Method, device and application for generating response output in response to voice input information

By dividing two analysis units in the voice input information, processing voice input information in non-target-oriented sessions, the problem of difficult response to unpredictable input in the prior art is solved, and effective response and response output to highly variable information is achieved optimization.

CN113228165BInactive Publication Date: 2024-11-08VOLKSWAGEN AG
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201980084669.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2018-12-18
Filing Date
2019-11-11
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

Prior art is difficult to effectively respond to unpredictable voice input information in non-objective oriented speech sessions, especially when using terms that have not yet appeared in session history or training, and machine learning-based approaches rely on a large amount of expert knowledge and dynamics. The dictionary update increases the complexity of the system and the difficulty of learning.

Method used

By dividing two analysis units in the voice input information, one is used to process frequently occurring common content to generate response partial output, and the other is used to determine the target information of the predefined entity and embed it in the final response output, reducing analysis The learning scope and functional requirements of the unit.

Benefits of technology

It realizes effective response to highly variable and unpredictable information in voice input, reduces the learning cost and complexity of the system, and optimizes the response output quality of non-target-oriented sessions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113228165B_ABST
    Figure CN113228165B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for generating a response output (A) in response to voice input information (E), the method comprising: a) determining whether the voice input information (E) includes at least one predefined entity; and when this is the case: b) generating a voice information set in which the entity is not included based on the voice input information (E); c) generating an information set, wherein the information set includes information in a knowledge database (14) assigned to the entity; d) using a first analysis unit (22): determining at least one target information important about the entity from the information set; e) using a second analysis unit (24): generating a response partial output based on the determined target information; and f) generating a response output (A) based on the response partial output. The present invention also relates to a system (10) and an application of such a system (10).
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method and apparatus for generating a response output in response to speech input information. More precisely, the present invention relates to the field of computer-assisted speech analysis and in particular speech conversation systems, in which speech input information is detected and a response output is generated based thereon. Background Art

[0002] It is known that speech input information is analyzed by means of computer-aided systems and in particular analysis units included therein and that a response output is generated based thereon. In this case, a distinction can be made between pure speech recognition and an analysis of a plausible response (response output) to the recognized speech input information, wherein the analysis can include, for example, a detection of the semantic content of the speech input information. In this context, speech input information can be understood as information in text or audio format, which is generated based on the speech input and / or describes the content of the speech input.

[0003] Within the framework of so-called goal-oriented conversations, users can, for example, verbalize questions in order to obtain the desired information. In general, conversations can be called goal-oriented if they are directed toward a defined purpose, as is the case, for example, when asking for information. The purpose would then be to obtain information. Likewise, there are non-goal-oriented conversations, in which users engage in a conversation with a computer-aided system (conversational system) without a previously defined purpose, for example, defined by the desired information. This can also be referred to as chatting with such a system.

[0004] Processing non-target-oriented conversations is more technically demanding, because the processes of these non-target-oriented conversations, for example in the form of so-called conversation states, are more difficult to estimate. In order to be able to respond to the corresponding various possible conversation states and generate suitable response outputs, the conversation system must be able to analyze or understand a large amount of information contained in the voice input information. This means that the system and especially the analysis unit of the system must be trained with the help of a correspondingly rich dictionary or a universal database. The latter is mainly applicable to systems based on machine learning methods and especially so-called deep learning methods. For example, in such a system, the corresponding words must be explicitly stored in the dictionary for each conceivable proper name, such as the proper names of people and cities, or for any numerical value, such as a numerical value in the form of a year or temperature value, so that the analysis unit can be trained with the help of the dictionary and later identify the corresponding words, thereby correctly detecting the semantic content. This increases the cost of establishment and makes the learning process of the corresponding analysis unit difficult.

[0005] Until now, attempts have been made to solve this so-called "out-of-vocabulary problem" by having the analysis unit recognize not only the words that have already been learned in the preparatory learning process, but also other words that have been used so far in the conversation history (i.e., words that have been used in the previous use of the analysis unit). However, in this case, new words that appear during the conversation history must also be constantly recorded and stored in a corresponding dictionary. Examples of such methods are found in the following scientific publications:

[0006] Raghu,D.&Gupta,N.(2018).Hierarchical Pointer Memory Network for TaskOriented Dialogue,arXiv preprint arXiv:1805.01216

[0007] Eric, M&Manning, CD (2017). A copy-augmented sequence-to-sequencearchitecture gives good performance on task-oriented dialogue. arXiv preprint arXiv: 1701.04024.

[0008] On the other hand, there are also systems that attempt to master non-goal-oriented conversations in an iterative manner rather than based on machine learning (see, for example, https: / / www.pandorabots.com / mitsuku / ). However, the achievable quality is greatly limited by the states and rules that have to be defined in advance (for example in the form of so-called expert knowledge).

[0009] Therefore, it is not possible to respond appropriately to unpredictable speech input information, especially in the context of non-goal-oriented conversations, with any of these known methods, especially when terms are used within these speech input information that have not yet appeared in the conversation history and / or in the training. In the known solutions, there is usually also a one-dimensional information flow: although the analysis unit can make requests to the dictionary for the purpose of speech recognition, the dictionary can also be dynamically supplemented based on the conversation history in the manner explained above. However, the (especially dynamically supplemented) entries of the dictionary cannot be used to generate a response output.

[0010] Another disadvantage is that at least the methods mentioned above that are not based on machine learning are characterized by the fact that a large amount of so-called expert knowledge or expert knowledge models are still used as analysis components. However, this deviates from the real purpose of processing non-goal-oriented conversations, because more predefined relationships in the form of expert knowledge must be defined and stored in advance. In this context, expert knowledge can generally be understood as the definition of logical relationships or rules, which can be used to determine the appropriate response (especially the response output) for the current state. Summary of the invention

[0011] It is therefore an object of the present invention to improve the computer-based generation of a response output in response to a speech input, in particular in the case of non-goal-oriented conversations.

[0012] This object is achieved by a method according to the invention, a device according to the invention and an application according to the invention. Advantageous further developments are described in the following description. It is to be understood that all features and definitions mentioned at the outset (alone or in any combination and unless otherwise indicated or obvious) can also be provided in the present solution or are equally applicable to the present solution.

[0013] The basic idea of ​​the invention is to divide the information of the speech input between two analysis units, which are preferably designed separately from each other (for example as separate software programs, software modules or computer program products), but which can also advantageously exchange information with each other. One of the analysis units can be designed to generate partial response outputs, wherein the latter can be generated in particular in response to frequently recurring and / or common contents of the speech input. The other analysis unit can be designed to determine target information, which is assigned to predefined entities within the speech input, wherein this target information can then become a component of the final response output. This target information can also be used as input variables for the first-mentioned analysis unit in order to generate a response output.

[0014] This has the following advantages: the contents of the speech input (that is to say the speech input information) can be predefined as entities, which according to the invention can have a higher variability or be associated with a large amount of other information that cannot be clearly predicted. Other entities can also be supplemented during operation and, for example, stored in the knowledge database, without requiring a new learning or training process of the system for this. This is important, for example, in order to supplement new information that appears during operation (such as newly released movies, etc.). Entities can especially be proper names or numerical values. Examples of information associated with an entity in the form of a movie title are, for example, director, participating actors, genre, movie awards, etc. The associated information itself can also be an entity.

[0015] At least one of these pieces of context information can then be determined by one of the analysis units as target information that is important in the current context. The analysis unit can be trained for this purpose (e.g. by means of a machine learning process), in particular in such a way that the analysis unit determines the target information based on the current speech context and a selection of the pieces of context information (information set or, in other words, a subset of the knowledge database). This can be understood in particular to mean that the analysis unit evaluates all pieces of context information with respect to their possible importance (e.g. in the form of a probabilistic statement) and then selects those pieces of context information with the highest importance evaluation as target information.

[0016] By this, it can be achieved that additional or target information matching the entity can be reasonably embedded in the final answer output. Therefore, at most the other analysis unit must be trained to process information associated with the predetermined and diverse entity, but the analysis unit that generates the above-mentioned answer part output does not have to be so. Such associated information does not have to be generated by the analysis unit itself, but can be called from the knowledge database. This reduces the required functional scope of the analysis unit and therefore also makes the learning process of the analysis unit easier.

[0017] That is, in other words, according to the present invention, it is possible to distinguish between contents of speech input that may have increasing variations (referred to herein as entities) and / or are associated with a large amount of other information, and between information or contents of speech input that often appear repeatedly and are not, for example, proper names. Then, for these two types of information / content, separate analysis units can be set, which are correspondingly optimized for the analysis tasks assigned to them respectively and the learning processes of these analysis units can be correspondingly carried out in a targeted manner. In addition, the embedding of a knowledge database is advantageous, and then one of these analysis units can select important target information from the knowledge database for embedding into the final response output. As described above, the knowledge database can also be continuously supplemented by new entities, without the need for these analysis units to enforce a new learning process.

[0018] In detail, a method for generating a response output in response to voice input information (the voice input information is, for example, a voice input converted into a text format) is proposed, the method having:

[0019] a) determining whether the speech input information includes at least one predefined entity;

[0020] And when this is the case:

[0021] b) generating a voice information set in which the entity is not included based on the voice input information;

[0022] c) generating an information set (entity information set), wherein the information set includes information in the knowledge database assigned to the entity;

[0023] d) using the first analysis unit: determining from the information set at least one important target information about the (determined predefined) entity;

[0024] e) using a second analysis unit: generating a response portion output based on the determined target information; and

[0025] f) generating a response output based on the response partial output.

[0026] In general, the method can be implemented in a computer-aided manner and / or by means of at least one (preferably digitally and / or electronically operated) computing unit (e.g. having a graphics card, a microprocessor or a general-purpose computer processor). For example, the method can be performed using a conventional PC.

[0027] At least one of the speech input information and the answer output can exist or be generated in the form of an audio file or as audio information. Additionally or alternatively, the speech input information and / or the answer output can exist in text form. In principle, it can be provided that the solution according to the present invention is used in a manner that the solution is embedded in a system that performs the conversion between audio information and text information. For example, a so-called ASR unit (Automatic Speech Recognition) can convert an audio speech input into text (then this corresponds to an exemplary speech input information according to the present invention), which is then processed by the solution according to the present invention. Additionally, a downstream TTS unit can convert the answer output generated according to the present invention and based on text into an audio output. However, it is also possible that the solution according to the present invention includes such an ASR and / or TTS unit or generally includes an audio and text conversion function. In addition, unless otherwise specified or obvious, the reference to the speech input information in the text can include not only the input in audio format but also the (audio) input converted into a text format (that is, the speech input information can generally exist in audio form or text form, wherein in order to provide measures according to the present invention, it is preferably at least temporarily converted into a text form).

[0028] As mentioned above, for example, names of articles, names of places, names of objects, names of blocks, names of work or names related to a subject and in particular proper names or numerical values ​​can be predefined as entities. To identify such entities, existing solutions can be used, such as so-called "named-entity recognition" or "named-entity-resolution" algorithms. In this way, for example, entities in the form of predefined proper names, such as those used, for example, for naming sports teams, films, places or persons, can be identified.

[0029] The speech information set can be generated in the form of a data record or generally as digital information. The speech information set may contain speech input information, wherein the speech input information exists at least partially in text form as a first recognition result in addition to the entity. The entity can be deleted or replaced, preferably by a so-called placeholder (or in other words by a template). The placeholder can, for example, contain a title, a general term or an identifier naming the entity. Generally speaking, it is preferred that the second analysis unit also takes the placeholder into account in order to determine the partial output of the response.

[0030] If one of these predefined entities is recognized in the (especially textual) speech input information, for example by means of a "named entity resolution" algorithm, then according to the invention, the corresponding placeholder can be embedded in the recognition result and / or in the text form instead of the entity. This can be achieved by means of so-called constructors.

[0031] Information can be stored in the knowledge database and can be associated and called up, for example, by means of so-called tags, identifiers, hashes or so-called key-value relationships. By means of the features listed above, a relationship can be established with the entity recognized (in the voice input information). If, for example, a movie title is recognized as an entity, then in this way, associated information in the form of the actors who acted therein, the year to which it belongs, the filming location, etc. can be determined, which is also stored in the knowledge database and is preferably linked to the entity via one of the above-mentioned relationships.

[0032] That is, generally speaking, a hash request may be directed to a knowledge database in order to combine information associated in the knowledge database with the entity into an information set.

[0033] In the information set, for example, all entries found in the information database can be collected as numerical values ​​and / or converted into vectors. Preferably, these numerical values ​​or vectors can be stored in a matrix, which the first analysis unit then uses as input information in order to identify the most important target information (preferably based on the analysis results of the second analysis unit as another input variable).

[0034] The first analysis unit and the second analysis unit can each be provided as a computer program product and in particular a computer program or software module. However, the first analysis unit and the second analysis unit can have (compared to each other) different input and / or output parameters or be configured to process and generate correspondingly different input and / or output parameters.

[0035] In general, it should be pointed out that, unless otherwise specified or obvious, the term analysis in connection with these analysis units can refer to analysis in the sense of determining a suitable response output (or at least a component of a response output), determining target information and / or detecting at least part of the semantic content of the speech input information. The target information can be those pieces of information from the information set that have the highest (assumed) importance in terms of entities. Here, importance can in particular refer to the currently existing conversation state or the importance can be determined (by the first analysis unit) based on the conversation history. More precisely, the first analysis unit can obtain the conversation history (also synonymously referred to as speech context in this article) and the information set as input parameters. Within the framework of the machine learning process, the first analysis unit can have been trained to then select one of the information in the information set as target information or evaluate the information as particularly important based on these input parameters. The information may include one or more entities that are considered important for generating a partial response output. For example, the speech context can be obtained from the second analysis unit. Then, the first analysis unit outputs the target information as an output parameter, more precisely, preferably outputs the target information as an output parameter to the second analysis unit.

[0036] The second analysis unit can receive the speech information set (especially including possible placeholders) and the determined target information as input variables. The second analysis unit can have been trained (for example within the framework of a machine learning process) to generate and output a partial response output based thereon.

[0037] In other words, at least and preferably exactly one piece of information contained in the information set and associated with the entity can be selected as the target information. Preferably, the target information has the current highest importance and / or is most suitable to be part of the answer output with respect to the current speech context. Generally speaking, it is preferred that the second analysis unit then determines the answer partial output knowing the determined target information and especially based on the determined target information. The target information can, but does not have to, become a component of the answer partial output and / or the answer output. In general, the target information may include at least one entity or a group of entities. For example, the target information may correspond to a row in the matrix discussed below, in which the entries of the information set are sorted or these entries form the information set.

[0038] If, for example, a film title is currently mentioned within the framework of the voice input information and the quality of the cast is generally highlighted, it may be important to select the actor names as target information associated with the film title identified as an entity in the next step and to make these actor names part of the answer output. If, on the other hand, it is clear from the voice input information that the user would like to still watch the film first, it may be more important to discuss possible playback locations or viewing channels for the film (i.e., for example, the terms DVD, cinema, or that become part of the answer output).

[0039] The second analysis unit can generate a response part output (e.g., as a text or audio output) based on the speech information set. In particular, the second analysis unit can determine the current conversation state and / or speech context (synonymous with conversation history in text). The conversation state can, for example, be a predefined state, such as asking questions about a defined subject area.

[0040] The final answer output can be generated with the aid of an optional reconstruction unit. In general, the answer output can be distinguished from the answer part output in that the answer output contains the answer part output, but if necessary also contains target information and / or entities embedded therein and / or additionally supplemented, which replace placeholders that may also be contained in the answer part output.

[0041] In other words, the answer output can be a combination of the answer part output and the target information and / or can include one or more target information inserted into the answer part output. In particular, possible placeholders in the answer part output can be replaced by the target information but also by other entities. In particular, the technical contribution of the solution according to the invention can be seen in that the important target information can be determined with limited effort and a sentence structure (answer part output) can be formed that matches the found target information using the second analysis unit. Then, the above-mentioned combination of these contents can be carried out to produce the answer output.

[0042] Therefore, according to one embodiment, it can be provided that: in the speech information set, the entity is replaced by a placeholder, and if the placeholder is also contained in the answer part output, the placeholder is then replaced by the entity. This can be done, for example, within the framework of step f). If the corresponding placeholder is not contained, it may also happen that the answer part output corresponds to the final answer output. Therefore, instead of the answer part output, it is also possible to say that the answer output information may be further processed by replacing the placeholder or the answer output information already corresponds to the final answer output. An extension of the present invention stipulates that the second analysis unit obtains the analysis results of the first analysis unit (for example, in the form of the determined target information or a group (subset) of entities forming the target information) in order to determine the answer part output. More precisely, these analysis results can form the input parameters of the second analysis unit. In general, the second analysis unit can be trained or taught to determine a suitable answer part output based on these analysis results.

[0043] According to an extension, the first analysis unit and / or the second analysis unit comprises an (artificial) neural network. In particular, a neural network comprising a plurality of layers (so-called deep) can be provided at least at the second analysis unit, wherein so-called input layers and output layers are linked via a plurality of layers. Each layer of the neural network can contain nodes in a known manner, and these nodes are linked to one or more nodes in adjacent layers via weighted connections. The weights of the nodes and the links can be taught in a known manner within a training or learning process of the neural network. In general, neural networks can define nonlinear relationships between input variables and output variables, and thus can determine suitable output variables based on the input variables provided. This nonlinearity can be caused or taken into account by so-called activation functions of the neural network. These networks can depict mathematical relationships between input variables and output variables via corresponding nodes and layers.

[0044] In this context, it can also be provided that the second analysis unit comprises a neural network in the form of a (preferably hierarchical) sequence-to-sequence model (Seq2Seq-Model1). The neural network can comprise, in a manner known per se, at least one encoder which receives an input variable (e.g. a speech information set). Furthermore, the encoder can determine a state based on the input variable and output this state to a decoder of the network. The decoder can generate a response partial output of the type explained herein and output this response partial output as an output variable.

[0045] Alternatively or additionally, the first analysis unit may comprise a neural network in the form of a feedforward network and / or with an attention mechanism. In general, a feedforward network may have no feedback and may therefore be a simple network, in particular having only two layers. The network may receive as input parameters an information set and a speech context (conversation history) and in particular may receive as input parameters an information set encoded or converted into a vector representation.

[0046] In general, the first analysis unit can determine a set (subset) of potentially important information from the knowledge database and then determine the target information from it. The latter is preferably implemented with the help of an attention mechanism. In principle, with the help of the attention mechanism, a value (score) can be calculated for each information of the subset based on the conversation history, which value indicates the assumed importance of the information, for example in terms of probability. Undervalued information cannot be selected as target information and has no influence on the output of the answer part, because these undervalued information are not taken into account by the second analysis unit.

[0047] According to an extension, the first analysis unit receives as input a vector representation (or in other words, a vector expression) of the information set. This can be achieved by so-called embedding, in which the information stored in the knowledge database (those information assigned to the entity) is converted into a vector representation. The result can be an information set evaluated or checked by the first analysis unit. Alternatively, the information set existing before the corresponding vector conversion can be regarded as an information set, and a vector representation of the information set can be generated based on the embedding.

[0048] Generally speaking, conversion to vector representation provides the following advantages: the analysis unit can be based on a mathematical model or a mathematical function and can therefore better process input variables in the form of vector representation. In this case, based on known methods such as "Word2Vec" or "GloVe", it is also advantageously possible to embed speech information.

[0049] According to another embodiment, the second analysis unit is not set up to determine the target information (or target entity) based on the information set in the knowledge database for the determined entity (itself). The second analysis unit may not be set up to determine important and / or assigned target information (or target entity), especially not to determine important and / or assigned target information (or target entity) based on the entries of the knowledge database and / or the above-mentioned information set. In particular, the analysis unit may not be trained for this and / or may not depict or recognize the corresponding relationship. However, the second analysis unit can be set up and especially trained to: understand the target information and / or generate a matching response part output based on this. However, the determination of the target information itself should preferably be achieved by the first analysis unit. This can be achieved: the second analysis unit is instead focused on other tasks and the training cost can be adapted accordingly. For example, although the second analysis unit should be able to handle placeholders or generally entity classifications, such as "movie titles" when necessary, it is not necessary to learn all entities or proper names specifically associated with it. Alternatively, these specific entities (preferably but not optionally the placeholders used for this purpose) can be specifically excluded from the analysis scope of the second analysis unit, which instead preferably generates a partial response output based at most on the placeholders or entity classifications as well as the target information and the speech context and is trained accordingly for this purpose.

[0050] Additionally or alternatively, provision can be made for the first speech analysis unit not to receive the speech information set as an input variable and / or not to be configured to independently generate the response partial output (or the response partial output).

[0051] The present invention also relates to a system for generating a response output in response to a voice input message, the system comprising:

[0052] a) a determination unit, the determination unit being configured to determine whether the speech input information includes at least one predefined entity;

[0053] b) a speech information set generating unit, the speech information set generating unit being configured to generate a speech information set in which the determined entity is not included based on the speech input information;

[0054] c) an information set generating unit, the information set generating unit being configured to generate an information set about the entity, wherein the information set contains information in the knowledge database assigned to the entity;

[0055] d) a first analysis unit, which is configured to determine at least one target information that is important with respect to the entity from the information set;

[0056] e) a second analysis unit configured to generate a response portion output based on the determined target information; and

[0057] f) A response output unit configured to generate a response output based on the response partial output.

[0058] Finally, the invention also relates to the use of a system according to the above aspects for generating a response output in the case of a non-goal-oriented conversation. It has been shown that it is particularly advantageous to provide the two analysis units according to the invention and thus divide the corresponding analysis tasks in the case of a corresponding non-goal-oriented conversation (between user and computer-aided system).

[0059] In general, the system can be set up to implement the method according to any of the above and below aspects. All extensions discussed in the context of having similar method features can also be provided with the same system features. In general, the system can be based on a computer or be called a computer-based speech conversation system. The units of the system can be implemented as software modules and / or can be implemented as components of a computer program product (for example in the form of individual software functions). Optionally, the system may include: a computing unit (for example a processor), on which these units can preferably be implemented; and / or a storage device, on which the knowledge database is preferably stored.

[0060] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings, wherein:

[0061] Figure 1 An overview of a system according to an embodiment of the present invention is shown;

[0062] Figure 2 Shows Figure 1 A detailed diagram of the system in ; and

[0063] Figure 3 Shows the use of Figure 1 and 2 Flow chart of the method according to the present invention implemented by a system.

[0064] exist Figure 1A system 10 according to an embodiment of the present invention is schematically shown in the figure. The system 10 is implemented as a computer program product, wherein the functions or functional blocks explained below can be implemented in a manner known per se by individual program modules. The system 10 can be implemented on a computing unit of a conventional PC. Only the knowledge database 14 explained below can be understood as a data collection, not primarily as a software module or a function that can be implemented by a computing unit, but can be stored as a data collection of the system 10. All units of the system 10 mentioned in this article can be correspondingly implemented as software modules, computer program products or software functions. However, the system 10 can also optionally include hardware components in the form of a computing unit (processor and / or graphics card) and / or a storage device for the knowledge database.

[0065] The system 10 receives, as an input parameter, speech input information E in the form of an audio input that has been converted into text form. In a determination unit 12, predefined entities contained in the speech input information can be identified, which is a proper name recognition module (NER - Named Entity Recogniser or Named Entity Resolver) and is based on a known algorithm. For this purpose, the determination unit 12 can use the knowledge database 14 explained below, in which the corresponding entities are stored. The recognition of entities in the speech input information (especially when these entities already exist in text form) is known per se and can be converted in practice by using conventional named entity recognizers, which can usually be based on neural networks.

[0066] The determination unit 12 can define the corresponding entities contained in the speech input information E as output parameters. The output signal of the determination unit 12 is supplied to a speech information set generation unit 16, which can also be referred to as a constructor. The speech information set generation unit 16 generates a speech information set (for example in the form of a text file) in which the entities are respectively replaced by placeholders or in general the identified entities are not present in the speech information set. The speech information set is supplied to an analysis block 18 as an input parameter.

[0067] This has the following advantages: Figure 1 The second analysis unit 24 in the analysis block 18 which is only schematically indicated in FIG. Figure 2) do not have to be set up for analyzing specific entities or based on specific entities, but at most are set up for processing corresponding placeholders. It is easy to understand that these entities have a significantly higher variation than placeholders, which should be understood as general terms for a large number of different entities. This can reduce learning or training costs. Instead, the corresponding second analysis unit 24 can focus on (and in particular be limited to) determining a general speech context, in which possible entities are completely excluded or are only considered as placeholders (for example in the form of "movie" (placeholder) instead of a specific movie title (entity) within a text file representing the speech input information). The analysis block 18 can also access the knowledge database 14, such as through Figure 1 As indicated by the corresponding arrows in .

[0068] The analysis block 18 outputs the answer partial output explained below as an output variable. If the answer partial output still contains placeholders, these are replaced by matching entities in the answer output unit 21, which can also be referred to as a reconstructor, and are thus combined with the answer partial output to form the answer output A. The answer output A can in turn be or include a text file and can be output to the user via an audio output device, for example, after conversion into an audio file.

[0069] exist Figure 2 In the Figure 1 Detailed view of the system 10 in FIG. In particular, further details of the analysis block 18 can be seen therein. The speech input information E in the form of text and the determination unit 12 are also shown, which can recognize entities within the input variable E with the aid of the knowledge database 14. The speech information generation unit 16 is also shown.

[0070] It should be noted that the access of the determination unit 12 to the knowledge database 14 is completely optional and can be performed, for example, to improve the recognition accuracy. In general and without being limited to this example and its other details, the determination unit 12 can also process information, for example, from the speech input information E, in order to improve the recognition accuracy. If, for example, the information contains the name of an actor, the probability that another entity in the input is a movie title increases.

[0071] The analysis block 18 is outlined with dashed lines and its individual functional blocks are shown. First of all, the vector generation unit 20 (embedder) can be seen, with which the information set explained below can be presented as a vector representation. To this end, the information set generation unit 25 integrated into the vector generation unit 20 can access the knowledge database 14 and first determine the information associated with the determined entity or assigned to the entity according to the knowledge database 14. This can be achieved by a so-called hash request. Subsequently, this information can be aggregated into information sets (subsets) and converted into vector representations.

[0072] The latter then forms the input variable for the first analysis unit 22. The first analysis unit is designed as a "feed-forward" neural network and can select the information contained therein as target information using an information set presented as a vector representation, based on which the second analysis unit 24 explained below should determine the partial response output. For this purpose, the speech context explained below is also taken into account, which is the analysis result of the second analysis unit 24 (see Figure 2 The neural network of the first analysis unit 22 then determines, as target information, information from the information set assigned to the entity that is important, in particular with respect to the current speech context. Preferably, the target information is that information of the information set that has been classified as important with the highest probability by the first analysis unit 22 or has received the highest importance assessment from the first analysis unit 22. The target information forms an output variable of the second analysis unit 22 and is output to the first analysis unit 24.

[0073] It can also be seen that the analysis block 18 comprises a second analysis unit 24 which receives the speech information set of the generation unit 16 and the target information of the first analysis unit 22 as input variables.

[0074] The second analysis unit 24 determines the partial answer output based on these input parameters and taking into account the placeholders embedded by the speech information set generation unit 16 and the constantly determined and updated speech context. The speech context indicates in particular which type of answer output A may be important and suitable from the user's perspective. The speech context is also output to the first analysis unit 22 in order to then obtain the target information determined therefrom.

[0075] The determined partial answer output contains neither the substance of the speech input information E nor the associated target information. The latter are determined by the first analysis unit 22 in the manner described above and are only embedded in the answer output part 21, more precisely in such a way that any placeholders still present in the partial answer output can be replaced, for example, by corresponding target information.

[0076] In the following, based on Figure 3 1 to depict an exemplary flow chart of the method according to the invention. In step S1, the speech input information E is received in text form and checked by the determination unit 12 for the presence of an entity. If no entity is found, the analysis block 18 determines the response output A solely with the aid of the first analysis unit 24, wherein this is the case in Figure 3 Not shown separately.

[0077] If it is determined that an entity exists, then in step S2 the speech information set generating unit 16 generates a speech information set in which the entities in the speech input information E are replaced by placeholders respectively.

[0078] In step S2, upstream and downstream (such as Figure 3 In step S5, or at least partially in parallel therewith, the information set generation unit 25 determines the information assigned to the entity according to the knowledge database 14. This information is summarized in the above-mentioned information set and converted into a vector representation by the vector generation unit 20 in step S6.

[0079] Next, the first analysis unit 22 determines in step S7 one of the information in the information set as target information, taking into account the speech context available from the second analysis unit 24. Next, in step S8, the second analysis unit 24 determines the answer part output based on the speech information set and the target information of the first analysis unit 22. If there are still placeholders in the answer part output, these placeholders can be reconstructed by the reconstruction unit 21 with optional access to the knowledge database 14 (see Figure 1 , 2 ) are replaced by entities.

[0080] An example of voice input information is "I really liked the first James Bond movie." The name "James Bond" as the title of the movie can be identified as an entity. Therefore, [movie] or [movie] can be used as a placeholder, and all information matching the movie can be determined from the knowledge database 14 as an information set. The first analysis unit 22 determines target information from the information set, which seems reasonable in terms of the voice context, such as the names of actors (here exemplarily X and Y). Next, the second analysis unit 24 will generate a matching response partial output in view of the one or more entities determined as target information and in view of the voice context, such as a rhetorical question in the form of "Do you prefer X or Y as the actor?"

[0081] According to an embodiment which is not shown separately and is an aspect of the invention independent of the shown exemplary embodiment, at least the first analysis unit (but optionally also the second analysis unit) can be based on an expert knowledge model instead of a neural network. Advantageously, no training data for the corresponding analysis unit is then required.

[0082] Reference numerals list

[0083] 10 System

[0084] 12 Determine the unit

[0085] 14 Knowledge Database

[0086] 16 Voice information set generation unit

[0087] 18 Analysis Block

[0088] 20 Vector generation unit

[0089] 22 First Analysis Unit

[0090] 24 Second Analysis Unit

[0091] 25 Information set generation unit

[0092] A Answer output

[0093] E Voice input information.

Claims

1. A method for generating a response output (A) in response to speech input information (E), the method comprising: a) determining whether the speech input information (E) includes at least one predefined entity; And when this is the case: b) generating a speech information set in which the entity is not included based on the speech input information (E); c) generating an information set, wherein the information set comprises information in a knowledge database (14) assigned to the entity; d) using a first analysis unit (22): determining from the information set at least one target information that is relevant to the entity, wherein the first analysis unit (22) does not receive the speech information set as an input variable and / or is not configured to generate a partial response output based thereon; e) using a second analysis unit (24): generating a response part output based on the determined target information and the speech information set; and f) generating a response output (A) based on the response partial output.

2. The method according to claim 1, characterized in that In the speech information set, the entity is replaced by a placeholder, and if the placeholder is also contained in the answer part output, the placeholder is then replaced by the entity.

3. The method according to claim 1 or 2, characterized in that: The first and / or second analysis unit (22, 24) comprises a neural network.

4. The method according to claim 3, characterized in that The second analysis unit (24) comprises a neural network in the form of a sequence-to-sequence model.

5. The method according to claim 3, characterized in that: The first analysis unit (22) comprises a neural network in the form of a feed-forward network and / or a neural network with an attention mechanism.

6. The method according to claim 1 or 2, characterized in that: The first analysis unit (22) obtains the vector representation of the information set as an input parameter.

7. The method according to claim 1 or 2, characterized in that: The second analysis unit (24) is not configured to determine target information for the determined entity based on the information set in the knowledge database (14).

8. A system (10) for generating a response output (A) in response to speech input information (E), the system comprising: a) a determination unit (12) configured to determine whether the speech input information (E) includes at least one predefined entity; b) a speech information set generating unit (16), which is configured to generate a speech information set in which the determined entity is not included based on the speech input information; c) an information set generating unit (25), which is configured to generate an information set about the entity, wherein the information set contains information in the knowledge database (14) assigned to the entity; d) a first analysis unit (22), which is configured to determine from the information set at least one target information item relevant to the entity, wherein the first analysis unit (22) does not receive the speech information set as an input variable and / or is not configured to generate a partial answer output based thereon; e) a second analysis unit (24) configured to generate a response part output based on the determined target information and the speech information set; and f) A response output unit (21) configured to generate a response output based on the response partial output.

9. Use of the system (10) according to claim 8 for generating a response output in a non-goal-oriented conversation.