DE102021101956A1 - Automatically extracting a knowledge of a knowledge giver - Google Patents

Abstract

Method for extracting knowledge from a knowledge provider, in which a chatbot transmits a question to a knowledge provider providing extractable knowledge and the chatbot receives an answer entered by the knowledge provider to the transmitted question as the extracted knowledge, device for extracting knowledge from a knowledge provider and computer program product for extracting a knowledge of a knowledge giver.

Description

The invention relates to a method for extracting knowledge from a knowledge provider, in which a chatbot transmits a question to a knowledge provider providing extractable knowledge, and the chatbot receives an answer entered by the knowledge provider to the transmitted question as the extracted knowledge. Furthermore, the invention relates to a device for extracting knowledge from a knowledge provider and a computer program product for extracting knowledge from a knowledge provider.

Chatbots are known in various configurations and are used to enable a computer to communicate automatically with a person using a natural language, i.e. to simulate a human interlocutor for the person. A chatbot is a program, for example a software program, which is stored or installed on the computer for this purpose.

Typically, a chatbot receives a natural language utterance from a person and transmits a natural language utterance to the person in response to the person's natural language utterance. The chatbot typically includes a human-machine interface (HMI) for receiving and transmitting natural language utterances.

The human-machine interface is configured for the person to input the natural language utterance, for example by typing text of the natural language utterance letter-by-letter into an editor of the human-machine interface. Further, the man-machine interface is configured to output the natural language utterance of the chatbot to the person, for example by displaying a text of the natural language utterance of the chatbot on a display screen of the man-machine interface.

The natural language utterances can be questions or answers. In one embodiment, the chatbot receives a question entered by the person and the chatbot transmits an answer to the transmitted question to the person. In an alternative embodiment of the chatbot, the chatbot transmits a question about the person and receives an answer entered by the person to the transmitted question.

In the first embodiment, the person can be referred to as a knowledge taker who extracts extractable knowledge provided by the chatbot. In the second embodiment, the person can be referred to as a knowledge provider who provides knowledge that can be extracted by the chatbot. Accordingly, the chatbot takes on a complementary role. Of course, in a further embodiment, the chatbot can be configured to act both as a knowledge giver and as a knowledge recipient and to switch flexibly between the two roles depending on the course of the conversation.

However, the chatbot is unsuitable for extracting the knowledge of the knowledge giver by a knowledge receiver when the knowledge giver and the knowledge receiver are different people from each other. Such constellations, in particular with a plurality of knowledge providers and/or a plurality of knowledge recipients, are often encountered in practice, particularly in companies.

In order to enable knowledge to be extracted efficiently even in complex constellations, those questions which are to be transmitted to the knowledge providers and answered by the knowledge providers must be determined from a plurality of questions provided by a plurality of knowledge recipients. In addition, a sequence of the specific questions and an assignment of each specific question to one or more knowledge providers must be specified.

So far, the extraction of the knowledge provided by at least one knowledge provider by at least one knowledge recipient has been possible in simple constellations only by means of another person acting as a so-called moderator. However, the extraction of knowledge mediated by a moderator is expensive because of the personnel expenditure and tends to decrease in efficiency or reliability with increasing complexity.

The invention is therefore based on the object of proposing a method for extracting knowledge from a knowledge provider, which can be carried out automatically and extracts the knowledge of the knowledge provider efficiently and reliably. Furthermore, it is an object of the invention to provide a device for extracting a knowledge of a knowledge provider and a computer program product for extracting a knowledge of a knowledge provider.

One subject of the invention is a method for extracting knowledge from a knowledge provider, in which a chatbot transmits a question about a knowledge provider providing extractable knowledge, and the chatbot transmits one of the knowledge sensgeber input response to the transmitted question receives as the extracted knowledge. In other words, the chatbot asks the question to the knowledge giver, and the knowledge giver answers the question asked to the chatbot. The chatbot thus acts as a knowledge taker. The chatbot can be implemented using the Rasa chatbot platform or another chatbot platform different from Rasa.

In the method according to the invention, a moderator module communicating with the chatbot determines the question and transmits the specific question to the chatbot. The moderator module controls the chatbot fully automatically and is designed as a program, for example a software program. The moderator module can be a program integrated into the chatbot or a program separate from the chatbot, which communicates with the chatbot by means of mutually corresponding interfaces of the moderator module and the chatbot.

The chatbot controlled by the moderator module acts as a moderator between the knowledge giver and a knowledge recipient. By determining the question to be transmitted to the knowledge provider, the moderator module can also specify a sequence of several questions to be transmitted and an assignment of each question to be transmitted to a knowledge provider.

In short, the moderator module extends the chatbot into a fully automated moderator with the ability to efficiently and reliably extract the extractable knowledge provided by the knowledge provider. The enhanced moderator can replace a person as the moderator, with comparatively lower knowledge extraction costs and higher knowledge extraction efficiency and reliability.

The moderator module preferably calculates a relevance value for each question-answer pair comprising a question and an answer from a plurality of stored question-answer pairs and assigns the calculated relevance value to each stored question-answer pair and determines the question as a question one Question-answer pairs with the highest assigned relevance value. Each stored question-and-answer pair includes an empty answer if the question is unanswered, or an answer that the chatbot provides before determining the question from the knowledge provider or another knowledge provider as an answer to the question comprised by the question-answer pair received, i.e. when the question has been answered. In short, each stored question-answer pair represents knowledge extracted prior to determining the question. The calculated relevance value can be understood as a probability that the respective question will be determined for transmission to the knowledge provider.

The higher the relevance value of a question-answer pair, the greater the probability that the moderator module determines the question of the question-answer pair as the question to be transmitted to the knowledge provider. The relevance value can be in a range from 0 to 1, for example, a question with an assigned relevance value 0 not being determined and a question with an assigned relevance value 1 being primarily determined. The stored question-answer pairs are in the storage area at the same time. As a result, the moderator module can easily and immediately compare relevance values of all stored question-answer pairs in order to determine the (next) question to be transmitted to the knowledge provider, i.e. to select from the questions comprised by the stored question-answer pairs.

The moderator module advantageously calculates the relevance value depending on the plurality of stored question-answer pairs. The moderator module accesses a memory area which includes the plurality of question-answer pairs, i.e. in which the plurality of question-answer pairs are stored.

Ideally, the moderator module calculates a similarity of an answer of a stored question-answer pair with an answer of another stored question-answer pair and calculates the relevance value assigned to the question-answer pair lower if the calculated similarity of the answer with each further answer is one stored question-answer pair is low, and calculates the relevance value assigned to the question-answer pair higher if the calculated similarity of the answer to another answer of a stored question-answer pair is high. The relevance value assigned to a stored question-answer pair depends on a similarity of the answer to answers comprised by other stored question-answer pairs and is correlated with the similarity.

The more similar the answers comprised by stored question-answer pairs are, the less specific they are. The respective questions for transmission to the knowledge provider should be determined accordingly. Non-specific answers include, in particular, blank answers, non-answers such as "I have no idea" or answers with generalities such as "It depends."

The more dissimilar an answer contained in one question-answer pair is to all the answers contained in other question-answer pairs, the more specific it is. Accordingly, the respective question does not have to be determined for transmission to the knowledge provider. In particular, the specific answers include answers with special technical terms that fit the field of the question of the question-answer pair.

This ensures that knowledge that has already been extracted is not extracted again. The efficiency of the knowledge extraction is higher, the more other, i.e. not yet extracted, knowledge is extracted by the knowledge provider per question.

The moderator module can determine a vector from a text of an answer of a question-answer pair and calculate the similarity of answers depending on a vector difference of the two determined vectors smaller when the vector difference is large and larger when the vector difference is small. To determine the vector of an answer of a question-answer pair, the moderator module can concatenate the text of the answer of the question-answer pair with a text of the question of the question-answer pair to form a text of the question-answer pair. A vector difference can be calculated particularly easily by the moderator module. A small vector difference corresponds to a small textual, in particular semantic, distance, a large vector difference corresponds to a large textual distance between the respective answer and other answers.

In one embodiment, the moderator module calculates each vector using a natural language understanding algorithm implemented by the moderator module, or the moderator module determines each vector difference using an artificial intelligence implemented by the moderator module. As a natural language understanding (NLU) algorithm, the moderator module can use the bag-of-words algorithm, paragraph2vec or another known NLU algorithm from the document classification. As artificial intelligence (AI; artificial intelligence, AI), the moderator module can use a neural network or another known AI algorithm.

In preferred embodiments, the chatbot transmits the received response to a knowledge seeker, receives a quality value entered by the knowledge seeker relating to the transmitted response, and transmits the received quality value to the moderator module. The quality value represents the knowledge recipient's satisfaction with the answer given by the knowledge provider. The quality value indicates how accurately or completely the knowledge provider has answered the question. For example, the quality value can be entered by the knowledge taker as a number of stars, where zero stars can mean an unusable quality of the answer and five stars can mean an optimal quality of the answer.

The moderator module can then assign the transmitted quality value to the stored question-answer pair and store the transmitted quality value with the question-answer pair as its quality value. In this way, the extracted knowledge represented by the question-answer pair is differentiated with regard to its usefulness. Furthermore, the moderator module can weight the quality value and, through the weighting, take account of the reliability, i.e. the evaluation competence, of the knowledge taker.

In these embodiments, the moderator module calculates the relevance score depending on the quality score of the question-answer pair, lower when the quality score is high and higher when the quality score is low. For example, the moderator module can weight a calculated similarity with the quality value. The moderator module can nevertheless calculate a very high relevance value for a question which is very similar to a stored question-answer pair because of a low quality value assigned to the question-answer pair, or for a question which corresponds to a stored question-answer pair is rather dissimilar, calculate a reduced relevance value because of a high quality value assigned to the question-answer pair.

The moderator module preferably determines the question earlier if the calculated relevance value is high and later if the calculated relevance value is low, or only if the calculated relevance value exceeds a predetermined relevance threshold. The moderator module can exclude questions with a relevance below the relevance threshold from being transmitted to the knowledge provider. If more relevant questions are transmitted to the knowledge provider before less relevant questions, the knowledge extracted per question is high at the beginning of the procedure and decreases as the procedure progresses.

In particular, this makes it more likely that an answer with a high quality value to a question with a high relevance value will be stored as a question-answer pair during the method. This question-answer pair can lead to lower relevance values of remaining stored questions and in this way result in fewer questions being transmitted to the knowledge provider overall. In this way, an efficiency of the knowledge extraction is increased.

In many embodiments, the chatbot receives a question entered by a knowledge taker and submits the received question to the moderator module, and the moderator module stores the submitted question along with a blank answer as a question-answer pair. In this embodiment, the questions are provided by the knowledge seeker. In this way, the knowledge recipient can extract the extractable knowledge provided by the knowledge provider. In this constellation, the chatbot communicating with the moderator module moderates the knowledge extraction by the knowledge seekers.

Alternatively or additionally, the moderator module receives a question entered by an administrator as a default question and stores the received question along with an empty answer as a question-answer pair. Each standard question need not be asked by a knowledge seeker, thereby increasing knowledge extraction efficiency. Furthermore, each standard question can ensure that a certain knowledge of the knowledge provider is extracted in any case, i.e. independently of a knowledge recipient, whereby the reliability of the knowledge extraction is increased.

In further embodiments, the chatbot transmits the received response to the moderator module and the moderator module stores the question and the transmitted answer as a question-answer pair. The saved question-answer pairs represent the previously extracted knowledge of the knowledge provider. If a question from a knowledge seeker has a vanishing or low relevance below the relevance threshold due to great similarity with a stored question-answer pair that is assigned a high quality value, the knowledge seeker can use the answer of the stored question-answer pair as an answer to his question to be submitted. The knowledge giver is not questioned again, whereby an efficiency of the knowledge extraction is increased.

The chatbot can receive questions from a plurality of knowledge seekers and/or transmit specific questions to a plurality of knowledge providers. In other words, the chatbot also takes over moderation between several knowledge takers and several knowledge providers, i.e. in complex constellations of knowledge extraction in which a person as a moderator would be extremely challenged or even overwhelmed. Consequently, the method expands the possible applications of knowledge extraction to complex constellations.

Another subject matter of the invention is a device for extracting knowledge from a knowledge provider, which comprises a computer and a chatbot installed in the computer. Devices of this type are suitable, for example, for polls or surveys in which opinions or information are to be extracted as extractable knowledge provided by knowledge providers by means of predetermined questions.

The device according to the invention also includes a moderator module installed in the computer and communicating with the chatbot via an interface, and is configured to carry out a method according to the invention. The moderator module turns the chatbot into a fully automatic moderator, which works more cost-effectively, efficiently and reliably than a person as a moderator.

Another subject matter of the invention is a computer program product for extracting knowledge from a knowledge provider, with a storage medium and a program which is stored in the storage medium and implements a chatbot. Such computer program products implement known chatbot platforms.

In the case of the computer program product according to the invention, the program also implements a moderator module which communicates with the chatbot and causes a computer to carry out a method according to the invention when it is carried out by a processor of the computer. The computer program product turns a computer into a fully automatic moderator.

An essential advantage of the method according to the invention is that a person is not necessary as a moderator of a knowledge extraction and the knowledge extraction takes place more cost-effectively, efficiently and reliably thanks to a fully automatic moderator. A further advantage can be seen in the fact that knowledge extraction is easily possible even in complex constellations with a plurality of knowledge providers and/or knowledge recipients. Finally, it is advantageous that in order to carry out the method it is only necessary to provide an existing chatbot platform with a moderator module.

It goes without saying that the features mentioned above and those still to be explained below can be used not only in the combination specified in each case, but also in other combinations or on their own, without departing from the scope of the present invention.

• 1 ein Blockdiagramm einer Vorrichtung nach einer Ausführungsform der Erfindung zum Extrahieren eines Wissens eines Wissensgebers;
• 2 in einem Flussdiagramm ein Verfahren nach einer Ausführungsform der Erfindung zum Extrahieren eines Wissens eines Wissensgebers, welches von der in 1 gezeigten Vorrichtung ausgeführt wird.

The invention is shown schematically in the drawings using an exemplary embodiment and is described in detail below with reference to the drawings. It shows

• 1 FIG. 14 is a block diagram of a device according to an embodiment of the invention for extracting a knowledge of a knowledge provider;
• 2 in a flowchart a method according to an embodiment of the invention for extracting a knowledge of a knowledge provider, which of the in 1 shown device is running.

1 1 shows a block diagram of a device 1 according to an embodiment of the invention for extracting a knowledge 42 from a knowledge provider 40. The device 1 comprises a computer 2, a chatbot 10 installed in the computer 2 and a chatbot 10 installed in the computer 2 and connected to the chatbot 10 a moderator module 20 communicating with an interface 11. The device 1 is configured to carry out the method described below and can be suitable for extracting the knowledge 42 of the knowledge provider 40 by a knowledge recipient 30.

The computer 2 includes a processor 3 and can include a computer program product 4 for extracting the knowledge 42 of the knowledge provider 40 . The computer program product 4 comprises a storage medium 5 and a program 6 stored in the storage medium 5. The program 6 implements the chatbot 10 and the moderator module 20 communicating with the chatbot 10 and causes the computer 2 to execute the method described below when it is processed by the processor 3 of the computer 2 is running.

2 shows in a flow chart a method according to an embodiment of the invention for extracting the knowledge 42 of the knowledge provider 40, which is derived from the in 1 shown device 1 is executed.

In the method, the chatbot 10 can receive a question 31 entered by the knowledge taker 30 and transmit it to the moderator module 20 . The moderator module 20 can store the submitted question 31 in a storage area along with an empty answer as a question-answer pair 70 to which the moderator module 20 has access. Alternatively or additionally, the moderator module 20 may receive a question 51 entered by an administrator 50 as a default question and also store the received question 51 in the storage area along with an empty answer as a question-answer pair 70 .

The moderator module 20 communicating with the chatbot 10 determines the question 21 and transmits the specific question 21 to the chatbot 10. Ideally, the moderator module 20 calculates for each question-answer pair 70 comprising a question 71 and an answer 72 from a plurality of stored questions -Answer pairs 70 a relevance value 710, assigns the calculated relevance value 710 to each stored question-answer pair 70 and determines question 21 as a question 71 of a question-answer pair 70 with the highest assigned relevance value 710.

The moderator module 20 advantageously calculates the relevance value 710 depending on the plurality of stored question-answer pairs 70.

To calculate the relevance value 710, the moderator module preferably calculates a similarity between an answer 72 of a stored question-answer pair 70 and an answer 72 of another stored question-answer pair 70 and calculates the relevance value 710 assigned to the question-answer pair 70 lower, if the calculated similarity of the answer 72 with every further answer 72 of a stored question-answer pair 70 is low, and calculates the relevance value 710 assigned to the question-answer pair 70 higher if the calculated similarity of the answer 72 with a further answer 72 of a stored question-answer pair is 70 in size.

To calculate the similarity, the moderator module 20 determines a vector from a text of an answer 72 of a question-answer pair 70 and calculates the similarity of answers 72 depending on a vector difference of the two determined vectors, smaller if the vector difference is large and larger if the vector difference is small.

Favorably, the moderator module 20 determines each vector using an algorithm implemented by the moderator module 20 for understanding a natural language and/or the moderator module 20 calculates each vector difference using an artificial intelligence implemented by the moderator module 20 .

The chatbot 10 then transmits the question 21 to the knowledge provider 40 providing the extractable knowledge 42.

In a further step, the chatbot 10 receives the answer 41 entered by the knowledge provider 40 to the transmitted question 21 as the extracted knowledge 42.

The chatbot 10 can transmit the received answer 41 to the moderator module 20 and the moderator module 20 can store the question 21 and the transmitted answer 41 as a question-answer pair 70 .

The chatbot 10 advantageously transmits the received response 41 to a knowledge taker 30, receives a quality value 32 entered by the knowledge taker 30 relating to the transmitted answer 41 and transmits the received quality value 32 to the moderator module 20. Furthermore, the moderator module 20 can assign the transmitted quality value 32 to the stored question-answer pair 70 and compare it with the question Answer pair 70 as its quality value 720, ie as a quality value 720 of the question-answer pair 70 store.

Advantageously, the moderator module 20 calculates the relevance value 710 depending on the quality value 720 of the question-answer pair 70, lower when the quality value 720 is high and higher when the quality value 720 is low.

Overall, moderator module 20 may determine question 21 earlier when calculated relevance value 710 is high and determine question 21 later when calculated relevance value 710 is low. Alternatively or additionally, the moderator module 20 can only determine the question 21 if the calculated relevance value 710 exceeds a predetermined relevance threshold.

Of course, in the method, the chatbot 10 can receive questions 31 from a plurality of knowledge recipients 30 and/or transmit specific questions 21 to a plurality of knowledge providers 40 or assign a specific question 21 to one or more knowledge providers 40 .

The method described can be carried out iteratively and/or asynchronously over a period of time, for example several weeks, i.e. during the time period, knowledge seekers 30 repeatedly enter questions 31 and knowledge providers 40 repeatedly enter answers 41 to the entered questions 31.

Within the scope of the method described, the moderator module 20 can suggest a question to be asked to the knowledge seekers 30 based on stored question-answer pairs 70, for example if a quality value 720 of a question-answer pair is too small. Likewise, moderator module 20 can suggest an answer to be given to knowledge providers 40 based on stored question-answer pairs 70 , for example if specific question 21 is very similar to a question 71 of a stored question-answer pair 70 . In this case, the answer 72 of the stored question-answer pair 70 can be considered as the answer to the specific question 21 . Of course, both the knowledge recipients 30 and the knowledge providers 40 can ignore suggested questions or suggested answers or modify them in a conventional way using an editor of the chatbot 10 .

Reference List

1

contraption

2

calculator

3

processor

4

computer program product

5

storage medium

6

program

10

chatbot

11

interface

20

moderator module

21

question

30

knowledge taker

31

question

32

quality value

40

knowledge giver

41

answer

42

knowledge

50

Administrator

51

question

70

question and answer pair

71

question

72

answer

710

relevance value

720

quality value

Claims (15)

A method for extracting knowledge (42) from a knowledge provider (40), in which - A chatbot (10) transmits a question (21) to an extractable knowledge (42) providing knowledge provider (40); - The chatbot (10) receives an answer (41) entered by the knowledge provider (40) to the transmitted question (21) as the extracted knowledge (42); - A moderator module (20) communicating with the chatbot (10) determines the question (21) and transmits it to the chatbot (10). procedure after claim 1 , in which the moderator module (20) calculates a relevance value (710) for each question-answer pair (70) comprising a question (71) and an answer (72) from a plurality of stored question-answer pairs (70), each stored question-answer assigns the calculated relevance value (710) to the pair (70) and determines the question (21) as a question (71) of a question-answer pair (70) with the highest assigned relevance value (710). procedure after claim 2 , in which the moderator module (20) calculates the relevance value (710) depending on the plurality of stored question-answer pairs (70). procedure after claim 3 , In which the moderator module (20) calculates a similarity of an answer (72) of a stored question-answer pair (70) with an answer (72) of a further stored question-answer pair (70) and the question-answer pair (70) assigned relevance value (710) lower if the calculated similarity of the answer (72) with each further answer (72) of a stored question-answer pair (70) is low, and the question-answer pair ( 70) associated relevance value (710) calculated higher if the calculated similarity of the answer (72) with a further answer (72) of a stored question-answer pair (70) is high. procedure after claim 4 , in which the moderator module (20) determines a vector from a text of an answer (72) of a question-answer pair (70) and calculates the similarity of answers (72) as a function of a vector difference between the two determined vectors, if the vector difference is large, and calculated larger when the vector difference is small. procedure after claim 5 , in which the moderator module (20) determines each vector by means of an algorithm implemented by the moderator module (20) for understanding a natural language and/or calculates each vector difference by means of an artificial intelligence implemented by the moderator module (20). Procedure according to one of claims 2 until 6 , in which the chatbot (10) transmits the received response (41) to a knowledge recipient (30), receives a quality value (32) entered by the knowledge recipient (30) and relates to the transmitted response (41) and the received quality value (32). transmitted to the moderator module (20), and in which the moderator module (20) assigns the transmitted quality value (32) to the stored question-answer pair (70) and stores it with the question-answer pair (70) as its quality value (720). . procedure after claim 7 , in which the moderator module (20) calculates the relevance value (710) as a function of the quality value (720) of the question-answer pair (70) lower if the quality value (720) is high, and calculates it higher if the quality value (720 ) is low. Procedure according to one of claims 2 until 8th , in which the moderator module (20) determines the question (21) earlier if the calculated relevance value (710) is high and determines the question (21) later if the calculated relevance value (710) is low and/or the moderator module ( 20) determining the question (21) only if the calculated relevance value (710) exceeds a predetermined relevance threshold. Procedure according to one of Claims 1 until 9 , in which the chatbot (10) receives a question (31) entered by a knowledge taker (30) and transmits it to the moderator module (20) and the moderator module (20) uses the transmitted question (31) together with an empty answer (72). stores a question and answer pair (70). Procedure according to one of Claims 1 until 10 wherein the moderator module (20) receives a question (51) entered by an administrator (50) as a default question and stores the received question (51) along with an empty answer (72) as a question-answer pair (70). . Procedure according to one of Claims 1 until 11 , in which the chatbot (10) transmits the received answer (41) to the moderator module (20) and the moderator module (20) stores the question (21) and the transmitted answer (41) as a question-answer pair (70). . Procedure according to one of Claims 1 until 12 , In which the chatbot (10) receives questions (31) from a plurality of knowledge recipients (30) and/or transmits specific questions (21) to a plurality of knowledge providers (40). Device (1) for extracting knowledge (42) from a knowledge provider (40), which has a computer (2), a chatbot (10) installed in the computer (2) and a chatbot (10) installed in the computer (2) and connected to the chatbot ( 10) via an interface (11) communicating moderator module (20) comprises and is configured, a method according to one of Claims 1 until 13 to execute. Computer program product (4) for extracting knowledge (42) from a knowledge provider (40), with a storage medium (5) and a chatbot (10) stored in the storage medium (5) and a moderator module (20) communicating with the chatbot (10). ) implementing program (6), which causes a computer (2), a method according to one of Claims 1 until 13 execute when it is executed by a processor (3) of the computer (2).

Images