Method, speech dialogue system and use of a speech dialogue system to generate a response output in response to speech input information

Abstract

Method for generating a response output (A) with a speech dialog system (10) in response to speech input information (E), with: a) Determine whether a speech input piece of information (E) includes at least one predefined entity; and if so: b) Generating a speech information set based on the speech input information (E) in which the entity is not included, but is replaced by a placeholder; c) Generating an information set, wherein the information set comprises information from a knowledge database (14) assigned to the entity; d) with a first unit of analysis (22): Identify at least one target piece of information from the information set that is relevant to the entity; e) with a second analysis unit (24): generating a partial response output based on the determined target information and the language information set; f) Replacing the placeholder with the entity, the target information, or other entities, provided the placeholder is still included in the response sub-output; and g) Generating an answer output (A) based on the answer partial output.

Description

[0001] The invention relates to a method, a speech dialogue system, and the use of a speech dialogue system for generating a response output in response to speech input information. More precisely, the invention relates to the field of computer-aided speech analysis and, in particular, speech dialogue systems in which speech input information is captured and response outputs are generated based on it.

[0002] It is known to analyze speech input information using computer-aided systems, and in particular the analysis units they comprise, and to generate response outputs based on this analysis. A distinction can be made between pure speech recognition and the analysis of meaningful responses (response outputs) to the recognized speech input information, whereby the analysis can, for example, include capturing the semantic content of the speech input information. Speech input information, as used here, can be understood as information in text or audio format that was generated based on speech input and / or that specifies its content.

[0003] Within the framework of a so-called goal-oriented dialogue, a user can, for example, formulate questions to obtain desired information. Generally, dialogues can be described as goal-oriented if they lead to a defined endpoint, as is the case when requesting information. The endpoint would then be receiving the information. Likewise, there are non-goal-oriented dialogues in which a user engages in a conversation with a computer-based system (dialogue system) without a predefined endpoint, such as one determined by the desired information. This can also be described as engaging in small talk with such a system.

[0004] Handling non-goal-oriented dialogues is technically more demanding, as their progression, for example in the form of so-called dialogue states, is more difficult to predict. To be able to react to a corresponding variety of possible dialogue states and generate appropriate responses, a dialogue system would have to be able to analyze or understand a large amount of information contained in the speech input. This would mean that the system, and especially an analysis unit, would have to be trained using correspondingly extensive dictionaries or general databases. The latter applies primarily to systems based on machine learning approaches, and especially so-called deep learning. For example, such systems would have to be trained for every conceivable proper name, such as that of people or cities, or for every numerical value, e.g.In the form of years or temperature values, corresponding words must be explicitly stored in the dictionary so that the analysis unit can be trained using this dictionary and later recognize the corresponding words in order to correctly grasp the semantic content. This increases the setup effort and complicates the learning process of such an analysis unit.

[0005] Current approaches to solving this so-called "out-of-vocabulary problem" involve enabling the unit of analysis to recognize not only words already learned in a preparatory learning process, but also additional words used in the dialogue history (i.e., words used in previous interactions with the unit of analysis). However, even in this case, it is necessary to continuously record new words appearing during the dialogue history and store them in a corresponding dictionary. Examples of such approaches can be found in the following scientific publications: Raghu, D. & Gupta, N. (2018). Hierarchical Pointer Memory Network for Task Oriented Dialogue, arXiv preprint arXiv: 1805.01216 Eric, M & Manning, CD (2017). A copy-augmented sequence-to-sequence architecture gives good performance on task-oriented dialogue. arXiv preprint arXiv: 1701.04024 On the other hand, there are also systems that attempt to master non-goal-oriented dialogues iteratively and not based on machine learning (e.g., https: / / www.pandorabots.com / mitsuku / ). However, the achievable quality is significantly limited by predefined states and rules (e.g., in the form of so-called expert knowledge).

[0006] With any of the known approaches, it is therefore not possible to respond appropriately to unpredictable speech input information, particularly in the context of non-goal-oriented dialogues, especially when this speech input information contains terms not yet present in the dialogue history and / or not encountered during training. Furthermore, the known solutions typically involve a one-dimensional information flow: While the analysis unit can query a dictionary for speech recognition purposes, which can also be dynamically supplemented based on the dialogue history as described above, the entries in this dictionary (especially the dynamically supplemented ones) cannot be used to generate response outputs.

[0007] Another disadvantage of the approaches mentioned above, which are not based on machine learning, is that they still rely heavily on so-called expert knowledge or expert knowledge models as components of their analysis. This, however, deviates from the actual goal of handling non-goal-oriented dialogues, as more predefined relationships in the form of expert knowledge need to be defined and stored beforehand. Expert knowledge, in this context, can generally be understood as the definition of logical relationships or rules that allow for the determination of appropriate responses (especially response outputs) given the existing states.

[0008] Further technological background information can be found in the following publications: Unger, Christina, et al. Template-based question answering over RDF data. In: Proceedings of the 21st international conference on World Wide Web. ACM, 2012. pp. 639-648. DOI: 10.1145 / 2187836.2187923 Zheng, Weiguo, et al. Question answering over knowledge graphs: question understanding via template decomposition. Proceedings of the VLDB Endowment, July 2018, Volume 11, No. 11, pp. 1373-1386. DOI: 10.14778 / 3236187.3236192

[0009] One object of the present invention is therefore to improve the computer-based generation of response outputs in reaction to speech inputs, especially in non-goal-oriented dialogues.

[0010] This problem is solved by a method with the features of claim 1, a speech dialogue system with the features of claim 8, and a use with the features of claim 9. Advantageous embodiments are specified in the dependent claims. It is understood that all features and definitions mentioned in the introduction (alone or in any combination and unless otherwise specified or apparent) are also provided for in the present solution or may also apply to it.

[0011] A fundamental concept of the invention is to divide the information from a speech input between two analysis units, which are preferably designed separately from one another (e.g., as separate software programs, software modules, or computer program products), but which can advantageously also exchange information with each other. One of the analysis units can be configured to generate a partial response output, the latter being generated in particular in response to frequently recurring and / or common content of the speech input. Another analysis unit, on the other hand, can be configured to determine target information that is assigned to predefined entities within the speech input, with this target information then becoming part of a final response output. The target information can also be used as an input for the first-mentioned analysis unit to generate the response output.

[0012] This has the advantage that the content of the speech input (i.e., the speech input information) can be predefined as entities, which are expected to be subject to higher variability or can be associated with a multitude of unpredictable additional information. Furthermore, additional entities can be added during operation and, for example, stored in a knowledge database without requiring new learning or training processes for the speech dialogue system. This is relevant, for example, for adding new information that arises during operation (such as newly released films or the like). The entity can be, in particular, a proper name or a numerical value. Examples of information associated with an entity in the form of a film title include the director, actors involved, the genre, film awards, and the like. The associated information can itself also be an entity.

[0013] At least one of these associated pieces of information can then be identified by one of the analysis units as relevant target information in the current context. The analysis unit can be trained for this purpose (for example, using a machine learning process), specifically in such a way that it identifies the target information based on the current language context and a selection of associated information (information set or, in other words, a subset of a knowledge database). This can be understood, in particular, as the analysis unit evaluating all associated information with regard to its potential relevance (for example, in the form of a probability rating) and then selecting the piece with the highest relevance rating as the target information.

[0014] This allows for the meaningful integration of relevant additional or target information into the final response output. Therefore, only the subsequent analysis unit needs to be trained to handle information associated with predefined and diverse entities, not the analysis unit that generates the aforementioned partial response output. Furthermore, such associated information does not need to be generated by the analysis unit itself, but can be retrieved from a knowledge database. This reduces the required functionality of the analysis unit and thus also facilitates its learning process.

[0015] In other words, the invention makes it possible to differentiate between speech input content that can be subject to increased variation (referred to herein as an entity) and / or can be associated with a multitude of further information, and between speech input information or content that recurs frequently and is, for example, not a proper name. Separate analysis units can then be provided for each of these types of information / content, optimized for their respective assigned analysis tasks, and whose learning processes can be carried out in a targeted manner. Furthermore, it is advantageous to integrate a knowledge database from which one of the analysis units can then select relevant target information for inclusion in a final response output.As mentioned above, the knowledge database can also be continuously supplemented with new entities without necessarily requiring new learning processes to be carried out by the analysis units.

[0016] In detail, a method for generating a response output in reaction to speech input information (which is, for example, speech input converted into text form) is proposed, including: a) Determine whether a speech input piece of information includes at least one predefined entity; and if so: b) Generating a speech information set based on the speech input information, in which the entity is not included; c) Generating an information set (entity information set), wherein the information set comprises information from a knowledge database assigned to the entity; d) with a first unit of analysis: Identify at least one target piece of information from the information set that is relevant with regard to the (identified predefined) entity; e) with a second analysis unit: generating a partial response output based on the determined target information; and f) Generating an answer output based on the answer partial output.

[0017] In general, the procedure can be carried out using a computer and / or at least one (preferably digitally and / or electronically operated) computing unit (for example, with a graphics card, a microprocessor, or a general-purpose computer processor). For example, the procedure can be carried out with a conventional PC.

[0018] At least one of the speech input information and response output can be in the form of an audio file or generated as audio information. Additionally or alternatively, the speech input information and / or the response output could be in text form. In principle, the solution according to the invention can be used in such a way that it is embedded in a speech dialogue system that performs conversions between audio and text information. For example, so-called ASR (Automatic Speech Recognition) units can translate audio speech input into text (which then corresponds to exemplary speech input information according to the invention), which is then processed by the solution according to the invention. Additionally, a downstream TTS (Text-to-Text) unit can translate the text-based response output generated according to the invention into audio output. However, it is also possible that the solution according to the invention integrates such ASR and / or TTS units.generally includes audio and text conversion functions. Furthermore, unless otherwise specified or apparent, references to speech input information herein may include both audio input and (audio) input converted into text form (i.e., the speech input information may generally be in audio or text form, with at least temporary conversion to text form being preferred for providing the measures according to the invention).

[0019] As indicated above, an entity can be, for example, a predefined name relating to an object, place, time, organization, work, or thing, and in particular a proper name or a numerical value. Existing solutions can be used to recognize such entities, for example, so-called "named-entity recognition" or "named-entity resolution" algorithms. In this way, entities in the form of predefined proper names, such as those used to name sports teams, films, places, or people, can be recognized.

[0020] The speech information set can be generated in the form of a data record or, more generally, as digital information. The speech information set can contain the speech input information, where, as a first recognition result, the speech input information, with the exception of the entity, is at least partially available in text form. The entity can either 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, generic term, or identifier that describes the entity. It is generally preferred that the second unit of analysis takes the placeholder into account to determine the partial response output.

[0021] If one of the predefined entities is recognized in the speech input information (especially in text form), for example by means of a "named-entity-resolution" algorithm, corresponding placeholders can be inserted into the recognition result and / or into the text form in place of this entity, according to the invention. This can be done by means of a so-called constructor.

[0022] Information can be stored in the knowledge base and associated with, for example, tags, identifiers, hashes, or key-value relationships, making it retrievable. The characteristics listed above allow a connection to be established to an entity recognized (in the speech input information). For example, if a film title is recognized as the entity, associated information such as actors, a corresponding year, a filming location, or similar details can be retrieved. This information is also stored in the knowledge base and preferably linked to the entity via one of the aforementioned relationships.

[0023] In general, a hash query can be directed to the knowledge database to compile information associated with the entity in the knowledge database into an information set.

[0024] The information set can, for example, express all entries found in the information database as numerical values ​​and / or collect them 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 to identify the most relevant target information (preferably based on an analysis result from the second analysis unit as a further input).

[0025] Both the first and second analysis units can each be provided as computer program products, and in particular as computer programs or software modules. However, they can have different input and / or output variables (compared to each other) or be configured to process and generate correspondingly different input and / or output variables.

[0026] In general, it should be noted that the term analysis in connection with the units of analysis can refer to analyzing in the sense of determining a suitable response output (or at least components of a response output), determining the target information and / or capturing the semantic content of at least parts of the speech input information, unless otherwise specified or apparent.

[0027] The target information can be that information from the information set which has the highest (presumed) relevance with respect to the entity. This relevance can relate specifically to the current dialogue state or can be determined from a dialogue history (by the first analysis unit). More precisely, the first analysis unit can receive the dialogue history (here also referred to as the language context) and the information set as input. The first analysis unit can be trained, within a machine learning process, to select one of the pieces of information from the information set as the target information based on this input, or to evaluate it as particularly relevant. This information can encompass one or more entities that are considered relevant for generating the response output.The language context can be obtained, for example, from the second analysis unit. The first analysis unit can then output the target information as an input, preferably to the second analysis unit.

[0028] The second analysis unit, on the other hand, can receive the language information set (especially including any placeholders) and the determined target information as input. This second analysis unit can be trained (for example, within a machine learning process) to generate and output a partial response based on this information.

[0029] In other words, at least and preferably exactly one piece of information from the information set contained in the information set and linked to the entity can be selected as the target information. This target information preferably has the highest current relevance and / or appears most appropriate, in relation to the current language context, to become part of the response output. It is generally preferred that the second unit of analysis determines the response sub-output only after being aware of, and especially depending on, this determined target information. The target information can, but need not, become part of the response sub-output and / or the response output. In general, the target information can comprise at least one entity or a group of entities. For example, the target information can correspond to a row from a matrix discussed below, in which the entries of the information set are sorted or which constitute the information set.

[0030] For example, if a film title is mentioned in the voice input and the quality of the cast is generally emphasized, it may be relevant in the next step to select the actor's name as target information, linked to the film title recognized as an entity, and to include this name in the response output. Conversely, if the voice input clearly indicates that a user wishes to see the film yet, it may be more relevant to discuss possible venues or viewing options (i.e., to include terms like DVD, cinema, or similar in the response output).

[0031] The second analysis unit, on the other hand, can generate the response output (e.g., as text or audio output) based on the speech information set. In particular, the second analysis unit can determine a current dialogue state and / or a speech context (here synonymous with the dialogue history). Dialogue states can be, for example, predefined states, such as asking a question on a defined topic.

[0032] Generating a final response output can be accomplished using an optional reconstruction unit. Generally, the final response output can differ from the partial response output in that it contains not only the partial response output, but also, if applicable, the target information inserted and / or added therein, and / or entities that may replace placeholders still present in the partial response output.

[0033] In other words, the response output can be a combination of response sub-output and target information and / or include the target information(s) inserted into the response sub-output. According to the invention, any placeholders in the response sub-output can be replaced by the target information or by other entities. A particular technical benefit of the solution according to the invention is that relevant target information can be determined with limited effort and that sentence structures (response sub-outputs) suitable for the identified target information can be generated using the second analysis unit. Subsequently, the aforementioned merging of these contents to generate the response output can then take place.

[0034] According to the invention, it can therefore be provided that the entity in the speech information set is replaced by a placeholder, which is subsequently replaced by the entity again, provided it is still contained in the partial response output. If no corresponding placeholders are contained, it is also possible that the partial response output corresponds to the final response output. Therefore, instead of a partial response output, one could also speak of response output information, which is potentially processed further by replacing the placeholders or which already corresponds to the final response output.

[0035] A further development of the invention provides that the second analysis unit receives analysis results from the first analysis unit (e.g., in the form of the determined target information or a subset of entities that constitute the target information) in order to determine the partial response output. More precisely, these analysis results can form an input for the second analysis unit. The second analysis unit can generally be trained, or learned, to determine suitable partial response outputs based on such analysis results.

[0036] According to a further training, the first and / or second analysis unit comprises an (artificial) neural network. In particular, at least the second analysis unit can employ a multi-layered (i.e., a so-called deep) neural network, in which an input layer and output layer are linked across multiple layers. Each layer of a neural network can contain nodes in a known manner, and these nodes are linked to one or more nodes in an adjacent layer via weighted connections. The weights and connections of nodes can be learned in a known manner within a training or learning process of the neural network. In general, neural networks can define a non-linear relationship between input and output variables and thus determine suitable output variables based on provided input variables.Nonlinearity can be caused by, or accounted for by, so-called activation functions of a neural network. These networks can represent a mathematical relationship between input and output variables through corresponding nodes and layers.

[0037] In this context, it may also be provided that the second analysis unit comprises a neural network in the form of a (preferably hierarchical) sequence-to-sequence model (Seq2Seq model). This can include, in a manner known per se, at least one encoder that receives input variables (e.g., the speech information set). Furthermore, the encoder can determine a state based on the input variables and output this to a decoder of the network. This decoder can generate a partial response output of the type described herein and output this as a signal.

[0038] Alternatively or additionally, the first analysis unit can comprise a neural network in the form of a feed-forward network and / or with an attention mechanism. The feed-forward network can generally be free of feedback and thus be a simple network, in particular one with only two layers. It can receive the information set and the language context (dialogue history) as inputs, and in particular an information set encoded or converted in vector representation.

[0039] In general, the first analysis unit can identify a subset of potentially relevant information from the knowledge database and then determine the target information from this subset. This latter step is preferably performed using the attention mechanism. In principle, the attention mechanism can calculate a score for each piece of information in the subset based on the dialogue history, indicating its presumed relevance, for example, in terms of probability. Information with low scores cannot be selected as target information and has no influence on the response output, as it is not considered by the second analysis unit.

[0040] According to a training course, the first unit of analysis receives a vector representation of the information set as input. This can be achieved through a process called embedding, in which information stored in the knowledge base (that which is assigned to the entity) is converted into vector representations. The result can be the information set evaluated or examined by the first unit of analysis. Alternatively, the information set as it exists before the corresponding vector conversion can be considered the information set, and a vector representation of this information set can be generated as a result of the embedding.

[0041] Converting to a vector representation generally offers the advantage that the analysis unit can be based on a mathematical model or function and therefore process the input variable in the form of a vector representation more effectively. Semantic information can also be advantageously integrated based on well-known approaches such as "Word2Vec" or "GloVe".

[0042] According to a further embodiment, the second analysis unit is not configured to determine target information (or target entities) for the identified entities themselves, based on an information set from the knowledge database. In particular, the second analysis unit cannot be configured to determine relevant and / or associated target information (or target entities), especially not based on entries in the knowledge database and / or the information sets described above. Specifically, the analysis unit cannot be trained for this purpose and / or cannot map or recognize corresponding relationships. However, the second analysis unit can be configured and, in particular, trained to understand the target information and / or to generate a suitable partial response output based on it. The determination of the target information itself, however, is preferably performed by the first analysis unit.This allows the second analysis unit to specialize in other tasks and adjust its training effort accordingly. For example, while the second analysis unit should be able to handle placeholders or general entity classifications, such as "film title," it should not need to learn all the specific entities or proper names associated with them. Instead, 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 then preferably generates partial response outputs based only on the placeholders or entity classifications, the target information, and the language context, and is trained accordingly.

[0043] Additionally or alternatively, it may be provided that the first language analysis unit does not receive the language information set as an input and / or is not configured to independently generate a (or the) partial response output.

[0044] The invention further relates to a speech dialogue system for generating a response output in reaction to speech input information, comprising: a) an investigative unit set up to determine whether a speech input piece of information includes at least one predefined entity; b) a speech information record generation unit that is configured to generate a speech information record based on the speech input information, in which a detected entity is not included; c) an information record generation unit configured to generate an information record for the entity, wherein the information record contains information from a knowledge database associated with the entity; d) a first analysis unit designed to identify at least one target piece of information from the information set that is relevant to the entity; e) a second analysis unit designed to generate a partial response output based on the identified target information; and f) a response output unit configured to generate a response output based on the response partial output.

[0045] Finally, the invention also relates to the use of a speech dialogue system according to the previous aspect for generating response outputs in non-goal-oriented dialogues. It has been shown that the inventive provision of the two analysis units, and thus the division of the respective analysis tasks, is particularly advantageous in corresponding non-goal-oriented dialogues (between a user and a computer-based system).

[0046] The speech dialogue system can generally be configured to execute a procedure according to any of the preceding and following aspects. All further developments discussed in the context of analogous procedure features can also be provided for the corresponding system features. The system can generally be computer-based or be described as a computer-based speech dialogue system. The units of the speech dialogue system can be implemented as software modules and / or as part of a computer program product (for example, in the form of individual software functions). The speech dialogue system can optionally include a computing unit (e.g., a processor) on which the units are preferably executable, and / or a storage device on which the knowledge database is preferably stored.

[0047] One embodiment of the invention is explained below with reference to the accompanying figures. These figures depict: Fig. 1 an overview of a speech dialogue system according to the invention in one embodiment; Fig. 2 a detailed representation of the speech dialogue system from Fig. 1; and Fig. 3. A flowchart of a process using the speech dialogue system of the Fig. 1 and Fig. 2 methods according to the invention.

[0048] In Fig. Figure 1 schematically shows a speech dialogue system 10 according to one embodiment of the invention. This speech dialogue system 10 is implemented as a computer program product in which the functions or function blocks described below can be implemented by individual program modules in a manner known per se. The speech dialogue system 10 can be executed on the processing unit of a conventional PC. Only the knowledge database 14 described below, as a data collection, cannot be understood primarily as a software module or a function executable by a processing unit, but rather as a data collection stored within the speech dialogue system 10. All units of the speech dialogue system 10 mentioned herein can accordingly be implemented as a software module, computer program product, or software function.The speech dialogue system 10 can optionally also include hardware components in the form of a computing unit (processor and / or graphics card) and / or a storage device for the knowledge database.

[0049] The speech dialogue system 10 receives speech input information E as an input signal in the form of audio input already converted to text. In a detection unit 12, which is a named entity recognition module (NER - Named Entity Recognizer or Named Entity Resolver) based on known algorithms, predefined entities contained in the speech input information can be recognized. For this purpose, the detection unit 12 can access a knowledge database 14, described below, in which corresponding entities are stored. The recognition of entities in speech input information (especially when it is already in text form) is a known concept and can be practically implemented using conventional named entity recognizers, which are typically based on neural networks.

[0050] As an output variable, the detection unit 12 can define corresponding entities contained in the speech input information E. The output information of the detection unit 12 is fed to a speech information record generation unit 16, which can also be referred to as a constructor. The speech information record generation unit 16 generates a speech information record (e.g., in the form of a text file) in which the entities are each replaced by a placeholder or, more generally, in which recognized entities are not present. This is fed to an analysis block 18 as an input variable.

[0051] This has the advantage that a second analysis unit 24, explained below (see Fig. 2), which is in the Fig. The analysis block 18, which is only schematically indicated, does not need to be set up for an analysis of or based on the specific entities, but at most for handling corresponding placeholders. It is understood that the entities are subject to significantly greater variation than the placeholders, which are to be understood as a general term for a number of different entities. This can reduce the learning and training effort. Instead, the corresponding second analysis unit 24 can be specialized (and in particular limited) to determining the general language context, whereby any entities are either completely excluded or only considered as placeholders (for example, in the form of "Film" (placeholder) instead of the specific film title (entity) within a text file representing the speech input information).Analysis block 18 can also access knowledge database 14, as indicated by the corresponding arrows in . Fig. 1 indicated.

[0052] As an initial output, analysis block 18 provides a partial response output, which is explained below. If this output still contains placeholders, these are replaced by suitable entities in a response output unit 21, which can also be called a reconstructor, and thus combined with the partial response output to form a response output A. The response output A can, in turn, be a text file and, for example, be output to a user via an audio output device after conversion to an audio file.

[0053] In Fig. 2 is a detailed representation of the speech dialogue system 10 from Fig. Figure 1 shows further details of the analysis block 18. The speech input information E, which is in text form, is again visible, as is the detection unit 12, which can recognize entities within the input variable E using the knowledge database 14. The speech information generation unit 16 is also shown.

[0054] It should be noted that access by Investigation Unit 12 to Knowledge Database 14 is purely optional and can be used, for example, to increase recognition accuracy. More generally, and not limited to the present example and its further details, Investigation Unit 12 can also process information, for example, from the speech input E, to increase recognition accuracy. If this input contains, for example, actor names, the probability increases that another entity in the input is a film title.

[0055] Analysis block 18 is outlined with a dashed line, and its individual function blocks are shown. A vector generation unit 20 (embedder) is visible, which can be used to represent an information set, explained below, as a vector representation. For this purpose, an information set generation unit 25, integrated into vector generation unit 20, can access the knowledge database 14 and first determine information that, according to the knowledge database 14, is associated with or assigned to the identified entity. This can be done via a so-called hash query. This information can then be summarized as an information set (subset) and converted into a vector representation.

[0056] The latter then forms an input for a first analysis unit 22. This unit is designed as a feed-forward neural network and, using the information set represented as a vector representation, can select a piece of information contained therein as target information. Based on this target information, the second analysis unit 24, described below, is to determine a partial response output. For this purpose, the language context, described below, is also taken into account, as it is an analysis result of the second analysis unit 24 (see the corresponding arrow-shaped connections between these units in [reference]). Fig. 2) Subsequently, the neural network of the first analysis unit 22 identifies a piece of information from the information set assigned to the entity that is particularly relevant in light of the current language context and serves as the target information. This is preferably the piece of information from the information set that the first analysis unit 22 is most likely to classify as relevant or that has received the highest relevance rating 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.

[0057] Furthermore, it can be seen that the analysis block 18 includes the second analysis unit 24, which receives the language information set of the generation unit 16 and the target information of the first analysis unit 22 as input.

[0058] The second analysis unit 24 determines a partial response output based on these input variables and taking into account placeholders inserted by the language information set generation unit 16, as well as a continuously determined and updated language context. The language context indicates, in particular, what type of response output A would likely be relevant and appropriate from the user's perspective. This language context is also output to the first analysis unit 22 in order to subsequently obtain the target information determined by it.

[0059] The determined response output does not contain either the entities of the speech input information E or the associated target information. The latter are determined by the first analysis unit 22 in the manner described above and are only integrated in the response output part 21, in such a way that any placeholders still present in the response output can be replaced, for example, by corresponding target information.

[0060] The following will be based on Fig. 3 An exemplary flowchart of a method according to the invention is described. In step S1, the speech input information E is received in text form and checked by the detection unit 12 for existing entities. If no entities are detected, the analysis block 18 determines the response output A solely with the help of the second analysis unit 24, whereby this case in Fig. 3 is not shown separately.

[0061] If the presence of an entity is determined, in step S2 a language information set is generated by the language information set generation unit 16, in which entities in the language input information E are each replaced by a placeholder.

[0062] Upstream to step S2, downstream (as in Fig. 3) or at least partially in parallel, the information record generation unit 25 determines in step S5 information that is assigned to the entity according to the knowledge database 14. This information is summarized in the aforementioned information record and converted into a vector representation in step S6 by the vector generation unit 20.

[0063] The first analysis unit 22 then determines, in step S7, taking into account the language context that can be obtained from the second analysis unit 24, one of the pieces of information from the information set as the target information. In step S8, the answer partial output is then determined by the second analysis unit 24 based on the language information set and the target information from the first analysis unit 22. If placeholders are still present in the answer partial output, these can be replaced by the reconstruction unit 21 with optional access to the knowledge database 14 (see corresponding arrows in Fig. 1, Fig. 2) be replaced by entities.

[0064] An example of speech input information is "I really like the very first James Bond film." The entity "James Bond" can be recognized as the film title. Therefore, [Film] or [movie] can be used as a placeholder, and all relevant information about this film can be retrieved from the knowledge database 14 as an information set. The first analysis unit 22 identifies target information from the information set that appears relevant in light of the language context, for example, actor names (here, for example, X and Y). The second analysis unit 24 then generates a suitable response output based on this identified target information and the language context, for example, a follow-up question such as "Did you prefer X or Y as the actor?"

[0065] According to an embodiment not shown separately, which represents an aspect of the invention independent of the illustrated embodiments, at least the first analysis unit (and optionally also the second) can be based on an expert knowledge model instead of a neural network. Advantageously, no training data is then required for a corresponding analysis unit. Reference symbol list 10 Voice Dialogue System 12 Investigation Unit 14 Knowledge database 16 Speech Information Set Generation Unit 18 Analysis block 20 vector generation unit 22 first analysis unit 24 second analysis unit 25 Information Set Generation Unit A Answer Output E Language input information

Claims

[1] Method for generating a response output (A) with a speech dialogue system (10) in response to speech input information (E), comprising: a) determining whether a speech input information (E) comprises at least one predefined entity; and if so: b) generating a language information record based on the language input information (E) in which the entity is not included but is replaced by a placeholder; c) generating an information set, wherein the information set comprises information associated with the entity from a knowledge database (14); d) with a first analysis unit (22): determining at least one item of target information from the information set that is relevant to the entity; e) with a second analysis unit (24): generating a partial response output on the basis of the determined target information and the speech information set; f) replacing the placeholder with the entity, with the target information or with other entities, provided that the placeholder is still contained in the response suboutput; and g) generating a response output (A) based on the response partial output. [2] Method according to claim 1, characterized by that the first and / or the second analysis unit (22, 24) comprise a neural network. [3] Method according to claim 2, characterized by that the second analysis unit (24) comprises a neural network in the form of a sequence-to-sequence model. [4] Method according to claim 2, characterized by 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. [5] Method according to one of the preceding claims, characterized by that the first analysis unit (22) receives as input a vector representation of the information set. [6] Method according to one of the preceding claims, characterized by that the second analysis unit (24) is not configured to determine target information on the determined entities on the basis of an information set from the knowledge database (14). [7] Method according to one of the preceding claims, characterized by that the first analysis unit (22) does not receive the speech information set as an input variable and / or is not configured to generate a response partial output based thereon. [8] Speech dialogue system (10) for generating a response output (A) in response to speech input information (E), comprising: a) a determination unit (12) configured to determine whether a voice input information item (E) comprises at least one predefined entity; b) a language information set generation unit (16) which is configured to generate a language information set on the basis of the language input information in which a determined entity is not included but is replaced by a placeholder; c) an information set generation unit (25) configured to generate an information set relating to the entity, the information set containing information associated with the entity from a knowledge database (14); d) a first analysis unit (22) configured to determine at least one item of target information from the information set that is relevant to the entity; e) a second analysis unit (24) which is arranged to generate a partial response output on the basis of the determined target information and the speech information set, f) a response output unit (21) configured to generate a response output based on the partial response output; and further configured to replace the placeholder with the entity, with the target information, or with other entities, provided that the placeholder is still contained in the partial response output. [9] Using a speech dialogue system (10) according to claim 8 for generating response outputs in non-target-guided dialogues.

Images