CN113360626A - Multi-scene mixed question-answer recommendation method for intelligent customer service robot - Google Patents

Multi-scene mixed question-answer recommendation method for intelligent customer service robot Download PDF

Info

Publication number
CN113360626A
CN113360626A CN202110751237.3A CN202110751237A CN113360626A CN 113360626 A CN113360626 A CN 113360626A CN 202110751237 A CN202110751237 A CN 202110751237A CN 113360626 A CN113360626 A CN 113360626A
Authority
CN
China
Prior art keywords
trigger
corpus
answer
question
value
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.)
Granted
Application number
CN202110751237.3A
Other languages
Chinese (zh)
Other versions
CN113360626B (en
Inventor
陈�光
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing Ronglian Qimo Technology Co ltd
Original Assignee
Beijing Ronglian Qimo Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing Ronglian Qimo Technology Co ltd filed Critical Beijing Ronglian Qimo Technology Co ltd
Priority to CN202110751237.3A priority Critical patent/CN113360626B/en
Publication of CN113360626A publication Critical patent/CN113360626A/en
Application granted granted Critical
Publication of CN113360626B publication Critical patent/CN113360626B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/332—Query formulation
    • G06F16/3329—Natural language query formulation or dialogue systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3343—Query execution using phonetics
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/3331—Query processing
    • G06F16/334—Query execution
    • G06F16/3344—Query execution using natural language analysis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/33—Querying
    • G06F16/335—Filtering based on additional data, e.g. user or group profiles
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/10—Text processing
    • G06F40/194—Calculation of difference between files
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Databases & Information Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Human Computer Interaction (AREA)
  • Acoustics & Sound (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention provides a multi-scene mixed question-answer recommendation method for an intelligent customer service robot, which comprises the following steps: prefabricating a knowledge base; prefabricating a task flow library; the method comprises the steps of obtaining a problem input by a user through a chat window of a visitor end; extracting a first trigger corpus corresponding to the problem from a knowledge base; extracting a second trigger corpus corresponding to the problem from the task flow library; and determining an answer to the question based on the first trigger corpus and the second trigger corpus. The multi-scene mixed question-answer recommendation method of the intelligent customer service robot is based on the credible NLP semantic similarity, the similarity data are sequenced, the rules that the priorities of multiple rounds are higher than the priorities of a single round are adopted, various thresholds are utilized for controlling and comparing, the final recommended answer or recommended question list of the robot is obtained, and the accurate intention or the fuzzy intention of a user is fully identified.

Description

Multi-scene mixed question-answer recommendation method for intelligent customer service robot
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a multi-scene mixed question and answer recommendation method for an intelligent customer service robot.
Background
At present, an intelligent customer service robot carries out interactive dialogue with a user by using an artificial intelligence algorithm, and the mission of the intelligent customer service robot is to help the customer solve the problem, reduce the workload of artificial customer service and improve the satisfaction degree of the user. Currently, the mainstream intelligent customer service robots in the industry (generally having three main capabilities, namely question-answer type, task type and chatty type) use NLP (natural language processing) technology to match questions provided by users with corpora (intentions) configured in the robots, calculate semantic similarity of the questions, and finally set recommendation rules and return the recommendation rules to the answers of the users. However, as enterprise service scenes become more complex, the requirement on the intelligent degree of the customer service robot becomes higher, and from practice, the method takes solving of user problems as a core, how to better give the user the best answer, and how to balance the priority of multi-turn conversations driven by tasks and single-turn conversations taking a knowledge base as a core becomes an important subject which needs to be researched by a robot question-answer recommendation algorithm.
Disclosure of Invention
One of the objectives of the present invention is to provide a multi-scenario hybrid question-answer recommendation method for an intelligent customer service robot, so as to solve the above problems.
The embodiment of the invention provides a multi-scene mixed question-answer recommendation method for an intelligent customer service robot, which comprises the following steps:
prefabricating a knowledge base;
prefabricating a task flow library;
the method comprises the steps of obtaining a problem input by a user through a chat window of a visitor end;
extracting a first trigger corpus corresponding to the problem from a knowledge base;
extracting a second trigger corpus corresponding to the problem from the task flow library;
and determining an answer to the question based on the first trigger corpus and the second trigger corpus.
Preferably, the pre-manufactured knowledge base comprises:
setting a plurality of knowledge points to obtain a common problem set;
setting a reply threshold and a difference threshold;
wherein each common problem in the common problem set comprises:
standard questions, similar questions, answers, and associated questions;
and the standard question and the similar question method are correspondingly provided with first trigger corpora.
Preferably, the pre-manufactured task flow library comprises:
setting a plurality of task flows and setting a plurality of second trigger corpora for each task flow;
and setting a task reply threshold.
Preferably, the extracting the first trigger corpus corresponding to the question from the knowledge base includes:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
and taking a first trigger corpus of which the first semantic similarity is more than a reply threshold as a first trigger corpus corresponding to the problem.
Preferably, a second trigger corpus corresponding to the problem is extracted from the task flow library; the method comprises the following steps:
calculating a second voice similarity of the third trigger corpus and the second trigger corpus;
and taking a second triggering language material with the second voice similarity above the task recovery threshold as a second triggering language material corresponding to the problem.
Preferably, the determining the answer to the question based on the first triggering corpus and the second triggering corpus includes:
sequencing the first similarity of the first trigger corpus from large to small to construct a first list, wherein the first list is SQ (SQ-SQ)1，SQ2，…，SQqTherein, SQ1>SQ2，SQ2>SQ3，…，SQq-1>SQq,SQqGreater than a reply threshold;
sorting the second similarity of the second trigger corpus from large to small to construct a second list, wherein the second list is SF ═ SF1，SF2，…，SFfIn which is SF1>SF2，SF2>SF3，…，SFf-1>SFf,SFfGreater than task returnA complex threshold value;
when SF1≥SQ1When it is time, trigger SF preferentially1A corresponding task flow;
when SF1<SQ1Then, according to the first list, calculating the difference between two adjacent first similarities in turn, wherein the calculation formula is as follows:
D(i-1)i＝SQi-1-SQi；
wherein D is(i-1)iThe difference value of the ith-1 st first similarity and the ith first similarity in the first list is obtained; SQi-1The number is the (i-1) th first similarity in the first list; SQiThe ith first similarity in the first list;
when D is present12When the difference value is greater than the threshold value, returning to SQ1Answers to the corresponding knowledge points;
when D is present12When the difference value is smaller than or equal to the difference threshold value, comparing the difference value of two subsequent adjacent first similarity with the difference threshold value, and when the first similarity is larger than the difference threshold value, marking as D(j-1)j(ii) a Return question list (SQ)1，SQ2，…，SQj) (ii) a And after the questions selected by the user based on the question list are obtained, returning corresponding answers.
Preferably, when SF1＝SQ1When 1, SF is triggered directly1A corresponding task flow;
when SF1<SQ1When the value is 1, then the SQ is directly recovered1Answers to the corresponding knowledge points;
and when the first trigger corpus and the second trigger corpus are both empty, performing bottom-of-pocket reply through a preset internet chatting library.
Preferably, the extracting the first trigger corpus corresponding to the question from the knowledge base includes:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
taking a first trigger corpus with the first semantic similarity above a reply threshold;
acquiring a first input record in a first time period preset before a user inputs a problem and/or a second input record input by opening a chat window this time;
extracting the first input record and/or the second input record to obtain a plurality of fourth trigger corpora;
acquiring a correlation problem of a first trigger corpus extracted based on the first semantic similarity;
acquiring a fifth trigger corpus of the associated problem;
calculating a third semantic similarity between the fifth trigger corpus and the fourth trigger corpus;
and correcting the first semantic similarity based on the third semantic similarity, wherein the correction formula is as follows:
Figure BDA0003146319930000041
wherein, SQ'0The corrected first semantic similarity; SQ0The first semantic similarity before correction; SGh,kThe third semantic similarity of the h fourth triggering corpus and the k fifth triggering corpus is obtained; m is the total number of the fourth trigger corpus; n is the total number of the fifth trigger corpus;
and taking the first trigger corpus of which the corrected first semantic similarity is more than a reply threshold as the first trigger corpus corresponding to the problem.
Preferably, the multi-scenario hybrid question-answer recommendation method for the smart client robot further includes:
when the first trigger corpus and the second trigger corpus are both empty, accessing a big data network platform;
sending the question to a plurality of data nodes on line;
receiving answers to be selected for the questions fed back by the data nodes;
calculating fourth semantic similarity between each answer to be selected and other answers to be selected;
acquiring a preset trust coefficient between a node corresponding to a preset answer to be selected and nodes corresponding to other answers to be selected;
acquiring a weight coefficient of a node corresponding to a preset answer to be selected and weight values of nodes corresponding to other answers to be selected;
and determining the effective value of each answer to be selected based on the trust coefficient, the weight value and the fourth voice similarity, wherein the calculation formula of the effective value is as follows:
Figure BDA0003146319930000042
wherein H is an effective value; mu.sjA preset trust coefficient is set between the node corresponding to the answer to be selected and the node corresponding to the jth other answer to be selected; SVjThe fourth semantic similarity between the answer to be selected and the jth other answer to be selected; t is0The weighted value of the node corresponding to the answer to be selected; t isdThe weighted value of the node corresponding to the ith other answer to be selected; tau is1Is a preset first weight coefficient; tau is2Is a preset second weight coefficient; epsilon1A first relation coefficient which is preset and corresponds to the influence of the answer; epsilon2A second relation coefficient which is a preset influence of the corresponding node;
and obtaining the answer to be selected with the maximum effective value as the answer of the question and outputting the answer to the user.
Preferably, the multi-scenario hybrid question-answer recommendation method for the smart client robot further includes:
obtaining the evaluation of answers to the questions by a preset number of users;
determining the rating value of the evaluation based on a preset evaluation rating template;
determining validity of the answer to the question based on the scoring value, the validity calculation formula being as follows:
Figure BDA0003146319930000051
wherein W is the effectiveness; pcA scoring value for the rating of the c-th user;
Figure BDA0003146319930000054
the credit coefficient is a preset c-th user;
when the validity is greater than a preset first validity threshold value, determining that the answer is valid; when the answer is determined to be valid, carrying out weight value up-regulation on nodes participating in providing the answer on the big data network platform; the up-regulation formula is as follows:
Figure BDA0003146319930000052
wherein, T'hThe adjusted weight value of the h node; t ishThe weight value of the h node before adjustment; rhohDetermining an adjustment index from a preset adjustment index table for the h node according to the adjustment history in a preset second time period; theta is a preset adjusting amplitude value; theta0Supplementing value for preset adjustment;
when the adjusted weight value calculated by the up-regulation formula is larger than a preset upper limit value, taking the upper limit value as the adjusted weight value;
when the validity is smaller than a preset second validity threshold value, determining that the answer is invalid; when the answer is determined to be invalid, carrying out weight value down regulation on nodes participating in providing the answer on the big data network platform; the downregulation formula is as follows:
Figure BDA0003146319930000053
wherein, t'sThe weight value of the adjusted s-th node is obtained; t is tsThe weighted value of the s-th node before adjustment is obtained; rhosDetermining an adjustment index from a preset adjustment index table for the s-th node according to the adjustment history in a preset second time period;
and when the adjusted weight value calculated by the down-regulation formula is smaller than the preset lower limit value, taking the lower limit value as the adjusted weight value.
The invention has the following beneficial effects:
firstly, based on credible NLP semantic similarity, through sequencing similarity data, adopting a rule that a plurality of rounds of priorities are higher than a single round of priorities, and utilizing various threshold value control comparison, obtaining a final recommended answer or recommended problem list of the robot, and fully realizing the identification of accurate intention or fuzzy intention of a user;
and secondly, by opening the self-defined setting of various thresholds, a user can flexibly control the question-answering logic of the robot according to different service attributes, and the usability and universality of the robot question-answering are greatly improved.
Additional features and advantages of the invention will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
The accompanying drawings, which are included to provide a further understanding of the invention and are incorporated in and constitute a part of this specification, illustrate embodiments of the invention and together with the description serve to explain the principles of the invention and not to limit the invention. In the drawings:
fig. 1 is a schematic diagram of a multi-scenario hybrid question-answer recommendation method for an intelligent customer service robot in an embodiment of the present invention.
Detailed Description
The preferred embodiments of the present invention will be described in conjunction with the accompanying drawings, and it will be understood that they are described herein for the purpose of illustration and explanation and not limitation.
The embodiment of the invention provides a multi-scene hybrid question-answer recommendation method for an intelligent customer service robot, which comprises the following steps of:
step S1: prefabricating a knowledge base;
step S2: prefabricating a task flow library;
step S3: the method comprises the steps of obtaining a problem input by a user through a chat window of a visitor end;
step S4: extracting a first trigger corpus corresponding to the problem from a knowledge base;
step S5: extracting a second trigger corpus corresponding to the problem from the task flow library;
step S6: and determining an answer to the question based on the first trigger corpus and the second trigger corpus.
The working principle and the beneficial effects of the technical scheme are as follows:
based on the mutual cooperation of the prefabricated knowledge base and the task flow base, the best answer of the questions proposed by the user is determined, and the accurate intention or the fuzzy intention of the user is fully recognized.
To achieve accurate prefabrication of the knowledge base, in one embodiment, the operations of prefabricating the knowledge base include:
setting a plurality of knowledge points to obtain a common problem set;
setting a reply threshold and a difference threshold;
wherein each common problem in the common problem set comprises:
standard questions, similar questions, answers, and associated questions;
and the standard question and the similar question method are correspondingly provided with first trigger corpora.
To implement the prefabrication of the task flow base, in one embodiment, the prefabrication task flow base comprises:
setting a plurality of task flows and setting a plurality of second trigger corpora for each task flow;
and setting a task reply threshold.
To achieve the obtaining of the first trigger corpus, in one embodiment, the extracting the first trigger corpus corresponding to the question from the knowledge base includes:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
and taking a first trigger corpus of which the first semantic similarity is more than a reply threshold as a first trigger corpus corresponding to the problem.
The working principle and the beneficial effects of the technical scheme are as follows:
the method comprises the steps of preprocessing a problem by means of NLP algorithm, performing word segmentation and the like, performing pairwise calculation on the problem and all trigger corpora of a common problem set Q in a knowledge base to obtain semantic similarity of each trigger corpus, taking similarity data above a threshold value, and sequencing the similarity data from large to small to obtain a semantic similarity list, namely a first list, of the knowledge base.
In one embodiment, a second trigger corpus corresponding to the problem is extracted from the task flow library; the method comprises the following steps:
calculating a second voice similarity of the third trigger corpus and the second trigger corpus;
and taking a second triggering language material with the second voice similarity above the task recovery threshold as a second triggering language material corresponding to the problem.
In one embodiment, determining the answer to the question based on the first trigger corpus and the second trigger corpus comprises:
sequencing the first similarity of the first trigger corpus from large to small to construct a first list, wherein the first list is SQ (SQ-SQ)1，SQ2，…，SQqTherein, SQ1>SQ2，SQ2>SQ3，…，SQq-1>SQq,SQqGreater than a reply threshold;
sorting the second similarity of the second trigger corpus from large to small to construct a second list, wherein the second list is SF ═ SF1，SF2，…，SFfIn which is SF1>SF2，SF2>SF3，…，SFf-1>SFf,SFfGreater than a task reply threshold;
when SF1≥SQ1When it is time, trigger SF preferentially1A corresponding task flow;
when SF1<SQ1Then, according to the first list, calculating the difference between two adjacent first similarities in turn, wherein the calculation formula is as follows:
D(i-1)i＝SQi-1-SQi；
wherein D is(i-1)iThe difference value of the ith-1 st first similarity and the ith first similarity in the first list is obtained; SQi-1The number is the (i-1) th first similarity in the first list; SQiThe ith first similarity in the first list;
when D is present12When the difference value is greater than the threshold value, returning to SQ1Answers to the corresponding knowledge points;
when D is present12When the difference value is smaller than or equal to the difference threshold value, comparing the difference value of two subsequent adjacent first similarity with the difference threshold value, and when the first similarity is larger than the difference threshold value, marking as D(j-1)j(ii) a Return question list (SQ)1，SQ2，…，SQj) (ii) a And after the questions selected by the user based on the question list are obtained, returning corresponding answers.
In one embodiment, when SF1＝SQ1When 1, SF is triggered directly1A corresponding task flow;
when SF1<SQ1When the value is 1, then the SQ is directly recovered1Answers to the corresponding knowledge points;
and when the first trigger corpus and the second trigger corpus are both empty, performing bottom-of-pocket reply through a preset internet chatting library.
In one embodiment, extracting the first trigger corpus corresponding to the question from the knowledge base includes:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
taking a first trigger corpus with the first semantic similarity above a reply threshold;
acquiring a first input record in a first time period preset before a user inputs a problem and/or a second input record input by opening a chat window this time;
extracting the first input record and/or the second input record to obtain a plurality of fourth trigger corpora;
acquiring a correlation problem of a first trigger corpus extracted based on the first semantic similarity;
acquiring a fifth trigger corpus of the associated problem;
calculating a third semantic similarity between the fifth trigger corpus and the fourth trigger corpus;
and correcting the first semantic similarity based on the third semantic similarity, wherein the correction formula is as follows:
Figure BDA0003146319930000091
wherein, SQ'0The corrected first semantic similarity; SQ0The first semantic similarity before correction; SGh,kThe third semantic similarity of the h fourth triggering corpus and the k fifth triggering corpus is obtained; m is the total number of the fourth trigger corpus; n is the total number of the fifth trigger corpus;
and taking the first trigger corpus of which the corrected first semantic similarity is more than a reply threshold as the first trigger corpus corresponding to the problem.
The working principle and the beneficial effects of the technical scheme are as follows:
when the first trigger corpus is extracted, the standard problem and the first similarity of the first trigger corpus corresponding to the similar question method are considered comprehensively, and a fifth trigger corpus of related problems of the first trigger corpus needs to be considered, so that the extraction accuracy is improved. The consideration of the relevant problems can be seen mainly from the previous problems of the user raising the problems, so that the problems are input by traversing a first input record in a first time period preset before the user inputs the problems and/or a second input record input by opening a chat window at this time; and segmenting the input record to extract a problem, extracting the corpus of the problem, and acquiring a fourth trigger corpus.
In one embodiment, the multi-scenario hybrid question-answer recommendation method for the smart client robot further includes:
when the first trigger corpus and the second trigger corpus are both empty, accessing a big data network platform;
sending the question to a plurality of data nodes on line;
receiving answers to be selected for the questions fed back by the data nodes;
calculating fourth semantic similarity between each answer to be selected and other answers to be selected;
acquiring a preset trust coefficient between a node corresponding to a preset answer to be selected and nodes corresponding to other answers to be selected;
acquiring a weight coefficient of a node corresponding to a preset answer to be selected and weight values of nodes corresponding to other answers to be selected;
and determining the effective value of each answer to be selected based on the trust coefficient, the weight value and the fourth voice similarity, wherein the calculation formula of the effective value is as follows:
Figure BDA0003146319930000101
wherein H is an effective value; mu.sjA preset trust coefficient is set between the node corresponding to the answer to be selected and the node corresponding to the jth other answer to be selected; SVjThe fourth semantic similarity between the answer to be selected and the jth other answer to be selected; t is0The weighted value of the node corresponding to the answer to be selected; t isdThe weighted value of the node corresponding to the ith other answer to be selected; tau is1Is a preset first weight coefficient; tau is2Is a preset second weight coefficient; epsilon1A first relation coefficient which is preset and corresponds to the influence of the answer; epsilon2A second relation coefficient which is a preset influence of the corresponding node;
and obtaining the answer to be selected with the maximum effective value as the answer of the question and outputting the answer to the user.
The working principle and the beneficial effects of the technical scheme are as follows:
this is not good for the user's experience if only a bottom-of-pocket reply is made. The problem can be sent to a big data network platform, and the best answer can be obtained by means of the powerful functions of the big data network platform. The best answer is mainly generated by comparing answers of all data nodes of the big data network platform, so that the accuracy of the answer is guaranteed. In addition, the pocket bottom reply can be combined, the pocket bottom reply can be carried out firstly, and the big data network platform is used for seeking help after the pocket bottom reply, so that the user experience is improved; for example: the bottom-pocketed reply can be that 'you are good and the question asked by you is professional, and the professional answers' page prompt switching is carried out, the answer of the big data network platform is determined in the switching process, and the switching is prompted to be completed after the answer is determined, so that the user experience is improved.
In one embodiment, the multi-scenario hybrid question-answer recommendation method for the smart client robot further includes:
obtaining the evaluation of answers to the questions by a preset number of users;
determining the rating value of the evaluation based on a preset evaluation rating template;
determining validity of the answer to the question based on the scoring value, the validity calculation formula being as follows:
Figure BDA0003146319930000111
wherein W is the effectiveness; pcA scoring value for the rating of the c-th user;
Figure BDA0003146319930000112
the credit coefficient is a preset c-th user;
when the validity is greater than a preset first validity threshold value, determining that the answer is valid; when the answer is determined to be valid, carrying out weight value up-regulation on nodes participating in providing the answer on the big data network platform; the up-regulation formula is as follows:
Figure BDA0003146319930000113
wherein, T'hThe adjusted weight value of the h node; t ishThe weight value of the h node before adjustment; rhohDetermining an adjustment index from a preset adjustment index table for the h node according to the adjustment history in a preset second time period; thetaThe amplitude value is a preset adjustment amplitude value; theta0Supplementing value for preset adjustment;
when the adjusted weight value calculated by the up-regulation formula is larger than a preset upper limit value, taking the upper limit value as the adjusted weight value;
when the validity is smaller than a preset second validity threshold value, determining that the answer is invalid; when the answer is determined to be invalid, carrying out weight value down regulation on nodes participating in providing the answer on the big data network platform; the downregulation formula is as follows:
Figure BDA0003146319930000114
wherein, t'sThe weight value of the adjusted s-th node is obtained; t is tsThe weighted value of the s-th node before adjustment is obtained; rhosDetermining an adjustment index from a preset adjustment index table for the s-th node according to the adjustment history in a preset second time period;
and when the adjusted weight value calculated by the down-regulation formula is smaller than the preset lower limit value, taking the lower limit value as the adjusted weight value.
The working principle and the beneficial effects of the technical scheme are as follows:
and adjusting the weight value of the participated node based on the effect of the answer so as to ensure the accuracy of the next answer determination. In addition, the up-regulation amplitude is lower than the down-regulation amplitude by adjusting the supplement value, namely the result that the answer given by the node is not ideal is more serious, the cautious degree of the answer given by the node is improved, and the accuracy of the final answer is ensured. Based on the preset adjustment index table, the same node cannot be adjusted for multiple times within the preset second time period, the influence of the node when an error occurs after the node is adjusted for multiple times is ensured, and the stability of the node for giving an answer is improved.
It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims (10)

1. A multi-scene mixed question-answer recommendation method for an intelligent customer service robot is characterized by comprising the following steps:
prefabricating a knowledge base;
prefabricating a task flow library;
the method comprises the steps of obtaining a problem input by a user through a chat window of a visitor end;
extracting a first trigger corpus corresponding to the problem from the knowledge base;
extracting a second trigger corpus corresponding to the problem from the task flow library;
and determining an answer to the question based on the first trigger corpus and the second trigger corpus.
2. The multi-scenario hybrid question-answer recommendation method for an intelligent customer service robot as claimed in claim 1, wherein the pre-made knowledge base comprises:
setting a plurality of knowledge points to obtain a common problem set;
setting a reply threshold and a difference threshold;
wherein each common problem in the common problem set comprises:
standard questions, similar questions, answers, and associated questions;
the standard question and the similar question method are both provided with the first trigger corpus correspondingly.
3. The multi-scenario hybrid question-answer recommendation method of smart client robot as claimed in claim 2, wherein pre-preparing a task flow library comprises:
setting a plurality of task flows, and setting a plurality of second trigger corpora for each task flow;
and setting a task reply threshold.
4. The multi-scenario hybrid question-answer recommendation method of smart client robot as claimed in claim 3, wherein said extracting the first trigger corpus corresponding to said question from said knowledge base comprises:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
and taking the first triggering language material with the first semantic similarity above the reply threshold as the first triggering language material corresponding to the question.
5. The multi-scenario hybrid question-answer recommendation method of smart client robot as claimed in claim 4, wherein said extracting a second trigger corpus corresponding to said question from said task flow library; the method comprises the following steps:
calculating a second voice similarity of the third trigger corpus and the second trigger corpus;
and taking the second triggering language material with the second voice similarity above the task reply threshold as the second triggering language material corresponding to the question.
6. The method as claimed in claim 5, wherein the determining the answer to the question based on the first trigger corpus and the second trigger corpus comprises:
sorting the first similarity of the first trigger corpus from large to small to construct a first list, wherein the first list is SQ ═ SQ1，SQ2，...，SQqTherein, SQ1＞SQ2，SQ2＞SQ3，…，SQq-1＞SQq，SQqGreater than the reply threshold;
sorting the second similarity of the second trigger corpus from large to small to construct a second list, wherein the second list is SF ═ SF1，SF2，...，SFfIn which is SF1＞SF2，SF2＞SF3，…，SFf-1＞SFf，SFfGreater than the task reply threshold;
when SF1≥SQ1When it is time, trigger SF preferentially1A corresponding task flow;
when SF1＜SQ1Then, according to the first list, calculating the difference between two adjacent first similarities in turn, wherein the calculation formula is as follows:
D(i-1)i＝SQi-1-SQi；
wherein D is(i-1)iThe difference value of the ith-1 st first similarity and the ith first similarity in the first list is obtained; SQi-1The number of the first similarities is the i-1 th in the first list; SQiThe ith first similarity in the first list;
when D is present12Returning to SQ if the difference value is greater than the difference value threshold value1Answers to the corresponding knowledge points;
when D is present12When the difference value is smaller than or equal to the difference value threshold value, comparing the difference value of two subsequent adjacent first similarity degrees with the difference value threshold value, and when the first similarity degree is larger than the difference value threshold value, marking as D(j-1)j(ii) a Return question list (SQ)1，SQ2，...，SQj) (ii) a And after the questions selected by the user based on the question list are obtained, returning corresponding answers.
7. The multi-scenario hybrid question-answer recommendation method of smart client robot as claimed in claim 6, wherein when SF is used1＝SQ1When 1, SF is triggered directly1A corresponding task flow;
when SF1＜SQ1When the value is 1, then the SQ is directly recovered1Answers to the corresponding knowledge points;
and when the first trigger corpus and the second trigger corpus are both empty, performing bottom-of-pocket reply through a preset internet chatting library.
8. The multi-scenario hybrid question-answer recommendation method of smart client robot as claimed in claim 2, wherein said extracting the first trigger corpus corresponding to said question from said knowledge base comprises:
preprocessing the problem to obtain a third trigger corpus;
calculating a first semantic similarity between the third trigger corpus and the first trigger corpus;
taking the first trigger corpus with the first semantic similarity above the reply threshold;
acquiring a first input record in a first time period preset before the user inputs the question and/or a second input record input by opening the chat window this time;
extracting the first input record and/or the second input record to obtain a plurality of fourth trigger corpora;
acquiring the associated problem of the first trigger corpus extracted based on the first semantic similarity;
acquiring a fifth trigger corpus of the associated problem;
calculating a third semantic similarity between the fifth trigger corpus and the fourth trigger corpus;
and correcting the first semantic similarity based on the third semantic similarity, wherein the correction formula is as follows:
Figure FDA0003146319920000031
wherein, SQ'0The corrected first semantic similarity is obtained; SQ0The first semantic similarity before correction; SGh，kThe third semantic similarity between the h-th triggering corpus and the k-th triggering corpus is obtained; m is the total number of the fourth trigger corpus; n is the total number of the fifth trigger corpus;
and taking the first trigger corpus of which the corrected first semantic similarity is more than the reply threshold as the first trigger corpus corresponding to the question.
9. The multi-scenario hybrid question-answer recommendation method of a smart client robot as claimed in claim 1, further comprising:
when the first trigger corpus and the second trigger corpus are both empty, accessing a big data network platform;
sending the question to a plurality of data nodes on line;
receiving answers to be selected for the questions fed back by the data nodes;
calculating fourth semantic similarity between each answer to be selected and other answers to be selected;
acquiring a preset trust coefficient between a node corresponding to the preset answer to be selected and nodes corresponding to other answers to be selected;
acquiring a preset weight coefficient of a node corresponding to the answer to be selected and weight values of nodes corresponding to other answers to be selected;
determining an effective value of each answer to be selected based on the trust coefficient, the weight value and the fourth voice similarity, wherein a calculation formula of the effective value is as follows:
Figure FDA0003146319920000041
wherein H is the effective value; mu.sjA preset trust coefficient is set between the node corresponding to the answer to be selected and the node corresponding to the jth other answer to be selected; SVjObtaining a fourth semantic similarity between the answer to be selected and the jth other answer to be selected; t is0The weighted value of the node corresponding to the answer to be selected; t isdThe weighted value of the node corresponding to the ith other answer to be selected; tau is1Is a preset first weight coefficient; tau is2Is a preset second weight coefficient; epsilon1A first relation coefficient which is preset and corresponds to the influence of the answer; epsilon2A second relation coefficient which is a preset influence of the corresponding node;
and acquiring the answer to be selected with the maximum effective value as the answer of the question and outputting the answer to the user.
10. The multi-scenario hybrid question-answer recommendation method of a smart client robot as claimed in claim 9, further comprising:
obtaining the evaluation of a preset number of answers of the user to the question;
determining the score value of the evaluation based on a preset evaluation score template;
determining validity of an answer to the question based on the scoring value, the validity calculation formula being as follows:
Figure FDA0003146319920000042
wherein W is the effectiveness; pcThe scoring value for the rating of the c-th said user;
Figure FDA0003146319920000051
the credit coefficient is a preset c-th user;
when the validity is larger than a preset first validity threshold value, determining that the answer is valid; when the answer is determined to be valid, the weight value of the node participating in providing the answer on the big data network platform is adjusted upwards; the up-regulation formula is as follows:
Figure FDA0003146319920000052
wherein, T'hThe adjusted weight value of the h-th node; t ishThe weight value of the h node before adjustment is obtained; rhohDetermining an adjustment index from a preset adjustment index table for the h-th node according to the adjustment history in a preset second time period; theta is a preset adjusting amplitude value; theta0To presetThe adjusted supplement value of (1);
when the adjusted weight value calculated by the up-regulation formula is larger than a preset upper limit value, taking the upper limit value as the adjusted weight value;
when the validity is smaller than a preset second validity threshold value, determining that the answer is invalid; when the answer is determined to be invalid, carrying out weight value down regulation on the nodes participating in providing the answer on the big data network platform; the downregulation formula is as follows:
Figure FDA0003146319920000053
wherein, t'sThe adjusted weight value of the s-th node; t is tsThe weight value of the s-th node before adjustment is obtained; rhosDetermining an adjustment index from a preset adjustment index table for the s-th node according to the adjustment history in a preset second time period;
when the adjusted weight value calculated by the down-regulation formula is smaller than a preset lower limit value, taking the lower limit value as the adjusted weight value.
CN202110751237.3A 2021-07-02 2021-07-02 Multi-scene mixed question-answer recommendation method for intelligent customer service robot Active CN113360626B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110751237.3A CN113360626B (en) 2021-07-02 2021-07-02 Multi-scene mixed question-answer recommendation method for intelligent customer service robot

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110751237.3A CN113360626B (en) 2021-07-02 2021-07-02 Multi-scene mixed question-answer recommendation method for intelligent customer service robot

Publications (2)

Publication Number Publication Date
CN113360626A true CN113360626A (en) 2021-09-07
CN113360626B CN113360626B (en) 2022-02-11

Family

ID=77538002

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110751237.3A Active CN113360626B (en) 2021-07-02 2021-07-02 Multi-scene mixed question-answer recommendation method for intelligent customer service robot

Country Status (1)

Country Link
CN (1) CN113360626B (en)

Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101373532A (en) * 2008-07-10 2009-02-25 昆明理工大学 FAQ Chinese request-answering system implementing method in tourism field
KR20170112102A (en) * 2016-03-30 2017-10-12 경기대학교 산학협력단 Apparatus and meyhod for large scale parallel temporal reasoning
CN110019149A (en) * 2019-01-30 2019-07-16 阿里巴巴集团控股有限公司 A kind of method for building up of service knowledge base, device and equipment
CN110532348A (en) * 2019-09-04 2019-12-03 网易（杭州）网络有限公司 Question and answer are to the generation method of data, device and electronic equipment
CN110795542A (en) * 2019-08-28 2020-02-14 腾讯科技（深圳）有限公司 Dialogue method and related device and equipment
US20200059558A1 (en) * 2017-12-13 2020-02-20 Genesys Telecommunications Laboratories, Inc. Systems and methods for chatbot generation
CN111026855A (en) * 2019-12-06 2020-04-17 易小博(武汉)科技有限公司 Intelligent customer service response method, system, controller and medium
CN111159363A (en) * 2018-11-06 2020-05-15 航天信息股份有限公司 Knowledge base-based question answer determination method and device
CN111625632A (en) * 2020-04-17 2020-09-04 北京捷通华声科技股份有限公司 Question-answer pair recommendation method, device, equipment and storage medium
CN112527965A (en) * 2020-12-18 2021-03-19 国家电网有限公司客户服务中心 Automatic question answering implementation method and device based on combination of professional library and chatting library
CN112632242A (en) * 2020-12-17 2021-04-09 竹间智能科技（上海）有限公司 Intelligent conversation method and device and electronic equipment
CN112988997A (en) * 2021-03-12 2021-06-18 中国平安财产保险股份有限公司 Response method and system of intelligent customer service, computer equipment and storage medium

Patent Citations (12)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101373532A (en) * 2008-07-10 2009-02-25 昆明理工大学 FAQ Chinese request-answering system implementing method in tourism field
KR20170112102A (en) * 2016-03-30 2017-10-12 경기대학교 산학협력단 Apparatus and meyhod for large scale parallel temporal reasoning
US20200059558A1 (en) * 2017-12-13 2020-02-20 Genesys Telecommunications Laboratories, Inc. Systems and methods for chatbot generation
CN111159363A (en) * 2018-11-06 2020-05-15 航天信息股份有限公司 Knowledge base-based question answer determination method and device
CN110019149A (en) * 2019-01-30 2019-07-16 阿里巴巴集团控股有限公司 A kind of method for building up of service knowledge base, device and equipment
CN110795542A (en) * 2019-08-28 2020-02-14 腾讯科技（深圳）有限公司 Dialogue method and related device and equipment
CN110532348A (en) * 2019-09-04 2019-12-03 网易（杭州）网络有限公司 Question and answer are to the generation method of data, device and electronic equipment
CN111026855A (en) * 2019-12-06 2020-04-17 易小博(武汉)科技有限公司 Intelligent customer service response method, system, controller and medium
CN111625632A (en) * 2020-04-17 2020-09-04 北京捷通华声科技股份有限公司 Question-answer pair recommendation method, device, equipment and storage medium
CN112632242A (en) * 2020-12-17 2021-04-09 竹间智能科技（上海）有限公司 Intelligent conversation method and device and electronic equipment
CN112527965A (en) * 2020-12-18 2021-03-19 国家电网有限公司客户服务中心 Automatic question answering implementation method and device based on combination of professional library and chatting library
CN112988997A (en) * 2021-03-12 2021-06-18 中国平安财产保险股份有限公司 Response method and system of intelligent customer service, computer equipment and storage medium

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
MACEDO R D 等: "An Intelligent Robotic Engine Using Digital Repository of the DSpace Platform", 《THE SIXTH INTERNATIONAL CONFERENCE ON INFORMATION, PROCESS, AND KNOWLEDGE MANAGEMENT》 *
YINIZHILIANLOVE: "ACL | 自然语言处理（NLP）之多轮对话建模优化（Utterance ReWriter)", 《HTTPS://BLOG.CSDN.NET/YINIZHILIANLOVE/ARTICLE/DETAILS/98400767》 *
廖美红: "大数据知识图谱的电商领域问答***设计", 《商场现代化》 *
高旭杨: "基于语句相似度计算的FAQ问答***设计", 《中国优秀博硕士学位论文全文数据库(硕士)信息科技辑》 *

Also Published As

Publication number Publication date
CN113360626B (en) 2022-02-11

Similar Documents

Publication Publication Date Title
US11631147B2 (en) Social recommendation method based on multi-feature heterogeneous graph neural networks
Peng et al. Composite task-completion dialogue policy learning via hierarchical deep reinforcement learning
CN108804689B (en) Question-answering platform-oriented label recommendation method integrating user hidden connection relation
CN110020426B (en) Method and device for distributing user consultation to customer service group
CN108351986A (en) Learning system, learning device, learning method, learning program, training data generating means, training data generation method, training data generate program, terminal installation and threshold value change device
CN108021934B (en) Method and device for recognizing multiple elements
CN109783617A (en) For replying model training method, device, equipment and the storage medium of problem
CN109063035A (en) A kind of man-machine more wheel dialogue methods towards trip field
CN110321291A (en) Test cases intelligent extraction system and method
Chen et al. Agent-aware dropout dqn for safe and efficient on-line dialogue policy learning
CN105868179A (en) Intelligent asking-answering method and device
CN113420058B (en) Conversational academic conference recommendation method based on combination of user historical behaviors
CN110019688A (en) The method that robot is trained
CN116757652B (en) Online recruitment recommendation system and method based on large language model
CN111626616A (en) Crowdsourcing task recommendation method
CN111309887A (en) Method and system for training text key content extraction model
CN110008332A (en) The method and device of trunk word is extracted by intensified learning
CN106776701A (en) The problem recommended for article determines method and apparatus
CN110689359A (en) Method and device for dynamically updating model
CN108595609A (en) Generation method, system, medium and equipment are replied by robot based on personage IP
CA3064116A1 (en) Systems and methods for performing automated interactive conversation with a user
CN110263232A (en) A kind of mixed recommendation method based on range study and deep learning
CN110059167A (en) Natural expression processing method, response method, equipment and the system of natural intelligence
CN113360626B (en) Multi-scene mixed question-answer recommendation method for intelligent customer service robot
CN109190116A (en) Semantic analytic method, system, electronic equipment and storage medium

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