Using generative adversarial network to train semantic parser of dialog system

Abstract

To provide a method for using a generative adversarial network (GAN) to train a semantic parser of a dialog system, a training system, a dialog system, and a program.SOLUTION: The method includes: accessing seed data including seed tuples, each of the seed tuples including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance; and training a semantic parser and a discriminator in a GAN. The semantic parser learns to map utterances to logical forms based on output from the discriminator, and the discriminator learns to recognize authentic logical forms based on output from the semantic parser. The semantic parser may then be integrated into a dialog system.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS This disclosure is incorporated herein by reference in its entirety. This provisional application claims priority to US Provisional Application No. 60 / 133,363, filed on May 13, 2006, which is incorporated herein by reference in its entirety.

[0002] Technical Field The present disclosure relates to dialogue systems, and more particularly to techniques for using generative adversarial networks to train semantic parsers for use in dialogue systems, such that the semantic parsers effectively learn to generate logical forms based on the adversarial context of the generative adversarial network. [Background technology]

[0003] An increasing number of devices now allow users to directly interact with the device using voice or spoken language. For example, a user can speak to such a device in natural language, in which the user can ask a question or make an utterance that requests an action to be performed. In response, the device uses an acoustic output to perform the requested action or answer the user's question. Because direct interaction using voice is a more natural and intuitive way for people to communicate with their surroundings, the popularity of such voice-based systems is growing at an astronomical rate. Summary of the Invention

[0004] A dialogue system is a speech-enabled system that can interact with a user, such as through speech input and acoustic output. Typically, a dialogue system includes one or more machine learning (ML) models, such as a semantic parser. The present disclosure relates to an approach for using a generative adversarial network (GAN) to train a semantic parser of a dialogue system. For example, in some embodiments, a GAN includes a semantic parser as its generator, and further includes a discriminator and an error minimization module. In some embodiments, the semantic parser and the discriminator are ML models that are trained as adversarial, such that both ML models are improved together.

[0005] In some embodiments, seed data is used as a training input to the GAN. The seed data may include a set of seed tuples, each of which includes an utterance and a logical form corresponding to the utterance. One embodiment described herein trains both the semantic parser and the classifier by first fixing the semantic parser while training the classifier based on the semantic parser output, then fixing the classifier while training the semantic parser based on the classifier output, and repeating until the seed data is exhausted.

[0006] While the semantic parser is fixed and therefore static, the semantic parser may generate logical forms based on randomly selected utterances. Each pair of a randomly selected utterance and a corresponding logical form as generated by the semantic parser may together form a generated tuple. The classifier may take as input the generated tuples. along with the seed tuples. For each such tuple, the classifier may output (i.e., predict) the probability that the tuple is genuine, or in other words, the probability that the tuple is a seed tuple. The error minimization module may generate a training signal to train the classifier based on the discrepancy of the classifier's prediction compared to an exact distribution (i.e., an exact sequence of values) that indicates whether the tuple is genuine.

[0007] If the classifier is fixed and therefore static, the semantic parser may generate logical forms based on utterances selected from seed tuples of the seed data. Each pair of an utterance from the seed data and a corresponding logical form as generated by the semantic parser may together form a tuple. The classifier may receive such tuples as input and may predict, for each tuple, the probability that the tuple is genuine. The error minimization module may generate a training signal to train the semantic parser based on the difference between the classifier's actual predictions of the genuineness compared to a desired set of predictions in which the generated logical forms are all predicted to be genuine. In other words, the semantic parser is trained with the goal of fooling the classifier into believing that the generated logical forms are all genuine.

[0008] As a result, the semantic parser and the classifier may be trained together as components of a GAN, where the semantic parser learns to generate logical forms and the classifier learns to recognize logical forms. Eventually, the classifier improves and so does the semantic parser, and the semantic parser becomes so good at generating logical forms that the generated logical forms are indistinguishable from logical forms that belong to the seed data. In this way, the semantic parser is trained to generate logical forms based on utterances.

[0009] In other embodiments, the dialogue system may include a voice input module for receiving voice input, a voice output module for outputting voice output to a user, and a dialogue processing module for taking or causing to take one or more actions based on an interpretation of the voice input and preparing an appropriate response to the user. The dialogue processing module may utilize a semantic parser obtained by using the above-mentioned technique for using GANs to train a semantic parser of the dialogue system.

[0010] The foregoing, together with other features and embodiments, will become more apparent with reference to the following specification, claims, and accompanying drawings. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram of a dialogue system incorporating a semantic parser trained in a generative adversarial network, according to certain embodiments described herein. [Diagram 2] FIG. 1 is a diagram of a training system configured to train a semantic parser and a classifier in a generative adversarial network, according to certain embodiments described herein. [Diagram 3] FIG. 1 is a diagram of a method for training a semantic parser together with a classifier in a generative adversarial network, according to certain embodiments described herein. [Figure 4] 4 is a diagram of a method for training a classifier as part of a training round in the method of FIG. 3, according to certain embodiments described herein. [Diagram 5] 4 is a diagram of a method for training a semantic parser as part of a training round in the method of FIG. 3 according to certain embodiments described herein. [Figure 6] FIG. 1 is a diagram of a distributed system for implementing certain embodiments described herein. [Figure 7]FIG. 1 is a block diagram of a cloud-based system environment in which training a semantic parser in a generative adversarial network according to certain embodiments described herein may be provided at least in part as a cloud service. [Figure 8] FIG. 1 is a block diagram of an example computer system that can be used to implement certain embodiments described herein. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] In the following description, for purposes of explanation, specific details are set forth in order to provide a thorough understanding of certain embodiments. It will be apparent, however, that various embodiments may be practiced without these specific details. The drawings and description are not intended to be limiting. The word "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any embodiment or design described herein as "exemplary" or "example" is not necessarily to be construed as preferred or advantageous over other embodiments or designs.

[0013] Voice-enabled systems that can interact with a user via speech input and acoustic output, also referred to as voice output, can come in a variety of forms. For example, such systems can be provided as standalone devices, as digital or virtual assistants, as voice-enabled services, etc. In each of these forms, the system can receive speech input, understand the speech input, generate a response or take an action in response to the speech input, and output the response using an acoustic output. In an embodiment, the dialogue functionality in such voice-enabled systems is provided by a dialogue system or infrastructure ("dialogue system"). The dialogue system is configured to receive speech input, interpret the speech input, maintain a dialogue, possibly take or cause one or more actions to be taken based on the interpretation of the speech input, prepare an appropriate response, and output the response to the user using an acoustic output.

[0014] Traditionally, a dialogue system includes various machine learning (ML) models, and a large amount of training data is required to train the various machine learning models of the dialogue system. One such ML model is a semantic parser subsystem, also called a semantic parser. In general, a semantic parser receives an utterance representing a voice input provided by a user, where the utterance is a textual representation of a natural language. The semantic parser maps the utterance to a logical form, where the logical form is a representation of the utterance that is converted into a logical-type language that conforms to an established grammar and is therefore parsable by a dialogue manager subsystem of the dialogue system. The dialogue manager subsystem then parses and processes the logical form to determine how to respond.

[0015] To train the semantic parser subsystem, a training system associated with the dialogue system typically requires a large number of tuples, each tuple including an utterance (e.g., a textual representation of a natural language) and a corresponding logical form. If an insufficient amount of training data is used, the semantic parser may not be as effective as desired and may introduce errors into the logical form and therefore into the workflow of the dialogue system. As a result of introducing errors into the logical form, the dialogue manager subsystem may not be able to generate an appropriate response to the user.

[0016] A generative adversarial network (GAN) is an architecture that trains a generator and a classifier in an adversarial manner. GANs are typically used in image analysis, especially in the field of machine vision. In an exemplary conventional GAN, an image generator generates images that are intended to be in a particular class (e.g., an image of a face), and a classifier determines whether the generated image is indeed in that class. By comparing the outputs of the image generator and the classifier with the exact distribution of outputs, the GAN provides training signals to the image generator and the classifier to train them. GANs can be used to determine whether the available training data is sufficient for individual training. This is particularly useful when it is smaller than ideal, since the adversarial nature can result in effective training with a relatively small set of training data.

[0017] In some embodiments of the training system described herein, a semantic parser of a dialogue system is trained in a GAN. Specifically, the semantic parser may behave as a generator of the GAN and thus may learn to generate logical forms, and the discriminator may learn to distinguish between authentic and inauthentic logical forms. An error minimization module may apply one or more objective functions to provide training signals for training the semantic parser and the discriminator. After training, the semantic parser may be used to map utterances to logical forms as part of a dialogue system. This aspect of training a semantic parser may be more effective than traditional approaches of training a semantic parser individually due to the adversarial nature of training in a GAN.

[0018] FIG. 1 is a diagram of an example of a dialogue system 100 according to an embodiment described herein that utilizes a semantic parser 114 trained in a GAN 155. The dialogue system 100 is configured to receive speech input 104, also referred to as voice input, from a user 102, such as through a voice input module. For example, the voice input module receives speech input from a microphone or other device capable of acquiring speech or voice. The dialogue system 100 may then interpret the speech input 104. The dialogue system 100 may maintain a dialogue with the user 102 and may take or cause to be taken one or more actions based on the interpretation of the speech input 104, perhaps by using a dialogue processing module that utilizes the semantic parser 114. The dialogue system 100 may prepare an appropriate response, such as through the use of the dialogue processing module, and may output the response to the user, such as through a voice output module, using a voice or audio output, also referred to as an acoustic output. For example, the audio output module may be a speaker or other device capable of outputting voice or audio. The dialogue system 100 is a special-purpose computing system that may be used to process large amounts of data, possibly using many computer processing cycles. The number of devices depicted in Fig. 1 is provided for illustrative purposes. A different number of devices may be used. For example, although each device, server, and system in Fig. 1 is shown as a single device, multiple devices may be used instead.

[0019] In one embodiment, the processing performed by the dialogue system 100 includes a speech input component 105, a wake-word detection (WD) subsystem 106, , an automatic speech recognition (ASR) subsystem 108, a named entity recognizer (NER) subsystem 112, and a semantic parser The system is realized by a pipeline of components or subsystems including a natural language understanding (NLU) subsystem 110 including a voice processing subsystem 114, a dialog manager (DM) subsystem 116, a natural language generator (NLG) subsystem 118, a text-to-speech (TTS) subsystem 120, and an audio output component 124. The subsystems listed above may be realized exclusively in software (e.g., using code, programs, or instructions executable by one or more processors or cores), exclusively in hardware, or in a combination of hardware and software. In some implementations, one or more of the subsystems may be combined into a single subsystem. Additionally or alternatively, in some implementations, functionality described herein as being performed by a particular subsystem may be realized by multiple subsystems.

[0020] The audio input component 105 includes hardware and software configured to receive the audio input 104. In some instances, the audio input component 105 may be part of the dialogue system 100. In some other cases, the voice input component 105 may be separate from and communicatively coupled to the dialogue system 100. The voice input component 105 may include, for example, a microphone coupled to software configured to digitize and transmit the voice input 104 to the wake word detection subsystem 106.

[0021] The wake word detection (WD) subsystem 106 is configured to listen and monitor a stream of acoustic input for input corresponding to a special sound or word or set of words called a wake word. Upon detecting a wake word for the dialogue system 100, the WD subsystem 106 is configured to activate the ASR subsystem 108. In one implementation, a user may be given the ability to activate or deactivate the WD subsystem 106 (e.g., by pressing a button) to cause the WD subsystem 106 to listen or stop listening for the wake word. When activated or operating in an active mode, the WD subsystem 106 is configured to continuously receive an acoustic input stream and process the acoustic input stream to identify an acoustic input, such as the speech input 104, that corresponds to the wake word. When an acoustic input corresponding to the wake word is detected, the WD subsystem 106 activates the ASR subsystem 108.

[0022] As described above, the WD subsystem 106 activates the ASR subsystem 108. In some implementations of the dialogue system 100, mechanisms other than wake word detection may be used to trigger or activate the ASR subsystem 108. For example, in some implementations, a push button on the device may be used to trigger the ASR subsystem 108 without requiring a wake word. In such implementations, it is not necessary to provide the WD subsystem 106. When the push button is pressed or activated, the voice input 104 received after the button activation is provided to the ASR subsystem 108 for processing. Additionally or alternatively, in some implementations, the ASR subsystem 108 may be activated upon receiving an input to be processed.

[0023] The ASR subsystem 108 is configured to receive and monitor the speech input 104 after a trigger or wake-up signal (e.g., a wake-up signal may be sent by the WD subsystem 106 upon detecting a wake word in the speech input 104, or a wake-up signal may be received upon activation of a button) and convert the speech input 104 to text. As part of its processing, the ASR subsystem 108 performs speech-to-text conversion. The speech input 104 may be in a natural language format, and the ASR subsystem 108 is configured to generate corresponding natural language text in the language of the speech input 104. This corresponding natural language text is referred to herein as an utterance. For example, the speech input 104 received by the ASR subsystem 108 may include one or more words, phrases, clauses, sentences, questions, etc. The ASR subsystem 108 is configured to generate an utterance for each spoken clause and provide the utterance to the NLU subsystem 110 for further processing.

[0024] The NLU subsystem 110 receives utterances generated by the ASR subsystem 108. The utterances received by the NLU subsystem 110 from the ASR subsystem 108 may include text utterances corresponding to spoken words, phrases, clauses, etc. The NLU subsystem 110 converts each utterance or sequence of utterances into a corresponding logical form.

[0025] In one implementation, the NLU subsystem 110 includes a named entity recognizer (NER) subsystem 112 and a semantic parser subsystem 114. The NLU subsystem 110 receives an utterance as input, identifies named entities in the utterance, and tags the utterance with information related to the identified named entities. The tagged utterance is then provided to the semantic parser subsystem 114, which is configured to generate a logical form for each tagged utterance, or for a set of tagged utterances. The logical form generated for an utterance may identify one or more intents corresponding to the utterance. An intent for an utterance identifies a purpose of the utterance. Examples of intents include "order pizza" and "find directions." An intent may, for example, identify an action requested to be performed. In addition to the intent, the logical form generated for an utterance may also identify slots, also called parameters or arguments, for the identified intent. For example, for a speech input of "I want to order a large pepperoni pizza with mushrooms and olives," the NLU subsystem 110 can identify the intent to order pizza. The NLU subsystem can also identify and fill slots (e.g., pizza_size (filled with L) and pizza_toppings (filled with mushrooms and olives)). The NLU subsystem 110 may generate the logical forms using machine learning based techniques, rules, or a combination of machine learning techniques and rules, which may be domain specific. The logical forms generated by the NLU subsystem 110 are then provided to the DM subsystem 116 for further processing.

[0026] 1, in some embodiments, a training system 150 described herein trains the semantic parser 114 as part of a generative adversarial network 155 to prepare the semantic parser subsystem 114, also referred to herein as the semantic parser 114, for its operation in the dialogue system 100. For example, the training system 150 utilizes the GAN 155 to train the semantic parser 114 to perform the above-mentioned task of determining a logical form based on one or more utterances. In some embodiments, the use of this adversarial approach to training the semantic parser 114 provides an improvement over conventional dialogue systems.

[0027] The DM subsystem 116 is configured to manage a dialogue with a user based on the logical form received from the NLU subsystem 110. As part of the dialogue management, the DM subsystem 116 is configured to track the dialogue state and initiate the execution of one or more actions or tasks or itself execute one or more actions or tasks to determine how to interact with the user. These actions may include, for example, querying one or more databases, generating execution results, or other actions. For example, the DM subsystem 116 is configured to interpret the intent identified in the logical form received from the NLU subsystem 110. Based on the interpretation, the DM subsystem 116 may initiate one or more actions that it interprets as being requested by the voice input 104 provided by the user. In an embodiment, the DM subsystem 116 performs dialogue state tracking based on the current and past voice input 104 as well as a set of rules (e.g., dialogue policies) configured for the DM subsystem 116. These rules may identify different dialogue states, conditions for transitions between states, actions to be taken when in a particular state, etc. These rules may be domain-specific. The DM subsystem 116 also generates responses to be communicated to users involved in the interaction. These responses may be based on the actions and their outcomes initiated by the DM subsystem 116. The responses generated by the DM subsystem 116 are provided to the NLG subsystem 118 for further processing.

[0028] The NLG subsystem 118 is configured to generate natural language texts corresponding to the responses generated by the DM subsystem 116. These texts may be generated in a format that allows them to be converted to speech by the TTS subsystem 120. The TTS subsystem 120 receives text from the NLG subsystem 118 and converts each of them into speech or voice audio, which may then be output as audio to a user via an audio or voice output component 124 (e.g., a speaker or a communication channel coupled to an external speaker) of the dialogue system. In some cases, the voice output component 124 may be part of the dialogue system 100. In some other cases, the voice output component 124 may be separate from the dialogue system 100 and communicatively coupled to the dialogue system 100.

[0029] As described above, the various subsystems of the dialogue system 100 working in concert provide functionality that allows the dialogue system 100 to receive speech input 104 and respond using speech output 122, thereby maintaining a dialogue with a user using natural language speech. The various subsystems described above may be implemented using a single computer system or using multiple computer systems working in concert. For example, for a device implementing a voice-enabled system, the subsystems of the dialogue system 100 described above may be implemented entirely on the device with which the user interacts. In some other implementations, some components or subsystems of the dialogue system 100 may be implemented on the device with which the user interacts, while other components may be implemented remotely from the device, perhaps on some other computing device, platform, or server.

[0030] 2 is a diagram of a training system 150 configured to train the semantic parser 114 of the dialogue system 100 through the use of a generative adversarial network 155, according to certain embodiments described herein. More specifically, in some embodiments, the semantic parser 114 acts as a generator of the GAN 155, and the training system 150 trains the semantic parser 114 along with a classifier 250 in the GAN 155. An error minimization module 260 in the GAN 155 may provide one or more training signals to the semantic parser 114 and the classifier 250 to train the semantic parser 114 and the classifier 250, and more specifically may use backpropagation to train the semantic parser 114. The data input subsystem 240 of the training system 150 may provide input, such as input based on the seed data 210, to the semantic parser 114, the classifier 250, and the error minimization module 260 to enable the semantic parser 114, the classifier 250, and the error minimization module 260 to perform the tasks described herein. After training, the semantic parser 114 may be used as part of a dialogue system 100 such as that shown in FIG.

[0031] In some embodiments, the training system 150 is implemented as a computing device, or a portion thereof, such as a server. The training system 150 may be implemented as a dedicated hardware device, or as program code, or a combination of both. For example, the operations described herein as being performed by the training system 150 may be embodied in program code implementing the training system 150, where such program code is executable by one or more processors. For example, the semantic parser 114, the classifier 250, the error minimization module 260, and other aspects of the GAN 155 may each be implemented as one or more software functions or dedicated hardware devices and may operate together to implement the training system 150 as described herein. The semantic parser 114 and the classifier 250 may each be implemented as a respective machine learning model, such as a neural network.

[0032] As shown in FIG. 2, the training system 150 may include a GAN 155, which includes a generator and a discriminator 250 that are trained together. In particular, the generator is a The semantic parser 114 may be the semantic parser 114 of the dialogue system 100, and thus the semantic parser 114 is trained within the GAN 155 before or during integration into the dialogue system 100. Specifically, the training system 150 may train the semantic parser 114 to map an utterance 220 to a logical form 230, and the training system 150 may train the discriminator 250 to determine whether the logical form 230 is authentic (i.e., shares a seed tuple in the set of seed data 210 with its corresponding utterance 220). Also, the error minimization module 260 updates the semantic parser 114 and the discriminator 250 to minimize the error between the expected prediction and the actual prediction made in the GAN 155. In some embodiments, as part of the GAN 155, the semantic parser 114 and the classifier 250 are trained together such that improving one of the semantic parser 114 and the classifier 250 improves the other. As a result, the semantic parser 114 learns to map utterances 220, such as utterances 220 tagged by the NER subsystem 112, to logical forms 230 usable by the DM subsystem 116.

[0033] In some embodiments, the semantic parser 114 uses a sequence-to-sequence (seq2seq) model to determine the logical form 230 based on the utterance 220. A neural network such as a GAN 155 is a neural network that trains semantic parsers, such as a GAN 155. Traditionally, semantic parsers are trained using training data that includes a set of tuples, where each tuple has an utterance and a corresponding logical form. However, as described herein, in some embodiments, the semantic parser 114 is trained as part of the GAN 155. As described further herein, the training utilizes seed data 210 that includes seed tuples, where each seed tuple has an utterance 220 and a corresponding logical form 230. However, the training also incorporates an output from a classifier 250.

[0034] The identifier 250 may be a binary classifier that maps tuples to classes; for example, the identifier 250 may be implemented as a neural network. The identifier 250 may receive as input a tuple that includes the utterance 220 and the logical form 136, and the identifier 250 may output an indication of whether the logical form 230 is an authentic representation of the utterance 220. For example, the identifier 250 may output a probability that the logical form 230 is authentic (i.e., paired with the utterance 220 in the seed data 210) and therefore accurate. In some embodiments, to achieve this, the identifier 250 may be trained in the GAN 155 together with the semantic parser 114, as described herein.

[0035] The error minimization module 260 may provide training signals to train the identifier 250 and, by backpropagation, to train the semantic parser 114. For example, to train the identifier 250, the error minimization module 260 may utilize a first objective function 265 to compare the output from the identifier 250, which is an indication of whether the received logical forms 230 are authentic, to a precise distribution of indications of authenticity for those logical forms 230. The error minimization module 260 may use the results of the first objective function 265 to train the identifier 250 to make better predictions of authenticity.

[0036] To train the semantic parser 114, the error minimization module 260 may utilize a second objective function 265, which may be the same as the first objective function 265, to compare the output from the classifier 250, which is indicative of a prediction of the authenticity of the logical forms 230 output by the semantic parser 114, to a selected distribution (i.e., a desired distribution) of indicia of authenticity of those logical forms 230. The selected distribution may be a distribution that indicates that all of the logical forms 230 output by the semantic parser 114 are authentic. This means that in some embodiments, the training system 150 may select a distribution whose outputs are always considered to be authentic. The objective function 265 is to provide a semantic parser 114 that is consistent with the objective function 265 of the GAN 155. One skilled in the art would understand how to configure such an objective function 265. In some embodiments, the semantic parser 114 and the classifier 250 are trained together in the GAN 155, but only one of the semantic parser 114 and the generator is actively updated at a given time based on the training signal from the error minimization module 260.

[0037] 3 is a diagram of a method 300 for training a semantic parser 114 together with a classifier 250 in a GAN 155, according to certain embodiments described herein. In some embodiments, the method 300, or a similar method, is performed before the semantic parser 114 is used in a dialogue system 100. Through training as described herein, the semantic parser 114 may learn to convert (i.e., map) utterances 220 to logical forms 230 that represent such utterances 220, which is a task performed by the semantic parser 114 as part of the dialogue system 100.

[0038] The method 300 depicted in FIG. 3, and other methods described herein, may be implemented in software (e.g., as code, instructions, or programs) executed by one or more processing units (e.g., processors or processor cores), in hardware, or a combination thereof. The software may be stored on a non-transitory storage medium, for example, on a memory device. The method 300 is intended to be exemplary and non-limiting. Although FIG. 3 depicts various activities occurring in a particular sequence or order, this is not intended to be limiting. In an embodiment, for example, these activities may be performed in a different order, or one or more activities of the method 300 may be performed in parallel. In an embodiment, the method 300 may be performed by the training system 150.

[0039] 3, in block 305, the data entry subsystem 240 of the training system 150 obtains seed data 210. The seed data 210 may include a set of seed tuples, each seed tuple including an utterance 220 and a corresponding logical form 230. In each seed tuple, the logical form 230 is an accurate and authentic representation or transformation of the utterance 220. It will be appreciated that there are various approaches to obtaining the seed data 210. For example, the seed data 210 may be determined manually, such as through crowdsourcing, and then stored in a memory device accessible by the data entry subsystem 240.

[0040] At block 310, the training system 150 begins a new round of training. Block 310 is the beginning of an iterative loop. In some embodiments, during each iteration of the loop, the training system 150 performs a round of training. Each round of training may include training the classifier 250 during one part of the round and training the semantic parser 114 during another part of the round. These two parts of the round may be performed in parallel, but in some embodiments, these two parts are performed sequentially so that the training of the semantic parser 114 in the second part can benefit from the training of the classifier 250 in the first part. Also, the training system 150 may utilize a different subset of the seed data 210 for each round of training the classifier 250 and the semantic parser 114, and thus for each iteration of the loop.

[0041] At block 315, the training system 150 trains the classifier 250 based on the first portion of the seed data 210 obtained at block 305. In some embodiments, this first portion of the seed data 210 has not yet been used for training in a previous round. During this training of the classifier 250, the semantic parser 114 updates the classifier 250 based on the classifier's predictions regarding the authenticity of the output from the semantic parser 114. may be fixed to allow without changing how the semantic parser 114 determines its output. The activities involved in training the classifier 250 during a training round are described in more detail below with reference to FIG.

[0042] At block 320, the training system 150 trains the semantic parser 114 based on the second portion of the seed data 210 obtained at block 305. In some embodiments, this second portion of the seed data 210 is different from the first portion used at block 310 (e.g., there are no overlaps in the seed tuples) and has not yet been used for training in a previous round. During this training of the semantic parser 114, the classifier 250 may be fixed to allow updates of the semantic parser 114 based on the classifier's predictions regarding the authenticity of the output from the semantic parser 114 without changing how the classifier 250 determines its authenticity. The activities involved in training the semantic parser 114 during a training round are described in more detail below with reference to FIG. 5.

[0043] At decision block 325, the training system 150 determines whether all of the seed data 210 obtained at block 305 has been selected and used for any round of training (i.e., any iteration of the loop) as described above. If not all of the seed data 210 has been used, the method 300 may return to block 310, where another round of training begins. However, if all of the seed data 210 has been used in training, the method 300 may end at block 330, at which point the semantic parser 114 may already be trained and ready for use in the dialogue system 100.

[0044] 4 is a diagram of a method 400 for training a classifier 250 as part of a round of training in a GAN 155, according to certain embodiments described herein. In particular, the training system 150 may perform this method 400, or a similar method, at block 315 of the method 300 described above, to train the classifier 250 within a single round of training. Thus, according to some embodiments, this method 400, or a similar method, may be performed once per iteration.

[0045] The method 400 depicted in FIG. 4, and other methods described herein, may be implemented in software (e.g., as code, instructions, or programs) executed by one or more processing units (e.g., processors or processor cores), in hardware, or a combination thereof. The software may be stored on a non-transitory storage medium, for example, on a memory device. The method 400 is intended to be exemplary and non-limiting. Although FIG. 4 depicts various activities occurring in a particular sequence or order, this is not intended to be limiting. In an embodiment, for example, these activities may be performed in a different order, or one or more activities of the method 400 may be performed in parallel. In an embodiment, the method 400 may be performed by the training system 150.

[0046] At block 405, the data input subsystem 240 of the training system 150 selects a first portion of the seed data 210, where the first portion has not yet been used in a previous round of training. In some embodiments, this first portion of the seed data 210 is a subset of the seed tuples in the seed data 210. For example, the first portion of the seed data 210 may be a proper subset of the seed tuples in the seed data 210, such that not all of the seed tuples of the seed data 210 are included in the first portion.

[0047] At block 410, the semantic parser 114 is frozen while the identifier 250 is open for updates. In other words, in some embodiments, the semantic parser 114 will remain static for the time being to allow the identifier 250 to be trained based on the current state of the semantic parser 114.

[0048] In some embodiments, blocks 415 and 420 are performed in parallel. At block 415, the data input subsystem 240 of the training system 150 inputs a stream of utterances 220 to the semantic parser 114, which generates a logical form 230 based on the stream of utterances 220 input to the semantic parser 114. The utterances 220 in the stream of utterances 220 may not be from the seed data 210, but may be, for example, utterances 220 randomly generated or randomly selected by the data input subsystem 240 or by some other component. By way of example and not by way of limitation, the utterances 220 may be selected from a corpus (e.g., one or more books or articles) written in a natural language. For each such utterance 220 received, the semantic parser 114 may generate a logical form 230. Initially in training, the output of the logical form 230 of the semantic parser may be poor. For example, the generated logical forms 230 may be random (e.g., a random arrangement of words and symbols). However, as training progresses throughout a round or across multiple rounds, the semantic parser 114 may be improved.

[0049] At block 420, which may occur in parallel with block 415, the data input subsystem 240 of the training system 150 may input a stream of tuples to the identifier 250, which may generate predictions of authenticity for those tuples. Each tuple in the stream of tuples may include an utterance 220 and a logical form 230. The data input subsystem 240 may provide this stream of tuples as a combination (e.g., random combination, or any combination) of (1) seed tuples from the first portion of the seed data 210 selected at block 405 and (2) tuples having logical forms 230 generated by the semantic parser 114 at block 415. Based on the stream of tuples, the identifier 250 may output predictions of authenticity corresponding to the tuples. In other words, given a tuple, the identifier 250 may predict whether the logical form 230 in the tuple is a genuine representation of the utterance 220 in the tuple, or in other words, the identifier 250 may infer whether the tuple is part of the seed data 210 rather than the logical form 230 generated by the semantic parser 114. The tuples received by the identifier 250 may include both seed tuples in the first portion of the seed data 210 and generated tuples, each of which includes the utterance 220 received by the semantic parser 114 and the corresponding logical form 230 generated by the semantic parser 114 in block 415.

[0050] In some embodiments, the classifier 250 is not informed which tuples are from the seed data 210 and therefore authentic, and which tuples are from the semantic parser 114 and therefore generated. For each tuple, the classifier 250 may classify the tuple based on its possible authenticity. To this end, the classifier 250 may output a score indicative of a judgment (i.e., a prediction) of the likelihood that the logical form 230 is an authentic representation of the corresponding utterance 220 in the tuple. For example, the score is the probability that the logical form 230 is an authentic representation of the corresponding utterance 220 in the tuple and therefore the tuple is from the seed data 210. Ideally, the classifier 250 outputs a high probability, such as 1, for authentic tuples (i.e., seed tuples) and a low probability, such as 0, for generated tuples. Early in training, the output of the classifier may be poor. For example, the probabilities output may be in a range (e.g., 0 to 1 (0 and 1 and However, as training progresses throughout a round or across multiple rounds, the classifier 250 may be improved.

[0051] At block 425, the error minimization module 260 of the training system 150 generates a training signal for training the classifier 250 based on the output of the semantic parser 114 and the classifier 250 as determined at blocks 415 and 420. In some embodiments, the data input subsystem 240 of the training system 150 provides the first portion of the seed data 210 as accessed at block 405 to the error minimization module 260 to enable the error minimization module 260 to determine the training signal.

[0052] For example, the error minimization module 260 may apply a first objective function 265 that compares the exact distribution of predictions (i.e., the exact sequence of values) to the actual distribution of predictions (i.e., the actual sequence of values) from the identifier 250 to determine the degree of discrepancy, or difference, between the two distributions. The error minimization module 260 may have access to a first portion of the seed data 210 and may therefore know which tuples received by the identifier 250 are genuine and which tuples were generated. Thus, the error minimization module 260 knows that the exact distribution assigns a value of 1 (i.e., 100 percent probability) to seed tuples and a value of 0 (i.e., 0 percent probability) to generated tuples that are not genuine. Thus, the error minimization module 260 may use the first objective function 265 to compare this exact distribution to the actual output of the identifier 250 to provide a training signal to the identifier 250. The training signal represents the error in the predictions made by the identifier 250. Based on the training signal, the training system 150 may update the classifier 250 to reduce the error between the correct distribution and the actual distribution. For example, a node of a neural network acting as the classifier 250 may be updated based on the training signal.

[0053] In some embodiments, as shown in FIG. 4, blocks 415 and 420 are followed by block 425. In that case, the error minimization module 260 may provide the training signal only after the identifier 250 has received all the tuples used to train the identifier 250 in this round of training. Thus, the identifier 250 may be updated once based on a batch including the entire first portion of the seed data 210 and the generated tuples described above. However, instead, block 425 may be performed in parallel with blocks 415 and 420 such that the identifier 250 is updated while the first portion of the seed data 210 and the generated tuples are still provided such that the training signal is sent back to the identifier 250 while the identifier 250 is evaluating the inputs.

[0054] At block 430, the training system 150 finishes training the classifier 250 in the current round of training, so the semantic parser 114 that was fixed at block 410 is unlocked as training progresses and does not need to remain fixed.

[0055] 5 is a diagram of a method for training a semantic parser as part of a round of training in a GAN 155, according to certain embodiments described herein. Specifically, training system 150 may perform this method 500, or a similar method, at block 320 of method 300 described above, to train semantic parser 114 within a single round of training. Thus, according to some embodiments, this method 500, or a similar method, may be performed once per iteration.

[0056] The method 500 depicted in FIG. 5, and other methods described herein, includes one or more processes. The method 500 may be implemented in software (e.g., as code, instructions, or programs) executed by a processor (e.g., a processor or processor core), in hardware, or in a combination thereof. The software may be stored on a non-transitory storage medium, for example, on a memory device. The method 500 is intended to be exemplary and non-limiting. Although FIG. 5 depicts various activities occurring in a particular sequence or order, this is not intended to be limiting. In an embodiment, for example, these activities may be performed in a different order, or one or more activities of the method 500 may be performed in parallel. In an embodiment, the method 500 may be performed by the training system 150.

[0057] At block 505, the data input subsystem 240 of the training system 150 selects a second portion of the seed data 210, where the second portion has not yet been used in a previous round of training. In some embodiments, this second portion of the seed data 210 is a subset of the seed tuples in the seed data 210. For example, the second portion of the seed data 210 may be a proper subset of the seed tuples in the seed data 210, such that not all of the seed tuples in the seed data 210 are included in the second portion. Additionally, in some embodiments, the second portion of the seed data 210 is different from the first portion. However, the second portion may alternatively be the same as the first portion of the seed data 210. In either case, the remainder of the seed data 210 other than the first and second portions may be used in subsequent iterations of the overall training method 300.

[0058] At block 510, the identifier 250 is frozen while the semantic parser 114 is open for updates. In other words, in some embodiments, the identifier 250 will remain static for the time being to allow the semantic parser 114 to be trained based on the current state of the identifier 250.

[0059] In some embodiments, blocks 515 and 520 are performed in parallel. At block 515, the data input subsystem 240 of the training system 150 inputs utterances 220 from the second portion of the seed data 210 selected at block 505 to the semantic parser 114, which generates logical forms 230 based on such utterances 220. Early in training, the output of the logical forms 230 of the semantic parser may be poor. For example, the generated logical forms 230 may be random (e.g., random arrangement of words and symbols). However, as training progresses throughout the round or across multiple rounds, the semantic parser 114 may be improved.

[0060] At block 520, which may be performed in parallel with block 515, the data input subsystem 240 of the training system 150 may input a stream of tuples to the identifier 250, which may perform a prediction of authenticity for each tuple in the stream. Each tuple in the stream of tuples may include an utterance 220 and a corresponding logical form 230. More specifically, each received tuple may include an utterance 220 from a respective seed tuple selected from the second portion of the seed data 210 as selected at block 505, along with a generated logical form 230 output by the semantic parser 114 based on that utterance 220 that is provided as an input to the semantic parser 114 at block 515. For each received tuple, the identifier 250 may classify the tuple based on its possible authenticity. Specifically, the identifier 250 may output a score indicative of the likelihood of authenticity. For example, the score determined by the classifier 250 for a tuple may be the probability that the logical form 230 is a genuine and therefore accurate representation of the corresponding utterance 220 in the tuple, or, in other words, that the tuple is It may be the probability that the seed tuple is from the seed data 210. In this respect, since the classifier 250 has already undergone some training as described with reference to Figure 4, while its output still requires improvement, the classifier 250 may continue to be improved during subsequent rounds of training.

[0061] At block 525, the error minimization module 260 of the training system 150 generates training signals for training the semantic parser 114 via back-propagation based on the outputs of the semantic parser 114 and the classifier 250 as determined at blocks 515 and 520. To this end, for example, the error minimization module 260 applies a second objective function 265 that determines the discrepancy between a selected distribution of predictions from the classifier 250 and an actual distribution of predictions from the classifier 250. The second objective function 265 may be, but need not be, the same as the first objective function 265 used in training the classifier 250. In some embodiments, the selected distribution of predictions is a series of predictions, since the training system 150 seeks to improve the semantic parser 114 such that it is able to trick the classifier 250 into predicting that all of the outputs of the semantic parser 114 are genuine. In some embodiments, such a distribution of predictions would indicate that the logical forms 230 generated by the semantic parser 114 are all genuine. Thus, the second objective function 265 may compare this selected distribution (e.g., of the set of predictions) with the actual distribution of the output from the identifier 250 in block 520 to determine the degree of difference between these distributions.

[0062] The error minimization module 260 may use a second objective function 265 to compare this selected distribution with the actual prediction of the classifier 250 to provide a training signal to the semantic parser 114. The training signal represents the error in the prediction made compared to the selected (i.e., desired) one. Based on the training signal, the semantic parser 114 may be updated to reduce the error between the selected distribution and the actual distribution. For example, the nodes of a neural network acting as the semantic parser 114 may be updated based on the training signal.

[0063] 5, blocks 515 and 520 are followed by block 525. In this case, the error minimization module 260 may provide a training signal only after the semantic parser 114 has received all of the utterances 220 in the second portion of the seed data 210. Thus, the semantic parser 114 is updated based on a batch that includes the entire second portion of the seed data 210. However, instead, block 525 may be performed in parallel with blocks 515 and 520 such that the training signal is sent back to the semantic parser 114 while the semantic parser 114 is evaluating the input, such that the semantic parser 114 is updated while the utterances 220 from the second portion of the seed data 210 are being provided.

[0064] Thus, as described above, the GAN 155 may be used to train the semantic parser 114 for use in the dialogue system 100. Various modifications may be made to the above-described approach, and such modifications are within the scope of this disclosure. For example, a semi-supervised GAN (SGAN) may be used as the GAN 155 to train both the supervised classifier, the unsupervised classifier, and the semantic parser 114. Even for GAN training, SGANs can be particularly useful when the seed data 210 is a relatively small set. Additionally or alternatively, the identifier 250 may incorporate information about a grammar to which the logical form 230 must conform. By referring to this grammar, the identifier 250 may more accurately determine whether the logical form 230 is authentic, which may prompt the semantic parser 114 to be improved more quickly.

[0065] 6 is a diagram of a distributed system 600 for implementing an embodiment. In the illustrated embodiment, the distributed system 600 includes one or more client computing devices 602, 604, 606, and 608 coupled to a server 612 via one or more communication networks 610. The client computing devices 602, 604, 606, and 608 may be configured to execute one or more applications.

[0066] In various embodiments, server 612 may be adapted to run one or more services or software applications that enable using backpropagation in GAN 155 to train semantic parser 114 of dialogue system 100 as described herein. For example, server 612 may run some or all aspects of training system 150 or some or all aspects of dialogue system 100.

[0067] In some embodiments, the server 612 may also provide other services or software applications, which may include non-virtual and virtual environments. In some embodiments, these services are provided as Software as a Service (SAS). The services provided by the client computing devices 602, 604, 606, and / or 608 may be provided to users of the client computing devices 602, 604, 606, and / or 608 as web-based or cloud services, such as under a Software as a Service (SaaS) model. Users operating the client computing devices 602, 604, 606, and / or 608 may then utilize one or more client applications to interact with the server 612 to utilize the services provided by these components. More specifically, for example, each of the client computing devices 602, 604, 606, and / or 608 may be an embedded device configured to execute the dialogue system 100 and further configured to communicate with the server 612 to enable the server 612 to train the semantic parser 114 of the dialogue system 100 through backpropagation in the GAN 155 as described herein.

[0068] In the configuration depicted in Figure 6, server 612 may include one or more components 618, 620, and 622 that implement the functions performed by server 612. These components may include software components, hardware components, or a combination thereof that may be executed by one or more processors. It should be understood that a variety of different system configurations are possible that may differ from distributed system 600. The embodiment depicted in Figure 6 is thus an example of a distributed system for implementing the system of the embodiments and is not intended to be limiting.

[0069] Users may use client computing devices 602, 604, 606, and / or 608 to interact with aspects of the dialogue system 100 provided by server 612 in accordance with the teachings of this disclosure. The client devices may provide an interface (e.g., a voice interface) that allows a user of the client device to interact with the client device. Although Figure 6 depicts only four client computing devices, any number of client computing devices may be supported.

[0070] Client devices may include various types of computers, such as PA devices, portable handheld devices, general purpose computers such as personal computers and laptops, workstation computers, wearable devices, gaming systems, thin clients, various messaging devices, sensors or other sensing devices, etc. The computing devices may include various types and versions of software applications and operating systems (e.g., Microsoft Windows, Apple Macintosh, UNIX or UNIX-like operating systems, Linux or Linux-like operating systems, e.g., Google Chrome™ OS), including various mobile operating systems (e.g., Microsoft Windows Mobile, iOS, Windows Phone, Android, Blackberry, Palm OS). Portable handheld devices may include mobile phones, smartphones (e.g., iPhone), tablets (e.g., iPad), personal digital assistants (PDAs), and the like. Wearable devices may include Google Glass head-mounted displays, and other devices. The gaming systems may include various handheld gaming devices, Internet-enabled gaming devices (e.g., Microsoft Xbox gaming consoles, Sony PlayStation systems, various gaming systems offered by Nintendo, etc., with or without a Kinect gesture input device), etc. The client devices may include various Internet-related applications, communication applications (e.g., email applications, short message service (SMS) applications, etc.), and the like. The network may be capable of running a variety of different applications, such as network applications, and may use a variety of communication protocols.

[0071] The network 610 may be any suitable network technology including, but not limited to, TCP / IP (transmission control protocol / Internet protocol), SNA (systems network architecture), IPX (Internet packet exchange), AppleTalk (registered trademark), and the like. The network 610 may be any type of network known to those skilled in the art that is capable of supporting data communications using any of a variety of available protocols, including without limitation, any protocol. By way of example only, the network 610 may be a local area network (LAN), an Ethernet or token ring based network, a wide-area network (WAN), the Internet, a virtual Network, virtual private network (VPN), Intranets, extranets, public switched telephone networks (PSTN), infrared networks, wireless networks (for example, The Institute of Electrical and Electronics (IEEE) 802.11 Protocol Bluetooth®, and / or any other wireless protocol , a network operating under any of the above), and / or any combination of these and / or other networks.

[0072] The servers 612 may be comprised of one or more general purpose computers, dedicated server computers (including, by way of example, PC (personal computer) servers, UNIX servers, mid-range servers, mainframe computers, rack-mounted servers, etc.), server farms, server clusters, or any other suitable configuration and / or combination. The servers 612 may include one or more virtual machines running a virtual operating system, or other computing architectures involving virtualization, e.g., one or more flexible pools of logical storage devices that may be virtualized to maintain virtual storage devices for the servers. In various embodiments, the servers 612 may be adapted to run one or more services or software applications that provide the functionality described in the preceding disclosure.

[0073] The computing systems in server 612 may run one or more operating systems, including any of those mentioned above, as well as any commercially available server operating system. Server 612 may also run any of a variety of additional server applications and / or mid-tier applications, including hypertext transport protocol (HTTP) servers, file transfer protocol (FTP) servers, common gateway interface (CGI) servers, JAVA servers, database servers, and the like. Exemplary database servers are provided by Oracle, Microsoft, Sybase, IBM (International Business Machines), and the like. The present disclosure includes, but is not limited to, those commercially available from manufacturers such as Sony Computer Entertainment Inc.

[0074] In some implementations, server 612 may include one or more applications for analyzing and consolidating data feeds and / or event updates received from users of client computing devices 602, 604, 606, and 608. By way of example, the data feeds and / or event updates may include, but are not limited to, Twitter feeds, Facebook updates, or real-time updates received from one or more third party information sources and continuous data streams that may include real-time events related to sensor data applications, financial tickers, network performance measurement tools (e.g., network monitoring and traffic management applications), clickstream analysis tools, automobile traffic monitoring, etc. Server 612 may also include one or more applications for displaying the data feeds and / or real-time events via one or more display devices of client computing devices 602, 604, 606, and 608.

[0075] The distributed system 600 may also include one or more data repositories 614, 616. In an embodiment, these data repositories may be used to store data and other information. For example, one or more of the data repositories 614, 616 may be used to store seed data 210 or other data required to train the semantic parser 114 of the dialogue system 100 by backpropagation in the GAN 155 as described herein. The data repositories 614, 616 may reside in various locations. For example, the data repository used by the server 612 may be local to the server 612 or may be remote from the server 612 and in communication with the server 612 via a network-based or dedicated connection. The data repositories 614, 616 may be of different types. In an embodiment, the data repository used by the server 612 may be a database, such as a relational database, such as databases provided by Oracle Corporation and other vendors. One or more of these databases may be adapted to allow data to be stored, updated and retrieved from the database in response to SQL-formatted commands.

[0076] In an embodiment, one or more of the data repositories 614, 616 may also be used by an application to store application data. The data repositories used by an application may be of different types, such as, for example, a key-value store repository supported by a file system, an object store repository, or a general storage repository.

[0077] In an embodiment, all or a portion of training the semantic parser 114 of the dialogue system 100 in the GAN 155 by backpropagation as described herein may be provided as a service via a cloud environment. FIG. 7 is a block diagram of a cloud-based system environment in which training the semantic parser 114 as described herein may be provided at least in part as a cloud service according to an embodiment described herein. In the embodiment depicted in FIG. 7, a cloud infrastructure system 702 may provide one or more cloud services that may be requested by users using one or more client computing devices 704, 706, and 708. The cloud infrastructure system 702 may include one or more computers and / or servers, which may include those described above for the server 612. The computers in the cloud infrastructure system 702 may be organized as general purpose computers, dedicated server computers, server farms, server clusters, or any other suitable configuration and / or combination.

[0078] The network 710 may facilitate communication and exchange of data between the client computing devices 704, 706, and 708 and the cloud infrastructure system 702. The network 710 may include one or more networks. The networks may be of the same type or of different types. The network 710 may support one or more communication protocols, including wired and / or wireless protocols, to facilitate communication.

[0079] The embodiment depicted in Figure 7 is merely one example of a cloud infrastructure system and is not intended to be limiting. It should be understood that in some other embodiments, cloud infrastructure system 702 may have more or fewer components than those depicted in Figure 7, may combine two or more components, or may have components in a different configuration or arrangement. For example, while Figure 7 depicts three client computing devices, in alternative embodiments, any number of client computing devices may be supported.

[0080] The term cloud services is generally used to refer to services that are made available to users on demand by a service provider's system (e.g., cloud infrastructure system 702) over a communications network such as the Internet. Typically, in a public cloud environment, the servers and systems that make up the cloud service provider's system are distinct from the customer's own on-premise servers and systems. The cloud service provider's systems are managed by the cloud service provider. Customers can thus utilize cloud services provided by the cloud service provider without having to purchase separate licenses, support, or hardware and software resources for the services. For example, the cloud service provider's system may host an application, and a user may order and use the application on demand over the Internet without the user having to purchase infrastructure resources to run the application. Cloud services are designed to provide easy and scalable access to applications, resources, and services. Several providers offer cloud services. For example, several cloud services, such as middleware services, database services, Java cloud services, etc., are offered by Oracle Corporation of Redwood Shores, California.

[0081] In one embodiment, the cloud infrastructure system 702 is a hybrid Service models, including Software as a Service (SaaS) models, Platform as a Service (PaaS) models, and infrastructure Infrastructure as a Service (IaaS) Model A cloud infrastructure system 702 may provide one or more cloud services using different models, such as under a single cloud service model, a single cloud service model, or a single cloud infrastructure model. Cloud infrastructure system 702 may include a suite of applications, middleware, databases, and other resources that enable the provision of various cloud services.

[0082] The SaaS model allows an application or software to be delivered to a customer as a service over a communications network such as the Internet without the customer having to purchase hardware or software for the base application. For example, the SaaS model may be used to provide customers with access to on-demand applications hosted by the cloud infrastructure system 702. Examples of SaaS services offered by Oracle Corporation include, without limitation, various services for human capital / capital management, customer relationship management (CRM), enterprise resource planning (ERP), supply chain management (SCM), enterprise performance management (EPM), analytical services, social applications, and the like.

[0083] The IaaS model is commonly used to provide infrastructure resources (e.g., servers, storage, hardware and networking resources) as a cloud service to customers to provide elastic computing and storage capacity. A variety of IaaS services are offered by Oracle Corporation.

[0084] The PaaS model is generally used to provide platform and environment resources that enable customers to develop, run, and manage applications and services as a service, without the customer having to acquire, build, or maintain such resources. Examples of PaaS services provided by Oracle Corporation include, but are not limited to, Oracle Java Cloud Service (JCS), Oracle Database Cloud Service (DBCS), data management cloud services, various application development solution services, etc.

[0085] Cloud services are generally provided on an on-demand, self-service basis, on a subscription basis, and in a resiliently scalable, reliable, highly available, and secure manner. For example, a customer may order one or more services provided by cloud infrastructure system 702 via a subscription order. Cloud infrastructure system 702 then performs processing to provide the services requested in the customer's subscription order. For example, a customer may subscribe to information services or other services provided by dialogue system 100 in a conversational format. Cloud infrastructure system 702 may be configured to provide one or more cloud services.

[0086] Cloud infrastructure system 702 may provide cloud services through different deployment models. In a public cloud model, cloud infrastructure system 702 may be owned by a third party cloud service provider and cloud services are offered to any public customer, where the customer may be an individual or a business. In some other embodiments, under a private cloud model, cloud infrastructure system 702 operates within an organization (e.g., within a business organization). The cloud infrastructure system 702 may be shared among several organizations in a community, and services may be provided for customers within the organization. For example, the customers may be various departments of a company, such as human resources, payroll, etc., or may be individuals within the company. In some other embodiments, under a community cloud model, the cloud infrastructure system 702 and the services provided may be shared by several organizations in an associated community. Various other models, such as hybrids of the above models, may also be used.

[0087] Client computing devices 704, 706, and 708 may be of different types (such as client computing devices 602, 604, 606, and 608 depicted in FIG. 6) and may be capable of operating one or more client applications. Users may use the client computing devices to interact with cloud infrastructure system 702, such as to request services provided by cloud infrastructure system 702. Attackers may use client devices to send malicious requests.

[0088] In some embodiments, the processing performed by cloud infrastructure system 702 may involve big data analysis. This analysis may involve using, analyzing, and manipulating large data sets to detect and visualize various trends, behaviors, relationships, etc. in the data. This analysis may be performed by one or more processors, possibly in parallel, processing the data, performing simulations using the data, etc. For example, big data analysis may be performed by cloud infrastructure system 702 to provide training of semantic parser 114 by backpropagation in GAN 155 as described herein. The data used for this analysis may include structured data (e.g., data stored in a database or structured according to a structured model) and / or unstructured data (e.g., data blobs (binary large objects)). stomach.

[0089] 7, cloud infrastructure system 702 may include infrastructure resources 730 utilized to facilitate the provision of various cloud services offered by cloud infrastructure system 702. Infrastructure resources 730 may include, for example, processing resources, storage or memory resources, networking resources, etc.

[0090] In one embodiment, infrastructure resources 730 may be bundled into sets of resources or resource modules (also referred to as “pods”) to facilitate efficient provisioning of these resources to support various cloud services offered by cloud infrastructure system 702 to different customers. Each resource module or pod may include a pre-integrated and optimized combination of one or more types of resources. In one embodiment, different pods may be pre-provisioned for different types of cloud services. For example, a first set of pods may be provisioned for database services, a second set of pods, which may include a different combination of resources than the pods in the first set of pods, may be provisioned for Java services, and so on. For some services, the resources allocated to provision the service may be shared between the services.

[0091] The cloud infrastructure system 702 itself is a cloud infrastructure There may be services 732 used internally that are shared by different components of the system 702 and facilitate provisioning of services by the cloud infrastructure system 702. These internal shared services may include, without any limitation, security and identity services, integration services, enterprise repository services, enterprise manager services, virus scanning and whitelist services, high availability, backup and recovery services, services to enable cloud support, email services, notification services, file transfer services, etc.

[0092] Cloud infrastructure system 702 may include multiple subsystems. These subsystems may be implemented in software, or hardware, or a combination thereof. As depicted in FIG. 7 , the subsystems may include a user interface subsystem 712 that allows a user or customer of cloud infrastructure system 702 to interact with cloud infrastructure system 702. User interface subsystem 712 may include a variety of different interfaces, such as a web interface 714, an online store interface 716 where cloud services offered by cloud infrastructure system 702 are advertised and available for purchase by consumers, and other interfaces 718. For example, a customer may use a client device to request one or more services offered by cloud infrastructure system 702 using one or more of interfaces 714, 716, and 718 (service request 734). For example, a customer may access an online store, browse cloud services offered by cloud infrastructure system 702, and place a subscription order for one or more services offered by cloud infrastructure system 702 to which the customer would like to subscribe. The service request may include information identifying the customer and the one or more services to which the customer would like to subscribe.

[0093] 7, cloud infrastructure system 702 may include an order management subsystem (OMS) 720 configured to process new orders. As part of this processing, OMS 720 may be configured to create an account for the customer if not already created, receive billing and / or accounting information from the customer that is used to bill the customer for providing the requested services to the customer, verify the customer information, and upon verification, book an order for the customer, and orchestrate various workflows to prepare the order for provisioning.

[0094] Once properly verified, OMS 720 may then invoke an order provisioning subsystem (OPS) 724 configured to provision resources for the order, including processing, memory, and networking resources. Provisioning may include allocating resources for the order and configuring the resources to facilitate the service requested by the customer order. The manner in which resources are provisioned for the order and the type of resources provisioned may depend on the type of cloud service ordered by the customer. For example, according to one workflow, OPS 724 may be configured to determine the particular cloud service being requested and identify the number of pods that may have been pre-configured for that particular cloud service. The number of pods allocated for the order may depend on the size / amount / level / scope of the service requested. For example, the number of pods allocated may be determined based on the number of users supported by the service, the duration for which the service is being requested, etc. The allocated pods are then customized for the particular requesting customer to provide the requested service. The number of pixels may be customized.

[0095] The cloud infrastructure system 702 may send a response or notification 744 to the requesting customer to indicate when the requested service will be ready for use. In some cases, information (e.g., a link) may be sent to the customer that enables the customer to begin using and taking advantage of the benefits of the requested service.

[0096] Cloud infrastructure system 702 may provide services to multiple customers. For each customer, cloud infrastructure system 702 is responsible for managing information related to one or more subscription orders received from the customer, maintaining customer data related to the orders, and providing the requested services to the customer. Cloud infrastructure system 702 may also collect usage statistics regarding the customer's use of the subscribed services. For example, statistics may be collected about the amount of storage used, the amount of data transferred, the number of users, the amount of system uptime and system downtime, etc. This usage information may be used to bill the customer. Billing may occur on a monthly basis, for example.

[0097] Cloud infrastructure system 702 may provide services to multiple customers in parallel. Cloud infrastructure system 702 may store information, possibly including sensitive information, for these customers. In one embodiment, cloud infrastructure system 702 includes an identity management subsystem (IMS) 728 configured to manage customer information and provide separation of managed information such that information related to one customer cannot be accessed by another customer. The IMS 728 may be configured to provide various security-related services, such as identity services, e.g., information access management, authentication and authorization services, services for managing customer identities and roles and associated capabilities, etc.

[0098] FIG. 8 is a block diagram of an exemplary computer system 800 that may be used to implement an embodiment. For example, in some embodiments, the computer system 800 may be used to implement any of the systems, subsystems, and components described herein. For example, multiple host machines may provide training and implement the semantic parser 114 of the dialogue system 100 in the GAN 155 as described herein. A computer system such as the computer system 800 may be used as a host machine. As shown in FIG. 8, the computer system 800 includes various subsystems, including a processing subsystem 804 that communicates with many other subsystems via a bus subsystem 802. These other subsystems may include a processing accelerator 806, an I / O subsystem 808, a storage subsystem 818, and a communication subsystem 824. The storage subsystem 818 may include a non-transitory computer-readable storage medium, including a storage medium 822 and a system memory 810.

[0099] Bus subsystem 802 provides a mechanism for allowing the various components and subsystems of computer system 800 to communicate with each other as intended. Although bus subsystem 802 is shown generally as a single bus, alternative embodiments of the bus subsystem may utilize multiple buses. Bus subsystem 802 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of a variety of bus architectures. For example, such architectures include the Industry Standard Architecture (ISA) bus, the Micro Channel architecture, and the like. (Micro Channel Architecture:MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) low culbus, and Peripheral Component Interconnect (Peripheral Component Interconnect: The PCI bus may be implemented as a mezzanine bus or the like, which is implemented according to the IEEE P1386.1 standard.

[0100] The processing subsystem 804 controls the operation of the computer system 800 and may include one or more processors, application specific integrated circuits (ASIs), and / or C), or field programmable gate array :FPGA). These processors may include single-core or multi-core processors. The processing resources of the computer system 800 may be organized into one or more processing units 832, 834, etc. A processing unit may include one or more processors, one or more cores from the same or different processors, a combination of cores and processors, or other combinations of cores and processors. In some embodiments, the processing subsystem 804 may include one or more special purpose co-processors, such as a graphics processor, a digital signal processor (DSP), etc. In some embodiments, some or all of the processing units of the processing subsystem 804 may be implemented using customized circuitry, such as an application specific integrated circuit (ASIC) or a field programmable gate array (FPGA).

[0101] In some embodiments, the processors in the processing subsystem 804 can execute instructions stored in the system memory 810 or on the computer readable storage medium 822. In various embodiments, the processors can execute various program or code instructions and can maintain multiple programs or processes running simultaneously. At any given time, some or all of the program code to be executed may reside in the system memory 810 and / or on the computer readable storage medium 822, possibly on one or more storage devices. Through suitable programming, the processing subsystem 804 can provide the various functionality described above. In instances where the computer system 800 is running one or more virtual machines, one or more processors may be assigned to each virtual machine.

[0102] In one embodiment, a processing acceleration unit 806 may optionally be provided to perform customized processing to accelerate the overall processing performed by computer system 800 or to offload some of the processing performed by processing subsystem 804.

[0103] I / O subsystem 808 may include devices and mechanisms for inputting information into computer system 800 and / or outputting information from or through computer system 800. In general, use of the term input device is intended to include all possible types of devices and mechanisms for inputting information into computer system 800. User interface input devices may include keyboards, pointing devices such as mice or trackballs, touch pads or touch screens integrated into displays, scroll wheels, click wheels, dials, buttons, switches, keypads, voice input devices with voice command recognition systems, microphones, and other types of input devices. User interface input devices also include motion sensing and / or gesture recognition devices, such as Microsoft Kinect® motion sensors, Microsoft Xbox® 360 game controllers, that allow a user to control and interact with the input device, and devices that provide an interface for receiving input using gestures and spoken commands. The user interface input device may also include an eye gesture recognition device, such as a Google Glass® blink detector, that detects eye activity from the user (e.g., "blinking" during picture taking and / or menu selection) and translates the eye gesture as input to the input device (e.g., Google Glass®). Additionally, the user interface input device may include a voice recognition sensing device that allows the user to interact with a voice recognition system (e.g., the Siri® navigator) through voice commands.

[0104] Other examples of user interface input devices include, without limitation, three-dimensional (3D) mice, joysticks or pointing sticks, game pads and graphic tablets, as well as audio / visual devices such as speakers, digital cameras, digital video cameras, portable media players, webcams, image scanners, fingerprint scanners, barcode readers 3D scanners, 3D printers, laser range finders, and eye-tracking devices. In addition, user interface input devices may include medical imaging input devices, such as, for example, computed tomography, magnetic resonance imaging, positron emission tomography, medical ultrasound, etc. User interface input devices may also include audio input devices, such as, for example, MIDI keyboards, digital musical instruments, etc.

[0105] In general, use of the term output device is intended to include all possible types of devices and mechanisms for outputting information from computer system 800 to a user or to another computer. User interface output devices may include non-visual displays such as a display subsystem, indicator lights, or audio output devices. The display subsystem may be a flat panel device such as one using a cathode ray tube (CRT), liquid crystal display (LCD) or plasma display, a projection device, a touch screen, etc. For example, user interface output devices may include, without limitation, a variety of display devices that visually convey text, graphics, and audio / visual information, such as monitors, printers, speakers, headphones, automobile navigation systems, plotters, audio output devices, and modems.

[0106] The storage subsystem 818 provides a repository or data store for storing information and data used by the computer system 800. The storage subsystem 818 provides a tangible, non-transitory computer-readable storage medium for storing data structures and basic programming that provide the functionality of some embodiments. The storage subsystem 818 may store software (e.g., programs, code modules, instructions) that, when executed by the processing subsystem 804, provide the functionality described above. The software may be executed by one or more processing units of the processing subsystem 804. The storage subsystem 818 may also provide a repository for storing data used in accordance with the teachings of this disclosure.

[0107] The storage subsystem 818 may include one or more non-transitory memory devices, including volatile and non-volatile memory devices. As shown in FIG. 8, the storage subsystem 818 includes a system memory 810 and a computer-readable storage medium 822. The system memory 810 may include multiple memories, including a volatile main random access memory (RAM) for storage of instructions and data during program execution, and a non-volatile read only memory (ROM) or flash memory in which fixed instructions are stored. In some implementations, a basic input / output system (BIOS), containing the basic routines that help to transfer information between elements within the computer system 800, such as during start-up, is typically stored in a ROM. RAM typically contains data and / or program modules currently being operated and executed by the processing subsystem 804. In some implementations, the system memory 810 may include a number of different types of memory, such as static random access memory (SRAM), dynamic random access memory (DRAM), etc. .

[0108] By way of example, and not by way of limitation, as depicted in FIG. 8, system memory 810 may include memory for various applications, such as a web browser, a mid-tier application, a relational database management system (RDBMS), etc. The system may load running application programs 812, program data 814, and operating system 816, which may include applications. By way of example, operating system 816 may include various versions of Microsoft Windows, Apple Macintosh, and / or Linux operating systems, various commercially available UNIX or UNIX-like operating systems (including, without limitation, various GNU / Linux operating systems, Google Chrome OS, etc.), and / or mobile operating systems such as iOS, Windows Phone, Android OS, BlackBerry OS, Palm OS operating systems, etc.

[0109] In one embodiment, software instructions or code implementing the training of the semantic parser 114 of the dialogue system 100 in the GAN 155 as described herein may be executed in the system memory 810.

[0110] The computer readable storage medium 822 may store programming and data structures that provide functionality of some embodiments. The computer readable storage medium 822 may provide storage of computer readable instructions, data structures, program modules, and other data for the computer system 800. Software (programs, code modules, instructions) that, when executed by the processing subsystem 804, provide the functionality described above may be stored in the storage subsystem 818. By way of example, the computer readable storage medium 822 may include non-volatile memory such as hard disk drives, magnetic disk drives, optical disk drives such as CD ROMs, DVDs, Blu-Ray® disks, or other optical media. The computer readable storage medium 822 may include, but is not limited to, Zip® drives, flash memory cards, universal serial bus (USB) flash drives, secure digital (SD) cards, DVD disks, digital video tapes, and the like. The computer readable storage medium 822 may also include flash memory-based solid-state drives (SSDs), enterprise flash drives, and the like. These may include SSDs based on non-volatile memory, such as solid state ROM, solid state RAM, dynamic RAM, static RAM, DRAM-based SSDs, magnetoresistive RAM (MRAM) SSDs, and hybrid SSDs that use a combination of DRAM-based SSDs and flash memory-based SSDs.

[0111] In an embodiment, the storage subsystem 818 may also include a computer readable storage medium reader 820, which may be further connected to a computer readable storage medium 822. The reader 820 may be configured to receive and read data from a memory device, such as a disk, a flash drive, or the like.

[0112] In some embodiments, computer system 800 may support virtualization techniques, including but not limited to virtualization of processing and memory resources. For example, computer system 800 may provide support for running one or more virtual machines. In some embodiments, computer system 800 may execute a program, such as a hypervisor, that facilitates configuring and managing virtual machines. Each virtual machine may be assigned memory, computational (e.g., processors, cores), I / O, and networking resources. Each virtual machine generally runs independently of other virtual machines. A virtual machine typically runs its own operating system, which may be the same or different from the operating systems executed by other virtual machines executed by computer system 800. Thus, multiple operating systems may be executed by computer system 800, possibly simultaneously.

[0113] The communications subsystem 824 provides an interface to other computer systems and networks. The communications subsystem 824 serves as an interface for receiving data from and transmitting data to systems other than the computer system 800. For example, the communications subsystem 824 may enable the computer system 800 to establish a communications channel over the Internet to one or more client devices to receive and transmit information to the client devices.

[0114] The communications subsystem 824 may support both wired and / or wireless communication protocols. For example, in one embodiment, the communications subsystem 824 may include a radio frequency (RF) transceiver controller for accessing wireless voice and / or data networks (e.g., using cellular technology, advanced data network technologies (e.g., 3G, 4G, or EDGE (enhanced data rates for global evolution), WiFi (IEEE 802.XX family of standards)), or other mobile communications technologies, or any combination thereof). Component, Global Positioning System (GPS) receiver component In some embodiments, the communications subsystem 824 can provide a wired network connection (e.g., Ethernet) in addition to or instead of a wireless interface.

[0115] The communications subsystem 824 can receive and transmit data in a variety of formats. For example, in some embodiments, in addition to other formats, the communications subsystem 824 may receive incoming communications in the form of structured and / or unstructured data feeds 826, event streams 828, event updates 830, and the like. For example, the communications subsystem 824 may receive web feeds, such as Twitter feeds, Facebook updates, Rich Site Summary (RSS) feeds, and the like. The social network may be configured to receive (or transmit) data feeds 826 in real time from users of the social network and / or other communication services, such as real-time updates from one or more third-party sources.

[0116] In an embodiment, the communications subsystem 824 may be configured to receive data in the form of a continuous data stream that may be continuous or infinite in nature with no apparent end, which may include an event stream 828 of real-time events and / or event updates 830. Examples of applications that generate continuous data may include, for example, sensor data applications, financial tickers, network performance measurement tools (e.g., network monitoring and traffic management applications), clickstream analysis tools, automobile traffic monitoring, etc.

[0117] Communications subsystem 824 may also be configured to communicate data from computer system 800 to other computer systems or networks. Data may be communicated in a variety of different formats, such as structured and / or unstructured data feeds 826, event streams 828, event updates 830, etc., to one or more databases that may be in communication with one or more streaming data source computers coupled to computer system 800.

[0118] Computer system 800 may be one of a variety of types, including a handheld mobile device (e.g., an iPhone® mobile phone, an iPad® computing tablet, a PDA), a wearable device (e.g., a Google Glass® head mounted display), a personal computer, a workstation, a mainframe, a kiosk, a server rack, or any other data processing system. Due to the ever-changing nature of computers and networks, the description of computer system 800 depicted in FIG. 8 is intended as merely illustrative. Many other configurations are possible having more or fewer components than the system depicted in FIG. 8. Based on the disclosure and teachings provided herein, one of ordinary skill in the art will appreciate other ways and / or methods for implementing various embodiments.

[0119] While specific embodiments have been described, various modifications, variations, alternative constructions, and equivalents are possible. The embodiments are not limited to operating in one particular data processing environment, but operate freely in multiple data processing environments. In addition, while certain embodiments have been described using a particular sequence of transactions and steps, it should be apparent to one skilled in the art that this is not intended to be limiting. Although some flow charts describe operations as a sequential process, many of the operations may be performed in parallel or simultaneously. In addition, the order of operations may be rearranged. A process may have additional steps not included in the figures. Various features and aspects of the above-described embodiments may be used individually or together.

[0120] Also, while certain embodiments have been described using a particular combination of hardware and software, it should be recognized that other combinations of hardware and software are possible. Certain embodiments may be implemented exclusively in hardware, exclusively in software, or using a combination thereof. The various processes described herein may be implemented on the same processor, or on different processors in any combination.

[0121] When a device, system, component, or module is described as being configured to perform a certain operation or function, such configuration may be accomplished, for example, by designing an electronic circuit to perform the operation, by programming a programmable electronic circuit (such as a microprocessor) to perform the operation, by executing, for example, computer instructions or code, or a processor or core programmed to execute code or instructions stored on a non-transitory storage medium, or any combination thereof. Processes can communicate using a variety of techniques, including but not limited to conventional techniques for inter-process communication, and different pairs of processes may use different techniques, or the same pair of processes may use different techniques at different times.

[0122] In this disclosure, specific details are given to provide a thorough understanding of the embodiments. However, the embodiments may be practiced without these specific details. For example, well-known circuits, processes, algorithms, structures, and techniques have been shown without unnecessary detail so as not to obscure the embodiments. This description provides exemplary embodiments. The above description of the embodiments is merely illustrative and is not intended to limit the scope, applicability, or configuration of other embodiments. Rather, the foregoing description of the embodiments will provide those skilled in the art with an enabling description for implementing various embodiments. Various changes may be made in the function and arrangement of elements.

[0123] The specification and drawings are therefore to be regarded in an illustrative rather than a restrictive sense. It will be apparent, however, that additions, deletions, and other modifications and changes may be made thereto without departing from the broader spirit and scope as set forth in the claims. Thus, although specific embodiments have been described, these are not intended to be limiting. Various modifications and equivalents are within the scope of the claims. The modifications and / or changes include all relevant combinations of the disclosed features.

[0124] According to one embodiment, the disclosed features are outlined as follows. Example 1: A method for training a semantic parser for a dialogue system is provided, the method including accessing seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance, the method further including training the semantic parser and a classifier in a generative adversarial network (GAN), where the semantic parser learns to map utterances to logical forms and the classifier learns to recognize genuine logical forms, the semantic parser is usable as part of the dialogue system.

[0125] Example 2: The method of Example 1, wherein the step of training a semantic parser and a classifier in the GAN includes inputting random utterances to the semantic parser, the semantic parser determining a generated logical form based on the random utterance, and inputting tuples to the classifier including the generated tuples and a subset of seed tuples selected from the seed data, wherein the generated tuples include the random utterance and the generated logical form, and the step of training a semantic parser and a classifier in the GAN further includes the classifier determining a predicted authenticity of the tuples.

[0126] Example 3: The method described in Example 2, wherein the step of training the semantic parser and classifier in the GAN further includes a step of comparing the predicted authenticity with the actual authenticity of the tuple, and a step of updating the classifier based on the step of comparing the predicted authenticity with the actual authenticity of the tuple.

[0127] Example 4: The method of example 3, wherein the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data.

[0128] Example 5: The method of Example 3, wherein the step of training the semantic parser and the classifier in the GAN further includes the step of inputting a seed utterance to the semantic parser, the seed utterance being selected from a seed tuple, and the step of training the semantic parser and the classifier in the GAN further includes the step of the semantic parser determining an additional generated logical form based on the seed utterance, the step of inputting the additional generated tuple including the seed utterance from the seed data and the additional generated logical form to the classifier, and the step of the classifier determining an additional predicted authenticity of the additional generated tuple.

[0129] Example 6: The step of training the semantic parser and the classifier in the GAN further includes the steps of comparing the additional predicted authenticity with the selected authenticity of the additional generated tuples, and updating the semantic parser based on the step of comparing the additional predicted authenticity with the selected authenticity of the additional generated tuples. How to.

[0130] Example 7: The method of example 6, wherein the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

[0131] Example 8: The method of Example 1, wherein the step of training a semantic parser and a classifier in the GAN includes the steps of training the classifier using a first portion of seed data based on output from the semantic parser while the semantic parser is static to determine an updated classifier, and training the semantic parser using a second portion of the seed data based on output from the updated classifier while the updated classifier is static to determine an updated semantic parser.

[0132] Example 9: The method of example 8, wherein the semantic parser is a sequence-to-sequence neural network and the classifier is a binary classifier.

[0133] Example 10: The method of Example 1, further comprising the steps of: a dialogue system receiving a voice input; applying an automatic speech recognition subsystem to the voice input to determine an utterance; applying a semantic parser to the utterance to determine a logical form; analyzing the logical form to generate a response text in response to the voice input; and applying a text-to-speech subsystem to the response text to determine an acoustic output for the dialogue system.

[0134] Example 11: A training system including a data input subsystem configured to access seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to each seed utterance, the training system further including a generative adversarial network (GAN), the GAN including a classifier configured to learn to recognize authentic logical forms, a semantic parser configured to learn to map utterances to the logical forms, and an error minimization module configured to train the semantic parser based on an output from the classifier that classifies authenticity of an output of the semantic parser, the semantic parser being usable as part of a dialogue system.

[0135] Example 12: The training system of Example 11, wherein the data input subsystem is configured to input random utterances to the semantic parser, the semantic parser is configured to generate logical forms based on the random utterances, the data input subsystem is configured to input tuples to the classifier including the generated tuples and a subset of seed tuples selected from the seed data, the generated tuples including the random utterances and the generated logical forms, and the classifier is configured to predict the authenticity of the tuples.

[0136] Example 13: The training system of Example 12, wherein the error minimization module is further configured to compare the predicted authenticity with an actual authenticity of the tuple and update the classifier based on comparing the predicted authenticity with the actual authenticity of the tuple.

[0137] Example 14: The method of example 13, wherein the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data.

[0138] Example 15: The data input subsystem is further configured to input a seed utterance to the semantic parser, the seed utterance being selected from the seed tuple, the semantic parser being configured to determine an additional generated logical form based on the seed utterance, and the data input subsystem is further configured to input the seed utterance and the additional generated logical form from the seed data. 14. The training system of example 13, further configured to input additional generated tuples, including the additional generated tuples, to the classifier, wherein the classifier is configured to determine additional predicted authenticities of the additional generated tuples.

[0139] Example 16: The training system of Example 15, wherein the error minimization module is further configured to compare the additional predicted authenticity with the selected authenticity of the additional generated tuple and update the semantic parser based on comparing the additional predicted authenticity with the selected authenticity of the additional generated tuple.

[0140] Example 17: The training system of example 16, wherein the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

[0141] Example 18: The training system of Example 11, wherein the semantic parser is a sequence-to-sequence neural network and the discriminator is a binary classifier.

[0142] Example 19: A computer program product for training a semantic parser of a dialogue system, the computer program product including a computer-readable storage medium having program instructions embodied thereon, the program instructions being executable by one or more processors to cause the one or more processors to perform a method, the method including accessing seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance, the method further including training a generative adversarial network (GAN) classifier based on the semantic parser of the GAN, the training the classifier including inputting a random utterance to the semantic parser, the semantic parser determining a generated logical form based on the random utterance, and inputting tuples to the classifier including the generated tuples and a subset of the seed tuples selected from the seed data, the generated tuples including the random utterance and the generated logical form, the training the classifier further including the classifier determining a generated logical form based on the random utterance. , determining a predicted authenticity of the tuple, comparing the predicted authenticity with an actual authenticity of the tuple, and updating the classifier based on the comparing the predicted authenticity with the actual authenticity of the tuple, the method further comprising: training a semantic parser of the GAN based on the classifier of the GAN, the training the semantic parser comprising inputting a seed utterance to the semantic parser, the seed utterance being selected from the seed tuple, the training the semantic parser further comprising: the semantic parser determining an additional generated logical form based on the seed utterance; inputting the additional generated tuple including the seed utterance from the seed data and the additional generated logical form to the classifier; the classifier determining an additional predicted authenticity of the additional generated tuple; comparing the additional predicted authenticity with the selected authenticity of the additional generated tuple; and updating the semantic parser based on the comparing the additional predicted authenticity with the selected authenticity of the additional generated tuple,A semantic parser is a computer program product that can be used in a dialogue system to map utterances to logical forms.

[0143] Example 20: The computer program product of Example 19, wherein the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data, and the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

Claims

1. 1. A method for training a semantic parser of a dialogue system, the method comprising: accessing seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance, the method further comprising: training a semantic parser and a classifier in a generative adversarial network (GAN), where the semantic parser learns to map utterances to logical forms and the classifier learns to recognize bona fide logical forms; The method, wherein the semantic parser is usable as part of a dialogue system.

2. The step of training the semantic parser and the classifier in the GAN comprises: inputting a random utterance into the semantic parser; the semantic parser determining a generated logical form based on the random utterance; and inputting tuples to the classifier, the tuples including the generated tuples and a subset of seed tuples selected from the seed data, the generated tuples including the random utterances and the generated logical forms, and training the semantic parser and the classifier in the GAN further comprises: The method of claim 1 , wherein the identifier comprises determining a predicted authenticity of the tuple.

3. Training the semantic parser and the classifier in the GAN further comprises: comparing the predicted authenticity with the actual authenticity of the tuple; and updating the classifier based on comparing the predicted authenticity with the actual authenticity of the tuple.

4. The method of claim 3 , wherein the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data.

5. Training the semantic parser and the classifier in the GAN further comprises: The step of training the semantic parser and the classifier in the GAN further comprises the step of inputting a seed utterance to the semantic parser, the seed utterance being selected from the seed tuple, the semantic parser determining additional generated logical forms based on the seed utterance; inputting an additional generated tuple including the seed utterance from the seed data and the additional generated logical form into the classifier; The method of claim 3 or claim 4, further comprising: the identifier determining additional predicted authenticity of the additional generated tuples.

6. Training the semantic parser and the classifier in the GAN further comprises: comparing said additional predicted authenticity with a selected authenticity of said additional generated tuple; Comparing the additional predicted authenticity with the selected authenticity of the additional generated tuple. and updating the semantic parser based on the comparing step.

7. The method of claim 6 , wherein the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

8. The step of training the semantic parser and the classifier in the GAN comprises: training the classifier using a first portion of the seed data based on output from the semantic parser while the semantic parser is static to determine an updated classifier; and training the semantic parser using a second portion of the seed data based on output from the updated classifier while the updated classifier is static to determine an updated semantic parser.

9. The method of claim 8 , wherein the semantic parser is a sequence-to-sequence neural network and the classifier is a binary classifier.

10. receiving a speech input from the dialogue system; applying an automatic speech recognition subsystem to the speech input to determine an utterance; applying the semantic parser to the utterance to determine a logical form; analyzing the logical form to generate response text in response to the speech input; The method according to any one of claims 1 to 9, further comprising the step of: applying a text-to-speech subsystem to the response text to determine an acoustic output for the dialogue system.

11. 1. A training system comprising: a data input subsystem configured to access seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance, the training system further comprising: A generative adversarial network (GAN), the GAN comprising: a classifier configured to learn to recognize genuine logical forms; a semantic parser configured to learn to map utterances to logical forms; and an error minimization module configured to train the semantic parser based on an output from the classifier that classifies an authenticity of an output of the semantic parser. The semantic parser is usable as part of a dialogue system.

12. the data input subsystem configured to input random utterances to the semantic parser; the semantic parser is configured to generate logical forms based on the random utterance; The data input subsystem selects the generated tuples and the seed data. a subset of seed tuples to the classifier, the generated tuples including the random utterance and the generated logical form; The training system of claim 11 , wherein the classifier is configured to predict authenticity of the tuples.

13. The error minimization module further comprises: Comparing the predicted authenticity with the actual authenticity of the tuple; The training system of claim 12 , configured to update the classifier based on comparing the predicted authenticity with the actual authenticity of the tuples.

14. The method of claim 13 , wherein the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data.

15. the data input subsystem is further configured to input a seed utterance to the semantic parser, the seed utterance being selected from the seed tuple; the semantic parser is configured to determine additional generated logical forms based on the seed utterance; the data input subsystem is configured to input an additional generated tuple including the seed utterance from the seed data and the additional generated logical form to the classifier; The training system of claim 13 or claim 14, wherein the classifier is configured to determine additional predicted authenticity of the additional generated tuples.

16. The error minimization module further comprises: comparing said additional predicted authenticity with the selected authenticity of said additional generated tuple; 16. The training system of claim 15, configured to update the semantic parser based on comparing the additional predicted authenticity with the selected authenticity of the additional generated tuples.

17. 17. The training system of claim 16, wherein the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

18. The training system of any one of claims 11 to 17, wherein the semantic parser is a sequence-to-sequence neural network and the classifier is a binary classifier.

19. 1. A computer program product for training a semantic parser of a dialogue system, the computer program product including a computer-readable storage medium having program instructions embodied thereon, the program instructions being executable by one or more processors to cause the one or more processors to perform a method, the method comprising: accessing seed data including seed tuples, each seed tuple of the seed data including a respective seed utterance and a respective seed logical form corresponding to the respective seed utterance, the method further comprising: Training a generative adversarial network (GAN) classifier based on the GAN semantic parser, the step of training the classifier comprising: inputting a random utterance into the semantic parser; the semantic parser determining a generated logical form based on the random utterance; inputting tuples including the generated tuples and a subset of seed tuples selected from the seed data to the classifier, the generated tuples being the generated logical form includes a dumb utterance, and the step of training the classifier further comprises: the classifier determining a predicted authenticity of the tuple; comparing the predicted authenticity with the actual authenticity of the tuple; and updating the identifier based on comparing the predicted authenticity with the actual authenticity of the tuple, the method further comprising: training the semantic parser of the GAN based on the classifier of the GAN, the step of training the semantic parser comprising: training the semantic parser further comprises inputting a seed utterance to the semantic parser, the seed utterance being selected from the seed tuple; the semantic parser determining additional generated logical forms based on the seed utterance; inputting an additional generated tuple including the seed utterance from the seed data and the additional generated logical form into the classifier; the classifier determining additional predicted authenticity of the additional generated tuples; comparing said additional predicted authenticity with a selected authenticity of said additional generated tuple; and updating the semantic parser based on comparing the additional predicted authenticity with the selected authenticity of the additional generated tuple; The semantic parser is usable in a dialogue system to map utterances to logical forms.

20. the actual authenticity of a tuple in the tuples is based on the tuple being a seed tuple in the seed data; 20. The computer program product of claim 19, wherein the selected authenticity of the additional generated tuple indicates that the additional generated tuple is authentic.

21. a voice input module for receiving voice input; a dialogue processing module, which utilizes a semantic parser obtained by using the training method according to any one of claims 1 to 10 to perform or cause to be performed one or more actions based on an interpretation of the speech input and prepares a response; and a voice output module for outputting said response.

Images