EP3652664A1

EP3652664A1 - Method for conducting dialog between human and computer

Info

Publication number: EP3652664A1
Application number: EP17780297.2A
Authority: EP
Inventors: Philipp Heltewig; Sascha Poggemann
Original assignee: Cognigy GmbH
Current assignee: Cognigy GmbH
Priority date: 2017-07-14
Filing date: 2017-07-14
Publication date: 2020-05-20
Also published as: US20200234700A1; WO2019011356A1; US11315560B2

Abstract

The invention relates to a method for conducting dialog between human and computer by means of a self-learning system, comprising: receiving user inputs in natural spoken text, associating synonyms and key words and making word associations, analyzing the user inputs with respect to sentence structure and sentence construction and recognizing same, associating key phrases, determining and classifying the user intention, checking if a confirmation is required, carrying out logical processing in a reasoning-and-answer preparation with formulation of a decision on the further dialog development, formulation or the dialog course with possible generation of an answer, wherein in order to recognize the user intention the statistical relevance of the association with a target intention is determined while at the same time all meta-information is retained, the classification occurs by means of a classification algorithm for intention clarification, logic and speech understanding are linked by means of meta-information in the decision tree, feedback to the self-learning system is generated by using earlier user questioning, a query to the user as to whether the recognized intention was correctly understood is generated if the statistical relevance is below a limit value, in this case an answer of the user is received and is then analyzed with respect to sentence structure and sentence construction and the user inputs are again divided into substantive content and into meta-information about the user and the user intention is again determined and classified, the recognition result is output to the self-learning system, the self-learning system develops automatic system recommendations and/or decisions for improving the intention recognition across a plurality of user interactions and keeps same at hand for later use, initiates the generation of the spoken-text answer to the user if the statistical relevance exceeds a limit value, and otherwise triggers a further inquiry.

Description

Procedure for dialogue between human and computer

The present method relates to an implementation on a computer that also includes a mobile phone, smartphone or other portable device having a digital

Computing unit may be that allows the user to conduct a natural, everyday-suitable conversation with a computer system. Through text or voice input, various steps for classification and subsequent speech comprehension are performed, whereupon context-based in the use case, a conversation between the user and the

Computer system is enabled. The special emphasis is placed on the recognition of user intention and their further processing. For this purpose, the method recognizes the meaning of the language.

Computer systems that deal with spoken or written natural language have long been state of the art. So there are dictation systems that can recognize spoken words and punctuation and translate into written text. On the other hand, there are a variety of reading systems that can announce a given text. These systems are unable to understand the meaning of the language.

Also navigation devices in vehicles are usually able to give the driver who has entered a destination instructions in spoken language for driving. But they do not know if the user understood the instructions. Even these systems can not understand the user, but are dependent on his destination input. From the start, they only allow targets that are stored in the system-immanent database.

For some time, sophisticated computer systems have been available that can analyze spoken questions and provide answers in spoken language. These are simple queries whose completed responses are stored in databases, and where the computer system is merely faced with the task of finding and speaking the correct one of these answers. The computer system does not really have to understand the meaning of the question, it is enough if it recognizes certain keywords by which the set of possible answers can be narrowed down so that the answer with the greatest numerical match of the keywords can be output in response.

For example, US Pat. No. 6,964,023 B2 describes a system for multimodal user conversions that is capable of converting spoken speech into text, resolving ambiguities and recognizing the mood of the user, for example by means of face recognition. If necessary, the system will ask questions to the user to clarify his instructions. In doing so, the system keeps asking until it detects the intent of the user. After that it leads System corresponding instructions, for example, it turns on the radio. The recognition process is based on face and speech recognition. An error probability is determined for each branch of the decision tree, both by comparison with a threshold and by linear regression. The decision trees are trained for word-dependent properties.

WO 02/073331 A2 provides a system architecture for commercial transactions in natural language in which users can inform themselves about products of a catalog. The system accesses a database for products and a database of user preferences and asks questions for clarification. The database of user preferences has an update function, but a self-learning function is missing.

US 8,600,747 B2 describes a method for managing a dialogue, which may also include spoken language, such as in the context of a telephone call. For this purpose, a large number of dialogue motivators is used, which is an error handling, a

Ambiguity treatment, acceptance and confirmation, missing information and

Continue to perform. Hereby order and payment transactions are possible, the system can ask questions to the user. However, it is not self-learning and primarily serves to transfer users to human dialogue partners.

DE 60 2004 006 641 T2 describes an audio dialogue system that has a

Speech recognition and an audio interface as well as a browser with access to content and determines activation phrases. By applying a latent semantic analysis, the similarity between the user utterances and the

Activation phrases determined as a numerical value. In the couplings of

Activation phrases and the contents of a language model is trained. A recognition of intention does not take place.

US 8,543,407 B1 describes a system which uses statistical modeling and formal logic to evaluate conversational input regarding data contents such as commands and dictations. For this purpose, a command dictionary and a dynamic grammar for identification, the clarification of ambiguity and the extraction of

Commands used, the system works context sensitive and can be more

Address user interfaces. However, it can not recognize user intentions and can not engage in any dialogue with the user.

WO 2014/047270 A1 describes a method and an apparatus for achieving the intent in an IVR - Interactive voice response - using natural language

predict and solves the problem that the general terms used by the user are subject to change during concretisation, which makes it a system make it difficult to ensure their consistency. The intention of the user is compared with earlier intentions and with the usual normal behavior of average users and estimated in its relative value. Depending on this value, value-based offers corresponding to preset threshold values are submitted to the user. Intentional recognition thus becomes a generation of intent.

US 8,849,652 B2 describes a system that includes voice-based and non-voice-based interfaces and user context, previous information and

Use user data to create a natural user environment for users requesting or issuing orders. However, it does not recognize users' intentions and can not adapt to them.

US Pat. No. 9,502,027 B1 describes a method for processing the output of a speech recognition. Here, after the conversion of the spoken language into text, the parsing, it is first examined whether it is a command. If this is the case, this command is executed, and if not asked by a query to clarify the ambiguity. However, the process can neither recognize intentions nor lead to a natural dialogue.

While the technical problems raised by this can now largely be considered solved, deep understanding and communication still pose considerable problems, especially those computer systems that have to process voice signals very quickly in real time while at the same time having limited resources.

Above all, the requirements result in "understanding" and "interpreting" the user input. The goal here is always to pass the Turing test. The test is then passed if the user can no longer distinguish from the answers or recognize whether he has communicated with a human or a machine. Apart from pure speech recognition, here too the laws of the human are

Communication, especially those that are dealt with in the transactional analysis of psychology, so must the computer system

• Respectfully respect "other people",

• Clarify expectations, especially the questions of what it is about and what you do

wants each other

• Obtaining information on the topic, not assessing it, asking opening questions instead of specifying alternatives, stopping the flood of information and having it structured and selected,

• paraphrasing or paraphrasing important conversation content, Make interpretations as such clear, use them sparingly,

identify, offer, if necessary,

communicate authentically and selectively according to the principle "Not everything that is true, I have to say, but everything I say must be true,"

observe the 50% rule, which means that at least half of the energy needed to solve the problem must lie with the other party,

Take stock and sum up what has been clarified and what has remained open, which questions have been raised and what next steps are pending,

Can give feedback.

The object of the invention is thus to enable the computer system to emulate human communication better and not only to capture and evaluate the language or even only their words and grammar, and thus with a

to engage human users in a dialogue.

In addition, the software should know some basic rules of strategic practice so that it can not be easily deceived or tricked.

The invention solves this problem by a method for dialogue between human and computer using a self-learning system, comprising

Receiving user input in natural language text,

• assignment of synonyms and keywords as well as word mappings,

Analyzing the user input on sentence structure and sentence structure and their recognition,

• assignment of key phrases,

• the determination and classification of the user intention,

• Checking whether a confirmation is required

• logical processing in a reasoning and answer preparation with the formulation of a decision on further dialogue design, formulation or dialogue with possible generation of a response,

• in which

• to determine the user intention, the statistical relevance of belonging to a target intention is determined while retaining all meta-information,

Classifying with a classification algorithm for declaring intent, linking decision tree, logic and speech understanding via meta-information,

• a feedback is generated using previous user surveys to the self-learning system, • if the statistic relevance falls below a threshold, a query is generated to the user as to whether the recognized intention was correctly understood,

• In this case, receive a response from the user and then to sentence structure and

Sentence structure is analyzed and repeatedly the user inputs are split into factual content and metadata about the user and again the user intention is determined and classified,

• the recognition result is delivered to the self-learning system,

• the self-learning system automatic system recommendations and / or

Decisions to Improve Intention Identification Over Several

Developed user interactions and ready for later use,

• initiates the generation of the voice text response to the user if statistical limits are exceeded, and otherwise induces further demand.

The individual method steps are explained in more detail below with reference to the known prior art.

First, the user is identified if possible. This can be done for example by a unique user ID. If the user has previously connected to the system, he will be recognized and the call will resume. If the user logs on again after an expiration time, the system generates a new instance based on the stored context information and lets the user start a new dialog. The user can connect via any interface such as chat or voice and then with appropriate identification and authentication, the appropriate user context can be loaded.

Such an application can be carried out in a manner similar to a forwarding of mobile radio signals between individual radio cells across channels and across the interfaces, thus permitting the transfer of dialogue between the individual channels. The system retains the context, the saved dialog status and the dialog history.

Receiving and then analyzing the user inputs to sentence structure and sentence structure and recognition is done by a natural language parser that outputs grammatical information of a sentence, such as identified subjects, further information about each word, and a tree structure, also called dependency tree. denotes the relation between the individual words.

This structure tree is then further analyzed in the "Dependency Tree Parsing" processing step, whereby subject-predicate-object structures are generated as far as possible. The processing of dependency trees and structural information extracted from it are combined with other existing meta information to use this information to determine the value of a user input. This so-called input object, which contains additional additional information, makes it possible to remove the corresponding meta information from the input and to buffer it in the conversation-dependent context.

Therefore, there is a transformation between the user input and the existing user context based on existing structural criteria, e.g. the language models as the basis of the "dependency tree parsing" instead.

Thereafter, the user intention is determined and classified in several stages. The intention of the user, also referred to as "intent", is to be understood in the following to mean a more far-reaching term than just a mere request, the request, referred to as a "query", representing a subset of the user intention. From the perspective of communication concerns the

User intention but not only the goal of presenting a fact, but it also includes the objective of the user himself, who either wants to buy something or wants to get a release or a permit or information, or even make a complaint or refund an ad. This will be the four

Language levels "fact level", "appeal level", "relationship level" and

"Self-disclosure level" relevant.

For all these levels, the computer system must be able to form a corresponding level of discussion or self-learning, and be able to execute these purposefully by means of structural criteria. This knowledge base is provided by a database of goal intentions, relevant context information, and external data sources.

In this case, the statistical relevance of belonging to a target intent while retaining all meta-information is determined to recognize the user intention. Subsequently, the classification is carried out with a classification algorithm for clarifying intent and meaning. In this step, the meta-information and the

Processing logic, also referred to as flow, to a basis for a

user-specific structured interviewing.

The information previously included as input meta-information was in

Structures, such as the data exchange format JSON, stored and will be used in the subsequent step for Intent detection and recognition of meaning. How the context is structured depends on the use case. This results in a further limitation of the previously determined meta-information to a specific Use case, a meaningful discussion according to the rules neat

Interviewing makes possible.

If the statistical relevance of a selected reference intent falls below a defined threshold, the system is able to formulate queries to the user in order to confirm or disprove the validity of the recognition. This type of feedback and the storage of user feedback creates a cycle in which the system is able to improve its own behavior, ie to learn independently. This in the form that after a certain number of queries pressed no further questions to any user must be asked. The system then learned to interpret and assign a previously unknown user request that did not match the reference intent.

It is not just a pure detection performed, but depending on

Conversation state is a targeted activation or deactivation more meaningful to the conversation of matching intents. For this purpose, an intent detection dependent on the conversation status is carried out with a conversation state-dependent delimitation of permitted or tolerated or meaningful intents. This achieves a serious improvement of the status-dependent intent detection. Paired with the one described above

Inquiry algorithm creates a procedure in which initially only a few structural

Information must be given and the system learns independently to lead the most natural possible dialogue with the user.

The intention recognition is done by statistical comparison. First, the

User statement in a parsing system broken down to triple that have the structure subject-predicate-object. The arrangement of the triples is displayed in a dependency tree. All eligible triples suitable for processing are stored in the system with each of the triples being assigned a statistical probability value. Within the framework of the learning system, this probability value is always adjusted, if several users consecutively cause the same triple, the probability is increased. The highest probability triple is chosen as Intent.

After intent recognition, the "Intent-Clarification" processing step, depending on the relevance of recognized and identified intentions, formulates a query to the user, so that any ambiguity of the user input by demand and re-capture of the processing steps described above are used to to get a clearer statement on intention and meaning of user input. In case of ambiguity of the user input in relation to the reference intents, the "Intent Clarification" clarifies whether the system has recognized the correct intent or not. When generating the inquiry, the following procedures are performed:

Each reference intent has one or more explicit queries associated with it. A script includes content from the meta-information and the context in a dynamically generated query. Thus, the query is adapted to the user input and generates a much more natural query or other feedback to the user.

After asking the user, the computer system waits for a response. If no response is given, the system again requests or aborts the user's request, if an answer comes, the computer system checks to see if it is being provided by the same user, authenticates it, and the analysis and processing proceeds again. Prior to this, it is ensured that the system waits for further user input in a valid state or that it actively asks the user for a conversation by asking questions.

After successful intent recognition, the next processing step of the "reasoning and answer preparation" is required: In this processing step, a context-based, structured dialog management suitable for the application is made possible by means of a logical decision matrix, a flow, a decision tree, a neural network or other implementation.

The decisive part of the "reasoning engine" is the graphical representation of a structure for dialogue management. About logical processing elements such as conditions such as "if", "eise", "then", "switch case" and user interaction elements such as "say", "say-once",

"asYNQuestion" etc. creates a logical framework for a case-specific dialogue. The "reasoning engine" manages the current conversation status of a discussion with the user, as well as previous discussions, according to the

User interaction to act as a conversation partner. Conversational context determined from previous information can be used as a basis for discussion.

This reasoning engine allows the meta-information from previous ones

Processing steps with decision criteria for meaningful, context-related

Use call handling. Previously executed processing steps provide the necessary meta-information of one or more user inputs for processing in the reasoning engine.

The intent discovery occurs before processing in the flow. On the one hand, a query to the user can be made directly from the intent detection. On the other hand, queries to the user can also be formulated directly from the flow. The difference is that the intent usually does not already contain all sorts of specifications, which are then queried by the flow. This separation between Intent and Flow follows from the Knowledge position of the user, who often can not sufficiently specify his concrete intention due to ignorance of the possible range of design, or because his decision-making process has not yet been completed and he reserves the right to alternatives.

For user interaction, replies or feedback to the user are generated from previous procedural steps or directly from the reasoning engine, allowing full bi-directional user interaction appropriate to the context of a conversation. To improve call quality, so-called "Natural Language Generation" [NLG] tools and frameworks can be used.

The system already enables partially dynamic generation of responses to the context or from stored meta-information. The queries are thus created from the context and output to the user. At this point, the flow generates the four language levels "fact level", "appeal level", "relationship level" and

"Self-disclosure level", so that the dialogues meet the communicative demands.

In further embodiments of the invention it is provided that the reception of user input is preceded by a conversion of spoken natural speech into text. Such a transformation is known in the art. In addition, this precedent step is used to gain further meta-information about the dialogue partner. Thus, it can be provided that meta-information of the user are included in the conversion of spoken language into text and provided separately to the algorithm of the intention recognition. As MetaInformationen the parameters hurry, emotion, language, dialect and person or combinations or parts thereof can be detected. This includes splitting user input into meta-information for factual content and meta-information about the user. Variations are known from market-known speech conversion applications that provide such meta-information in addition to the text that has been found.

Furthermore, these meta-information obtained from the natural speech of the user can serve as raw data for the intention recognition as well as for the subsequent logical processing and to this the parameters tonation, urgency and reassurance are delivered. So sometimes it can be concluded from the tonation on whether a complaint or a purchase request exists. In the case of urgency, it may be useful to lower the thresholds for statistical relevance. moreover

Such meta-information enables the flow of conversation appropriately

Adapting user behavior, such as between a purchase process, a

Consultation process or a complaint process is differentiated. Provided that the receipt of user input is preceded by a conversion of spoken natural speech into text, a further embodiment of the invention provides that the generation of the speech text response to the user is followed by a conversion of text into spoken natural language in order to lead to a unified dialogue , In such cases, it is necessary to uniquely authenticate each user at all times to ensure that the system responds to the same user who made the request so the dialogue continues with the same person.

The system therefore allocates an identification record to each user in one embodiment, allowing access to user-dependent context information. In this way, it is also possible to conduct an interface-independent dialogue, the context and the current conversation and dialogue status persist and are always consistent. When the user terminates the conversation by disconnecting from the system, it first checks if there is a technical error and becomes the context

cached. If the dialog has been finalized, the current dialog status will be deleted or deleted.

In addition to the feedback of spoken text, data can also be returned as control commands for an application. Speech interaction can thereby be achieved with real on-screen interaction, e.g. by displaying graphics or a device such as a robot, an "Internet of Things" device, or even a smartphone or a

Computer or a virtual or digital environment. These

User interface is preferably designed to be multiple

Can generate input signals consisting of text or speech and input data such as control commands and meta-information,

Control commands can also be processed directly from the meta information as data records, thereby influencing the conversation process from the context and using the flow elements, which control the conversation depending on the context, with the user. For example, a connected robot can greet the user when he detects movement. Through a possible face recognition such a robot could try to uniquely identify a user and his company. The system would then be able, based on existing contextual information, to create a dynamically generated greeting using the name and other context information and to convey it via the robot in the form of speech to the user.

Further embodiments of the invention relate to the way in which the self-learning system is embedded in the method for dialog guidance. Thus it can be provided that the meta-information automatically combines in the self-learning system, interpreted, consolidated and stored, whereby the user is responsible for the data protection concerns.

Self-learning systems are sometimes criticized for being unwanted

Automate and incorporate user input unreflectively. This leads to abuse and damage, such as when groups of users ask nonsensical questions to manipulate the system, or when the system learns prejudice and defamation. In order to prevent this, it can be provided that certain keywords have the

self-learning system is supported by an administrator who is involved in the occurrence of keywords or in the event of excessive demand. In the system described here, the flow and manageable context information ensure that the system is partially self-learning, but does not change its actual behavior, since it still operates on the basis of predefined structures. The

Administrator can only intervene controlling and bring in new elements for enhanced or enhanced user dialogues.

Further embodiments of the invention relate to the speech analysis. It can be provided that the self-learning system is supported by a Synonymmerkennung. The synonym recognition goes hand in hand with the elaboration of further context information, so that key words can be identified, and thus the input can be extended via temporary meta information according to the recognized key words.

Furthermore, it can be provided that by analyzing the sentence structure of

User input in dependency parsing a dependency mapping is worked out.

Further embodiments of the invention relate to the integration and query of external databases. Thus, it can be provided that, in order to determine conversation-relevant data, such as context information or meta-information, corresponding, linked databases are queried. These may be, for example, product databases of manufacturers or goods and their storage location in a shopping center. If these databases are foreign-language, a translation program can also be used.

In a further embodiment of the invention, it is provided that the system for

Generating issues via a markup language has access to the meta-information and the context and thus can include that information in the feedback and general processing.

The invention will be explained in more detail with reference to a figure. Fig. 1 shows an overview block flow diagram of the entire process including the options.

The invention will be explained in more detail by way of example, but the invention is not limited to this example.

Fig. 1 shows the user input 1 as text. The user can make this input either through spoken language, which requires a subsequent program to translate from speech to text, or via a smartphone or a computer or other device. The conversion can also be integrated by means of voice control in a smartphone, computer or other device. An upstream automatic translation is possible. For unique user identification also a kind of authentication or user identification is required. The user input, as well as the entire subsequent dialog, can also be mixed via different input channels.

In the first processing step "Synonym Mapper" 2, a synonym recognition takes place, wherein the text is checked for synonymous phrases given by the administrator. If such synonyms are found, Synonym Mapper creates 2 metadata that replaces the original text with these synonyms without deleting the original data. This meta-record of the original sentence of the user is successively extended in the subsequent processing steps.

In the second processing step, "NLP" 3 (Natural Language Processing), the user-defined metadata-extended input to sentence structure and sentence structure is analyzed. In detail, first

• determine number words and replace them with numbers,

• removes words and expands contractions,

• spaces removed

• missing punctuation added,

and it is analyzed whether the sentence is a question and if so, whose peculiarities are extracted.

Thereafter, the result is put in an NLP parser, which assigns words and punctuation of a text to parts of speech, which is also referred to as "POS-tagging", and the dependencies are determined to determine the grammatical and syntactic relationships between the sentence components to output as data. Specially trained machine learning models can be used for the generic identification of sentence structures. Following this, the data generated by the parser is parsed to identify the subject, object, their relationship to each other, and their properties. In addition, the record type is determined, ie for example, whether it is a question or a statement. These data are added as metadata to the record of the previous step.

In the third processing step "Keyphrase Mapping" 4, a second pass of a synonym recognition takes place using the metadata obtained in the previous step. In addition, the frequency of their occurrence is counted. In addition, the negativity is evaluated, it could also be that the user does not intend something exactly, so does not want. Furthermore, a categorization is carried out so that the right context databases and possibly external databases can be involved for further processing. Furthermore, tags are given. "Keyphrase Mapping" 4 thus expands the existing meta-dataset by these meanings.

In the fourth processing step "Intent Mapper" 5 is identified what the user

wanted to express when he made his user input. The following scoring algorithm is used for this.

The "Intent Mapper" 5 first accepts all valid ones defined by the administrator

Intent definitions and for each example sentence A provided by the administrator, it generates a pair P consisting of the sentence S of the user including the

For each of these pairs, the scoring algorithm assigns a similarity value W by making the following comparisons:

If A and S are identical, W gets the value 1. This also applies to the lemmatized form of S and the form of S in which synonyms have been replaced.

If no identity of A and S has been established, a number of

Property comparisons are made, each giving a similarity value W between 0 and 1. These properties include subjects, relationships, objects, as well as their lemmas, questioner-who-what-where, circumstance, if it is a question, conditions, negative adverbs and possibly other aspects.

The calculation follows the example of the following scheme, where EQ stands for the result shares of the properties of the respective total value W:

If subject (A) and subject (S) are identical then EQ = 1, otherwise EQ = 0. These

Calculation can be done on the lemmatized version of the corresponding words. Now, the results of this first calculation are merged as follows.

W = EQ (Subject) ^* 0.1 + EQ (Verb) ^* 0.2 + EQ (Object) ^* 0.3 + EQ (Question) ^* 0.2 +

EQ (negativity) ^* 0.2 The number of parameters used here may vary depending on the use and also the weightings are exemplary.

The scoring algorithm then checks for certain cases whether a value 0 must be assigned, for example if the example sentence is a question, but the user input is not, or if the subject, relationship and object do not match. In addition, if the negativity is determined, the system could assign the value W to 0.

The scoring values W, of the different pairs P, are then collected, stored in a table and sorted by size. This table is added to the metadata record.

Now, for each example sentence A and the user input S, this calculation is performed and the result of the calculation for the user input S is compared with all active intents and thus with a certain number of example sentences A. The example sentence with the highest relevance is selected. By belonging the example sentence to a specific intent, the intent assignment takes place.

In the next processing step "Examination of the Intent Score" 6 it is checked if a

Inquiry to the user is required, and if so, how. For this purpose the administrator sets two threshold values between 0 and 1. The first value is the query threshold RT and the second threshold is the validity threshold CT, the threshold CT is higher than that of the RT.

The two thresholds RT and CT divide the range of scoring values W into three ranges. If the first threshold RT is not reached, the program will explain to the user that he did not understand it and politely address to him the request 7, his input

reformulate. If the second threshold CT is reached or exceeded, the intention is considered valid and a query is waived, then the program continues with the flow 8.

If the first threshold value RT is reached, but the second threshold CT is not yet reached, the program takes advantage of its learning system and its

Databases provide a deeper investigation. The learning system inserts after the Intent Mapper 5 and is connected to a database 100, which consists of three parts:

User data 101, context data 102 and general, external data 103, in addition, the learning system is possibly connected to the APIs 104.

The part of the user data 101 contains data about the respective user and data of all users from previous user inputs and has a learning database individually for each user as well as a common learning database for all users together. The respective learning databases contain the already processed data previous dialogs, including related meta-information, and a counter that indicates how many times each phrase containing a recognized intent has occurred by the user and by all other users. The learning system acts as follows:

After the scoring value W has been determined, it is first checked in query 6 whether CT was reached or exceeded or not. In this case, the program continues without the intervention in the learning system with the next processing step 8. If W is between RT and CT, it is checked in the user data 101 in the examination of previous confirmations 10 whether the user has previously confirmed such an input sentence in a comparable context. If this is the case, no query is formulated to the user, rather the system "thinks" that the user again means the advised input and increments the counter for that detected sentence by 1. The program also continues with flow 8.

If not, the program takes the example sentence A, which belongs to the highest score value W, formulated in the query 1 1 from a confirmation question and asks the user if this is his intention.

If the user denies this in his answer 12, the program confirms that the intention has not been understood and, if other example sentences have also reached RT, whose score W was lower than the previous one, then asks the user if these example sentences then his intention was better. Otherwise, the program continues with flow 8, but sets a "flag" in the input object, indicating that the user's intent was not properly determined, so that Flow 8 can respond to the input and continue the user dialog ,

Of course, the user can then also at any time resort to a reformulation, for which then the original process starts all over again. If no confirmation is reached by this request, the learning system checks for each of the (inconclusive) queries whether the user's record has already been deposited as a learning sentence in the learning database LDB.

If this is the case, the associated counter is decremented by 1, unless it was already 0. Thus, the program reduces the risk that misunderstandings will be handed down.

If the user agrees, it is confirmed that this phrase means the intent and the counter is incremented by 1 for that user and set. In addition, in the processing of the answer 14, it is checked whether this sentence has already been confirmed as intent in previous queries of other users and stored as a learning sentence in the learning database LDB. If such a learning sentence already exists, its counter is increased by 1, if not, it is created as a learning sentence in the learning database LDB. Thereafter, in the step "Examination for example sentences" 15 it is checked whether the learned sentence is suitable as a new example sentence for further cases. In this case, if the counter in the learning database LDB exceeds a value LT specified by the administrator, the program removes the learning set from the learning database in the "Modification LDB" step 16 and, in the future, leads it as an example sentence so that a direct recognition takes place. Here, the new example sentence is marked as such to allow the administrator to retrieve this example sentence and, if necessary, manually delete it. This avoids abuse.

The program now continues with the processing step "Flow" 8. Flow 8 is a decision tree that enables you to create meaningful statements during a conversation with a user. The flow 8 uses the stored context data 102 and the status of the conversation to be a human-like one

Reach user dialog. The context is used to store information about everything related to the current user and the dialogue with him, for example the name of the user after he has named it. The context data 102 is stored and is then available for future user conversations. The status determines where you are in the conversation, for example, at the beginning, the end, or in a specific processing step. The status information is crucial for adequate answers, since identical user inputs can have completely different meanings depending on the current dialog status.

The Flow 8 has direct configuration access to external data sources, external data 103, and external APIs 104, and can include these in context and conversation.

The flow 8 consists of a number of different nodes that determine the logical structure to structure a conversation and generate the desired outputs. These are a start node that determines the beginning of a flow, function nodes that perform a specific function, for example, output a statement to a user or make an entry in a database, logic nodes that structure the flow within the flow, for example, if / then / else or

switch / case / default, and code nodes containing executable code, such as JavaScript. Using these nodes, the flow 8 can be the current one

Enter user input, the context and the status, preferably by means of a script language activated at runtime.

Such a runtime activated scripting language typically consists of a markup language tagging portions of text or data. Furthermore, this activated during runtime scripting language to form corresponding requests to databases with external data 103 and allow out of the flow node direct queries to these data sources. Synchronous in the form that any requests must be answered until the flow 8 continues to process, or asynchronously in the form that the immediate processing of further flow requests. The

Processing in the flow 8 after an asynchronous database request takes place as soon as the requested external data 103 is present, the flow of the still valid flow 8 is continued with the new data now available. In a similar procedure, external programming interface APIs 104 can also be accessed here.

Each flow starts with a start node followed by any number of child nodes. All nodes are executed in sequence starting from the starting node as they are reached. Logic nodes leading to branches control the process and ensure that only relevant nodes are traversed during a call. Thus, it may be that some nodes in the course of a dialogue repeatedly or not at all. This constellation results in a free but structured dialogue.

It is particularly advantageous to use a markup language within the Flow 8 that allows to call or change the meta-information on user input, the context and the status directly within the decision tree. Access is e.g. via different nodes or via free programming.

In the following, the interaction will be described in more detail by means of a simple example. Consider a larger group of visitors

Shopping centers that are a little hungry. The first visitor speaks to the system with the words: "I want a snack, where is this?". At first this will be

converted sentence into known text by means of known technology, since the access can be made by several consecutive users, the currently active dialog is assigned an identification number and a new default conversation context is loaded. Subsequent user interactions within an active dialog take place under

Use of this conversation context. If a new user then leads a conversation, a new default conversation context is loaded again.

This text input 1 starts the system. In the key phrase Mapper 2, the term "snack" is searched for synonyms or generic terms, for example the term "pizza" is found for this and it is loaded from the user-added dictionary. There, a keyword with a unique identifier, which can be used later in the flow, and a freely definable number of associated synonyms, created. in the

Subsequent step NLP 3 examines the word by word entry and the relationship between the words. In this case, it is recognized that they are two half-sentences, the last of which is a question that addresses a location. Subject of the first half sentence is the user himself, object is a food item. A number was found: the word "one". The verb could indicate a relationship state that contains an intention. But the questioning sentence could also be an intention.

In the step "Keyphrase Mapping" 4 it is first checked whether something is negated, this is not the case here. The categorization is specified, items are a food and a location.

In the step "Intent Mapper" 5, the example sentences A for food and for

Location information loaded and pairs are formed. For example there are sandwiches, pizza and chips with corresponding example sentences A from the

Synonym database. For the location can not be because of the reference "the"

Example sentence are loaded because the subject follows only from the first half-sentence, which is first to clarify. For each of these example sentences A, a scoring value W is assigned. These scoring values can be roughly proportional to the typical sales of the local

Other procurement modalities, such as references to current advertising activities are conceivable.

In the present example case, there is no example sentence A with "snack", so none of the example sentences has the scoring value W = 1. The remaining examples should have the values W-0.6 for chips, W = 0.7 for sandwiches and W = 0.8 for pizza, which was previously set by the administrator, were also set a CT value for no longer confirming queries and RT = 0.5 for unintelligible user input.

In the following step "Checking the Intent Score" 6 it is checked whether the highest scoring value of the example sentences exceeds the value CT. This is not the case here. Then it is checked whether this highest value is at least above RT, this is the case. The system then checks in step 13 whether there is already a confirmation from a previous user input for this user and for such a user input. If this is the case, no further search is made for a confirmation, but the user input is regarded as confirmed and the system proceeds to the flow 8 as the next step.

Care must be taken with such skipping of the acknowledgment, and the flow 8 must check from the context as well as the status retrospectively whether the

Skipping the confirmation was really appropriate, for example, the same user would have come back later, and can ask for a snack. Since he now knows the place where he could get pizza, then his expectation would be that now he is offered something else, otherwise he would not have asked. Therefore, the second, identical question does not automatically have the same expectation regarding the answer, but can also express a different intention depending on the status, this clarification is not the subject of the examination of the Intent Score 6, but of the Flow 8.

If there is no confirmation from this user of the past, the program initiates inquiry 1 1 associated with the example set of the highest scoring value W. In the present example this is the sentence, "Would you like a pizza?". The program expects a yes-no answer. If the answer is "yes", the program checks for the second half-sentence whether there is an example sentence with a value of 1 for this, in this example this is assumed and no further inquiry is necessary. The program also continues to flow here, but previously saves the sentence in its learning database LDB, which is part of the user database 101 so that it can recognize the question next time.

In addition, the program checks whether it already knows the question about the snack from other user queries. If this were the case, it would increment the counter in the learning database by 1 for all users, but in the present example, the sentence must first be recreated in the learning database. For this reason, the query, whether the

Threshold value LT for transfer to the database for example records in the database for reference or example sentences for intents has also been exceeded.

The program then proceeds to flow 8, which generates a response where to get a pizza, possibly in the form of an enumeration, if there are several places to do it that it knows. Possibly. It also asks for the pads you want and immediately places a complete order that the system can then forward to a digital ordering system. In this case, the flow 8 can access by configuration to further application-specific, external data 103 or programming interface APIs 104.

In the present case, the second user asks "I also want a snack but not pizza". The first steps are analogous, but there is no second half sentence with a local question, and there is also a negative that is identified in the Keyphrase Mapping 4 if it has not already been identified in the previous step NLP 3. Since a negation to pizza is included, the example sentence with the pizza is given the scoring value 0. In the following step "Intent Mapping" 5 you will only be asked "Would you like a sandwich?". If the user affirms this, then the procedure continues analogously as in the previous case. Only in the flow 8 does the program determine that there was no asking for a location and ask further questions about it, possibly it also offers users a corresponding direct order.

The learning database then has two different ones for the question about the snack

Confirmations with the meter reading 1. Unless the group has other similar requests the system learns what the users mean and want by increasing or decreasing the counter, and by including all contextual and meta-information.

LIST OF REFERENCE NUMBERS

1 user input

2 synonym mapper

3 NLP

4 Keyphrase Mapping

5 Intent Mapper

6 Examination of the Intent Score

7. Please

8. Flow

9. (empty)

10. Examination of previous confirmations

1 1. Consultation to users

12. User response

13 Examination of the answer

14 Processing the answer

15 exam for example sentences

16 modification LDB

100 database

101 user data

102 contextual data

103 External data

104 External Programming Interface API

Claims

claims

1 . Method for dialogue between human and computer using a self-learning system, comprising

Receiving user input (1) in natural language text,

• assignment (2) of synonyms and keywords as well as word mappings,

Analyzing (3) the user inputs on sentence structure and sentence structure and their recognition,

• assignment (4) of key phrases,

Determining and classifying the user intention (5),

• Check (6) if confirmation is required

• logical processing (8) in a reasoning-and-answer preparation with the formulation of a decision on further dialogue design, formulation or dialogue process with possible generation of a response, characterized in that

• to detect user intent (5) determine the statistical relevance of belonging to a target intent while retaining all meta-information,

The classification (5) is carried out with an intentional classification algorithm linking decision tree, logic and speech understanding via meta-information,

• a feedback using previous user surveys to the

self-learning system is generated

• if the statistic relevance falls below a threshold, a query (11) to the user is generated as to whether the recognized intention was correctly understood,

• In this case receive a response from the user (12) and then on

Sentence structure and sentence structure is analyzed (13) and the user input in factual content and in meta-information about the user is split and repeatedly and again the user intention is determined and classified,

• the recognition result is delivered to the self-learning system,

• the self-learning system automatic system recommendations and / or

Develop decisions for improving the recognition of intent through multiple user interactions and make them available for later use,

• if the statistical limit value is exceeded, the generation of the

Voice text response (8) to the user, and otherwise induces further demand.

2. The method according to claim 1, characterized in that the reception of user input (1) is preceded by a conversion of spoken natural speech into text.

A method according to claim 2, characterized in that meta-information of the user is included in the conversion of spoken language into text and provided separately to the algorithm of intent recognition.

4. The method according to claim 3, characterized in that as meta-information, the parameters hurry, emotion, language, dialect and person or combinations or parts thereof are detected.

5. The method according to any one of claims 3 or 4, characterized in that the

Users are split into meta-information for factual content and meta-information about the user.

6. The method according to any one of claims 3 to 5, characterized in that as

Meta-information the parameters tonation, urgency and reassurance are given to the intention recognition.

7. The method according to any one of claims 1 to 6, characterized in that the

Meta information in the self-learning system is automatically combined, interpreted, consolidated and stored.

8. The method according to any one of claims 1 to 6, characterized in that the

Generating the voice text response to the user following a conversion of text into spoken natural language.

9. The method according to any one of claims 1 to 8, characterized in that the self-learning system is supported by an administrator.

10. The method according to any one of claims 1 to 9, characterized in that the self-learning system is supported by a Synonymmerkennung.

1 1. Method according to one of claims 1 to 10, characterized in that a dependency mapping is developed by analyzing the sentence structure of user input in dependency parsing.

12. The method according to any one of claims 1 to 1 1, characterized in that the

Determination of conversation-relevant data databases (100, 101, 102, 103) or APIs (104) are queried.

13. The method according to any one of claims 1 to 12, characterized in that the

System for generating expenses via a markup language access to has the meta-information and the context and thus can include this information in the feedback and general processing.

14. The method according to any one of claims 1 to 13, characterized in that a targeted activation or deactivation is made more meaningful, Intents suitable for conversation by a dependent on the conversation status intent detection is performed with talk state dependent delimitation of permitted or tolerated or meaningful intents.

15. The method according to any one of claims 1 to 14, characterized in that as

User interface is a robot, an Internet of Things device, a smartphone, a computer, a virtual or digital environment, or a combination of it is used.

16. The method according to any one of claims 1 to 15, characterized in that the

User interface generates multiple input signals consisting of text or speech and input data.