, the template problem is converted into voice through the voice synthesis technology, and a user is asked; and converting the answer of the user into characters for analysis through a voice recognition technology.
5. Character input
In the answer judgment, if a certain question is answered conveniently and already answered, but the voice recognition module does not recognize the content of the answer of the user, or the question which is easy to make mistakes in voice recognition is identified, such as: "ask for your name is? The user can select the keyboard to input answers, and then the answers of the user do not pass through the voice recognition module any more, but pass through the character input module, and directly take the input of the user on the keyboard as answers. Therefore, the problems of inaccurate information extraction and the like caused by inaccurate speech recognition are avoided.
6. One-key generation of interview draft
In the interview process, the voice recognition technology is applied to convert the answers of the users into characters, keywords in the answers of the users are extracted in a regular matching or template matching mode, nonsense words in the answers of the users are removed, and then templates of corresponding logic branches are filled in the answers of the users, so that the interview draft is generated.
7. Emotion discrimination and adjustment
Word2vec models are trained on 60 ten thousand Chinese corpus of microblog data, word2vec is used by each word in the user answers (user answers) to obtain corresponding word vectors, sentence-level vectors of the user answers are obtained through two layers of LSTMs, and the vectors are input into a trained emotion classifier to obtain emotion labels (happy, neutral or difficult to pass) of the user answers. If the emotion label is a happy or neutral emotion label, performing interviewing according to the process described above; if the answer of the user is subjected to emotional classification and the result is difficult, the joke related to the current topic is searched as a response, and then the question is asked for the user.
And (3) training of an emotion discriminator: the method comprises the steps of segmenting words from a Chinese emotion analysis data set, training word vectors on the data set, inputting the word vectors corresponding to each word into a two-layer neural network (LSTM) to extract sentence information, taking the hidden state of neurons of the last layer of network as compressed sentence information to represent, obtaining probability distribution of the sentences corresponding to negative type (injury), neutral type and positive type (distraction) emotions through a Softmax layer, taking the type with the highest probability as a prediction tag, calculating loss between the prediction tag and a real tag, and training the network.
After the emotion discriminator is trained, in the actual application stage, the emotion label of the sentence can be obtained by inputting the user answer.
8. Speed of speech selection function
During conversation, if the speech speeds of the two parties are matched, the communication can be smoother. Therefore, after the user answers the first question, the speech rate output by the interview robot is adjusted according to the speech rate of the user speaking.
The invention has the beneficial effects that:
the interview template is constructed and used, so that interaction with a user can be accurately carried out; the self-built knowledge base is used, so that the interactive content is enriched, the atmosphere is activated, and the user experience is improved; and combining the technology of extracting the key words to automatically generate the interview draft. The invention can solve the problems of few chatting turns, unclear chatting logic architecture and inaccurate voice recognition of few situations during interview, and improves the interview effect.
Drawings
Fig. 1 is a flow chart of a method for automatically generating an interview draft by an intelligent robot apparatus according to the present invention.
FIG. 2 is a schematic diagram of an emotion determination framework provided in the present invention;
wherein, Long Short-Term Memory (LSTM) represents Long and Short Term Memory network; the Softmax layer is a classification layer, probability distribution of sentences corresponding to negative classes (casualties), neutral and positive classes (happily) is obtained through the Softmax layer, and the class with the highest probability is used as a prediction label.
Detailed Description
The technical solutions according to the invention are further illustrated below with reference to examples and figures, without in any way limiting the scope of the invention:
fig. 1 is a flow of a method for automatically generating an interview draft by an intelligent robot apparatus according to the present invention, and as shown in fig. 1, the present invention implements an automatic short news manuscript writing method based on intelligent template selection to automatically generate a short news manuscript; the method comprises the following steps:
11) building templates
According to the interview requirements and cautions, a set of complete interview templates is constructed manually. The interview template mainly relates to modules of character basic information, education background, working condition, marriage and love condition, future prospect and the like, and different modules or different logic branches in the same module are selected for different people to ask questions.
In the basic information module of the person, the name, the sex, the native place, the birth date and the like of the person are mainly asked. And replying the question according to the corresponding information of the user. If the user answers: 'I is a black-roughneck finish man', the system can randomly answer 'Xinjiang roughneck chicken with lattice flowing water' or 'lattice feels that Xinjiang is beautiful like blood mixing' or 'Xinjiang has a caray oilfield', and the like. According to different years and months of birth of the user, the invention can also reply the answer words with corresponding year characteristics and constellations.
After obtaining the basic information of the person, 3 logic branches are divided according to the age:
for users under 22 years old, their childhood memories, learning experiences, love experiences and hobbies are mainly interviewed. And combined with the user's answer, enter the corresponding logical branch. If the user likes music, the user's favorite music, singer, etc. is inquired when inquiring about the user's interests and hobbies; if the user likes to watch a movie, the user is asked about his favorite actors, the type of movie, and the like.
For users 22-30 years old, the study experience, the work experience, the marriage and love situation, the hobbies and the future prospect and the like of the users are mainly interviewed. And selects corresponding logic branch line questions according to different conditions of users. For example, when inquiring about the marriage condition of the user, if the user is not married, the opinion of the user on the marriage and the prospect in the future are interviewed; and if the user is married, inquiring whether the user has children after interviewing the opinion of the user on the marriage, and if the user has children, further interviewing the opinion of the user on the educational development of the children.
For users over 30 years old, the system mainly adopts the work experience, the hobbies and interests, the marriage and love conditions, the future prospect and the like. And for the answers of different users, corresponding logic branches are entered for further questioning. If the user is asked about the work experience of the user, the user will interview his view of the slot jump if he has the experience of the slot jump.
12) Building a knowledge base
In order to keep the interview process comfortable and pleasant in atmosphere and high in flexibility and improve the experience of a user, a knowledge base is built, and interaction is carried out with the user in a mode of extracting information from the knowledge base and filling the information into a template.
The main knowledge base is constructed as follows: (1) the province level city and county level city of provinces and provinces of China are listed, and as long as the user speaks a place name, the interview robot replies an answer with regional characteristics. The relation between the place names is stored in a dictionary form, each province or direct prefecture city is used as a key, and the corresponding city or county is used as a value; the feature description corresponding to province or direct prefecture city is also stored in a dictionary mode, the place name is key, and the corresponding description is value. (2) The year knowledge base records the years of the occurrence of the important events of nearly one hundred years, and after the user answers the birth year, the invention replies the information related to the corresponding year. The year knowledge base is represented by a dictionary, the year is key, the important event is value, and the corresponding value is found as an answer according to the year of the answer of the user. (3) And the constellation knowledge base is used for replying the information containing the corresponding constellation information to the user by the robot according to the birth date of the user. The dictionary indicates that the constellation is key and the description corresponding to the constellation is value. (4) And the visiting robot obtains the information of the birth date of the user and then tells the user which celebrities and users have the same birth date historically. The month and day are keys, the celebrity corresponding to the birthday and the occupation are values, one key may correspond to a plurality of values, and three values are randomly selected each time to serve as answers of the replying user. (5) The career table stores various careers and related characteristics thereof, a certain characteristic of the careers of the users is randomly extracted to answer according to the responses of the users, and if the users are programmers, the career table comprises the following steps: "listen and speak programmers to overtime, wonder" or "programmers can all be rigorously" etc. Key is a career, value is a candidate description corresponding to the corresponding career, and one of a plurality of values is randomly selected for replying each time. (6) The problem-pursuing database stores a plurality of keywords which may appear in the interviewing process and related pursuits thereof. Such as: the question corresponding to the keyword "movie" is "which movie you like most is like? "," who the favorite shadow star is ", etc. Key is a movie, value is a relevant question, and a question is randomly selected from the value to be asked each time.
2) Selecting interview content
If the interview is just started, basic information such as user name, gender and the like is asked; if the answer is in the interview process, extracting the key words according to the answer of the last question of the user to judge whether the answer is needed to be asked again (if the user does not answer the last question), if the user answers the question, judging the emotional tendency of the answer of the user, inputting the answer of the user to judge the emotion of the user as shown in figure 2, if the emotion is happy or neutral, selecting the answer from a template or a knowledge base, if the emotion is difficult, searching a joke related to the current conversation theme to reply, and then interviewing.
In FIG. 2, Long Short-Term Memory (LSTM) represents a Long Short-Term Memory network, which is a time-recursive neural network suitable for processing and predicting important events with relatively Long intervals and delays in time sequence; LSTM plays a more important role in the relevant tasks of natural language processing. And (3) training of an emotion discriminator: specifically, words are segmented from a Chinese emotion analysis data set, word vectors are trained on the data set, the word vector corresponding to each word is input into a two-layer neural network (LSTM) to extract sentence information, the hidden state of a neuron of the last layer of network is taken as compressed sentence information to be represented, probability distribution of the sentences corresponding to negative type (casualty), neutral type and positive type (open heart) emotions is obtained through a Softmax layer, the category with the highest probability is taken as a prediction tag, loss between the prediction tag and a real tag is calculated, and the network is trained. After the emotion discriminator is trained, in a practical stage, the user answering is input, and the emotion label of the sentence can be obtained.
3) Speech synthesis
And converting the interview content into voice from characters by using a voice synthesis technology.
4) Speech recognition
The user's answer is converted from speech to text by speech recognition techniques.
5) Character input
For a question corresponding to an answer that is easy to make mistakes in speech recognition (for example, "ask your name is.
6) Keyword extraction
And extracting the keywords in the user response by regular matching or template matching. If the type of the answer is special, if the answer is a number when the interview age is long, extracting by matching the number in the answer language of the user; if the answer type is the common text, extracting the answer according to all possible reply sentence patterns corresponding to the question, such as: "what name you call? ", possible replies are: "I call XXX", "I name is XXX", "I surname X (X), name X (X)", etc., and correspondingly extracted: the keyword (name) can be obtained by following the words such as "call" or "yes" or "last name" and "first name".
7) Generating an interview draft
If a certain logic branch line has a plurality of templates, one template is randomly selected, and the extracted keywords are filled in the templates of the corresponding logic branch lines according to the templates and the user answers to generate the interview draft.
It is noted that the disclosed embodiments are intended to aid in further understanding of the invention, but those skilled in the art will appreciate that: various substitutions and modifications are possible without departing from the spirit and scope of the invention and appended claims. Therefore, the invention should not be limited to the embodiments disclosed, but the scope of the invention is defined by the appended claims.
Claims (10)
1. An intelligent interview method for automatically generating interview manuscripts comprises the steps of constructing an interview template, interviewing according to conditions of questions answered by users and different sub-modules or logic branch lines provided by the sub-modules of the interview template, extracting keyword information from answers of the users, retrieving corresponding content and performing follow-up interview; automatically generating an interview draft in a mode of extracting keywords and matching the interview template; the method comprises the following steps:
1) constructing an interview template: the interview template comprises a plurality of sub-modules of various information of interview objects, including character basic information, education background, working condition, marriage and love condition and future prospect sub-modules; constructing an interview logic branch line according to each submodule; each logic branch line can be used as a sub-template of the interview template;
2) constructing a knowledge base module which comprises sub-modules of geography, year, constellation, celebrity birth date and career table;
3) and in the interview process, the following operations are performed:
31) selecting different sub-modules or different interview logic branches of the same sub-module for different people to ask questions;
32) interaction is carried out with the user in a mode of extracting information from the knowledge base module and filling the information into the interview template;
33) extracting keywords, judging and asking for questions; the method specifically comprises the following steps:
331) for the answer of the user in the interviewing process, filtering and analyzing the answer of the user, judging whether the user answers the current question or not, and identifying whether the current answer of the user contains the content required by the interviewing or not;
332) according to the sentences answered by the user to the question, extracting answers through matching to obtain keywords replied by the question;
333) searching in the existing question-pursuing question library according to the keywords, and pursuing the question if finding the corresponding question; if the question-pursuing question bank does not have corresponding entries, carrying out online retrieval, carrying out screening sequencing according to the retrieved entries and the first entries based on the retrieval result, combining the questions and the answers of the users, and selecting the entry with the highest similarity to the current questions and the answers of the users to pursue the questions;
4) and performing voice synthesis and voice recognition and user interaction in the interview process: through a voice recognition and voice synthesis interface, converting the interview template problem into voice through a voice synthesis technology, and asking a user; or the answer of the user is converted into characters for analysis through a voice recognition technology;
5) in the interview process, for the problem that accurate character reply is difficult to obtain by voice recognition, the input of a user on a keyboard can be directly used as an answer by selecting a character input mode;
6) judging emotion in the interview process, inputting user answering to obtain an emotion tag, and adjusting the interview process according to the emotion tag; the following operations are performed:
61) training a word2vec model by using Chinese corpora; training an emotion discriminator by utilizing a Chinese emotion analysis data set;
62) inputting each word in the user answer into the trained word2vec to obtain a corresponding word vector; obtaining a sentence-level vector of the user answer through two layers of LSTMs;
63) inputting the sentence-level vector into a trained emotion classifier to obtain an emotion label of a user answer;
64) determining a subsequent interview flow according to the emotion label category;
7) and (3) selecting and adjusting the speech rate in the interview process: during interview conversation, after the user answers the question, the speech rate output by the interview robot is adjusted according to the speech rate of the user;
8) generating an acquisition draft by one key: in the interview process, a voice recognition technology is applied to convert the answers of the user into characters, keywords in the answers of the user are extracted in a regular matching or template matching mode, nonsense words in the answers of the user are removed, and then templates of corresponding logic branches are filled in the answers of the user, so that an interview draft can be generated;
through the steps, the intelligent interview of automatically generating the interview draft is realized.
2. An intelligent interview method for automatically generating interview draft according to claim 1, wherein said logic branch is a path of questions asked at the time of interview, represented as a plurality of sub-templates of the interview template, and when a question has different answers, sub-questions to be subsequently asked are determined based on the answers.
3. The intelligent interview method for automatically generating the interview draft according to claim 1, wherein in the interview process, three logic branches are divided according to the age of the interview user in the character basic information:
the first logic branch line aims at users under 22 years old, and interview questions mainly aim at childhood memories, learning experiences, love experiences and interests; associating corresponding information to further ask questions in combination with the user answers;
the second logic branch line aims at 22-30 years old users, and interview questions mainly aim at learning experiences, work experiences, marriage and love conditions, interests and future prospects; and for different conditions of the user, associating corresponding information to further ask a question;
the third logic branch line is used for interview and inquiry of users over 30 years old mainly aiming at work experience, hobbies, marriage and love conditions and future prospects; and for the answers of different users, corresponding information is associated for further questioning.
4. The intelligent interview method for automatically generating interview drafts as claimed in claim 1, wherein specifically, constructing the knowledge base module comprises:
A. a geographical knowledge base is constructed in a dictionary form, and a geographical name list of a region and corresponding information description with region characteristics are stored;
B. constructing a year knowledge base in a dictionary form, and storing the information of the important events in the last century and the years of occurrence of the events;
C. the constellation knowledge base stores constellations and corresponding descriptions thereof in a dictionary mode;
D. constructing a celebrity birth date table, and storing celebrities and careers thereof born every month historically;
E. and constructing a career table and storing each career and related characteristics thereof.
5. The intelligent interview method for automatically creating interview draft as claimed in claim 1, wherein in particular, step 333) pre-trains word vectors and calculates the similarity between sentences using the trained word2vec model, thereby obtaining the vocabulary entry with the highest similarity to the current question and user answer.
6. The intelligent interview method for automatically generating interview draft according to claim 1, wherein the emotion discriminator is trained to specifically segment sentences from a Chinese emotion analysis dataset; training word vectors on the data set; inputting the word vector corresponding to each word into a two-layer neural network LSTM to extract sentence information; taking the hidden state of the neuron of the last layer of network as the compressed sentence information representation; obtaining probability distribution of emotion corresponding to the sentence through a Softmax layer; the category with the highest probability is used as a prediction label; and calculating a loss function loss between the predicted label and the real label, thereby training the emotion discriminator network.
7. The intelligent interview method for automatically generating interview draft according to claim 1, wherein the probability distribution of the emotion is divided into three categories, namely negative category, neutral category and positive category; the negative class represents heart injury; the positive category represents distraction.
8. An interview intelligent robot device for realizing the intelligent interview method for automatically generating the interview draft according to any one of claims 1 to 7, which is characterized by comprising an interview template, a knowledge base module, an answer judging module, a keyword extracting module, an answer chasing module, a voice synthesis module, a voice recognition module, an emotion judging and adjusting module and an interview draft one-key generating module; the interview intelligent robot device also comprises a character input module, a speech speed selection module and a retrieval module;
the interview template comprises a plurality of sub-modules of various information of interview objects, including character basic information, education background, working condition, marriage and love condition and future prospect sub-modules;
the knowledge base module is used for storing knowledge which may be needed in the interview process and comprises sub-modules of geography, year, constellation, celebrity birth date and career table;
the answer judging module is used for filtering and analyzing the user answer, judging whether the user answers the current question or not and identifying whether the current reply of the user contains the content required by the interview or not;
the keyword extraction module is used for extracting answers through matching according to the sentence patterns replied by the user to the question after the answer judgment module identifies that the current reply of the user contains the content required by the interview, so as to obtain the keywords replied by the question;
the answer question-chasing module is used for adding related questions to carry out question-chasing according to the answers of the user;
the voice synthesis module is used for converting the selected intelligent answer or interview words from the text into voice in the interview process based on voice interaction;
the voice recognition module is used for converting the answer of the user from voice to text, and the text can be used for analyzing the answer of the user and extracting key words;
the emotion judging module and the adjusting module are used for detecting the moods of users responding to a certain question so as to further determine an interview strategy;
the one-key generation module of the interview draft comprises the steps of submitting an interview question, converting user responses into characters, judging answers, extracting key words, extracting information required by the interview draft and filling the extracted information into the interview draft, or finding out descriptions corresponding to the key words and filling the descriptions into the interview draft, so that all interview drafts related to the interview can be generated;
the character input module is used for taking the input of the user on the keyboard as an answer when the user selects the keyboard input;
the speech rate selection module is used for selecting corresponding speech synthesis speech rate according to the average speech rate of the user to answer the question, so that the speech rates of the human-computer and the two parties are close to each other;
the retrieval module is used for updating the keywords at any time in the interviewing process, and performing follow-up after obtaining the related information through online retrieval.
9. The intelligent interview robot device according to claim 8, wherein the basic information of people submodule of the interview template is used for asking basic information of people and replying questions according to corresponding information of users; the education background submodule is used for interviewing the education background of the characters and replying the questions according to the corresponding information of the user;
the working condition submodule is used for interviewing the working experience and personal experience of the character, answering the user according to the corresponding reply of the user and forming interaction with the user;
the marriage condition submodule is used for interviewing the emotional experience of the user, including the love experience and the marital condition of the user, and replying the user according to the reply of the user;
and the future prospect submodule is used for interviewing the prospect of the user for the future after knowing the basic information of the user and intelligently replying the user according to the answer of the user.
10. The intelligent interview robot device according to claim 8, wherein the geographic sub-module of the knowledge base module is used for storing place names, and can establish association and use associated information with the interview users in place; the year submodule is used for storing major events occurring every year and can be used in association with the birth date of the interview people; the constellation submodule is used for storing the birth date corresponding to each constellation and the characteristics corresponding to each constellation, and can associate the corresponding constellation with the birth month and day of the user; the celebrity birth date submodule is used for storing the birth date and the related information of the celebrity and can be used in association with the birth date of the user; and the occupation table submodule is used for storing occupation and corresponding related characteristics of the occupation and can be used for establishing association with the occupation information of the user.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910109002.7A CN109918650B (en) | 2019-02-03 | 2019-02-03 | Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910109002.7A CN109918650B (en) | 2019-02-03 | 2019-02-03 | Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109918650A CN109918650A (en) | 2019-06-21 |
| CN109918650B true CN109918650B (en) | 2020-10-23 |
Family
ID=66961467
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910109002.7A Active CN109918650B (en) | 2019-02-03 | 2019-02-03 | Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109918650B (en) |
Families Citing this family (13)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110211591B (en) * | 2019-06-24 | 2021-12-21 | 卓尔智联(武汉)研究院有限公司 | Interview data analysis method based on emotion classification, computer device and medium |
| CN112182169A (en) * | 2019-07-01 | 2021-01-05 | 北京邮电大学 | Social engineering robot simulation method and device based on user attributes |
| CN112307156A (en) * | 2019-07-26 | 2021-02-02 | 北京宝捷拿科技发展有限公司 | Cross-language intelligent auxiliary side inspection method and system |
| US11288459B2 (en) * | 2019-08-01 | 2022-03-29 | International Business Machines Corporation | Adapting conversation flow based on cognitive interaction |
| CN110930106A (en) * | 2019-10-14 | 2020-03-27 | 平安科技（深圳）有限公司 | Information processing method, device and system of online interview system |
| CN110910283A (en) * | 2019-10-18 | 2020-03-24 | 平安科技（深圳）有限公司 | Method, device, equipment and storage medium for generating legal document |
| CN111046133B (en) * | 2019-10-29 | 2023-07-25 | 平安科技（深圳）有限公司 | Question and answer method, equipment, storage medium and device based on mapping knowledge base |
| CN113110767A (en) * | 2020-01-10 | 2021-07-13 | 阿里巴巴集团控股有限公司 | Interaction method and device |
| CN111506770B (en) * | 2020-04-22 | 2023-10-27 | 新华智云科技有限公司 | Interview video gathering generation method and system |
| CN111818290B (en) * | 2020-06-29 | 2021-07-09 | 广州快决测信息科技有限公司 | Online interviewing method and system |
| CN112185383A (en) * | 2020-09-30 | 2021-01-05 | 深圳供电局有限公司 | Processing method and system for customer service return visit |
| CN113208592B (en) * | 2021-03-29 | 2022-08-16 | 济南大学 | Psychological test system with multiple answering modes |
| CN113988070B (en) * | 2021-10-09 | 2023-05-05 | 广州快决测信息科技有限公司 | Investigation problem generation method, investigation problem generation device, computer equipment and storage medium |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080177623A1 (en) * | 2007-01-24 | 2008-07-24 | Juergen Fritsch | Monitoring User Interactions With A Document Editing System |
| CN105975466A (en) * | 2015-11-04 | 2016-09-28 | 新华通讯社 | Method and device for machine manuscript writing aiming at short newsflashes |
| CN108062359B (en) * | 2017-11-28 | 2021-03-16 | 歌尔科技有限公司 | Method for robot interview and draft writing and robot interview and draft writing |
| CN108831450A (en) * | 2018-03-30 | 2018-11-16 | 杭州鸟瞰智能科技股份有限公司 | A kind of virtual robot man-machine interaction method based on user emotion identification |
| CN109255697A (en) * | 2018-08-15 | 2019-01-22 | 普信恒业科技发展（北京）有限公司 | A kind of automatic credit estimation method and system based on artificial intelligence |
-
2019
- 2019-02-03 CN CN201910109002.7A patent/CN109918650B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN109918650A (en) | 2019-06-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109918650B (en) | Interview intelligent robot device capable of automatically generating interview draft and intelligent interview method | |
| CN107943998B (en) | Man-machine conversation control system and method based on knowledge graph | |
| CN106649704B (en) | Intelligent conversation control method and system | |
| CN115238101B (en) | Multi-engine intelligent question-answering system oriented to multi-type knowledge base | |
| CN110301117B (en) | Method and apparatus for providing response in session | |
| CN107944027B (en) | Method and system for creating semantic key index | |
| US8818926B2 (en) | Method for personalizing chat bots | |
| US11823074B2 (en) | Intelligent communication manager and summarizer | |
| CN110462730A (en) | Promote with the end-to-end communication of multilingual and automation assistant | |
| CN108446286A (en) | A kind of generation method, device and the server of the answer of natural language question sentence | |
| WO2017112813A1 (en) | Multi-lingual virtual personal assistant | |
| CN108804698A (en) | Man-machine interaction method, system, medium based on personage IP and equipment | |
| O’Shea et al. | Systems engineering and conversational agents | |
| US20200183928A1 (en) | System and Method for Rule-Based Conversational User Interface | |
| Hung et al. | Towards a method for evaluating naturalness in conversational dialog systems | |
| US20190295199A1 (en) | Intelligent legal simulator | |
| Park et al. | Systematic review on chatbot techniques and applications | |
| KR20200089914A (en) | Expert automatic matching system in education platform | |
| CN109871441A (en) | One kind knowledge neural network based of leading answers system and method | |
| Yordanova et al. | Automatic detection of everyday social behaviours and environments from verbatim transcripts of daily conversations | |
| Matsuyama et al. | Automatic expressive opinion sentence generation for enjoyable conversational systems | |
| CN117556802B (en) | User portrait method, device, equipment and medium based on large language model | |
| US20230350929A1 (en) | Method and system for generating intent responses through virtual agents | |
| KR20130068624A (en) | Apparatus and method for recognizing speech based on speaker group | |
| CN117235354A (en) | User personalized service strategy and system based on multi-mode large model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |