Computer-Implemented Method, Computer Program Product, and Computer System (Multi-Language Intent Recognition)

Abstract

To provide a computer implemented method, a program, and a system which allow better training when an amount of speech training data is limited.SOLUTION: The computer comprises accessing one or more intents and associated entities from a limited amount of speech to text training data in a single language; locating speech to text training data in one or more other languages by using the accessed one or more intents and associated entities to locate speech to text training data in the one or more other languages different from the single language; and training a neural network based on the limited amount of speech to text training data in the single language and the located speech to text training data in the one or more other languages.SELECTED DRAWING: Figure 6

Description

Technical Field

[0001] The present invention generally relates to intent recognition, and more particularly to multilingual training for speech intent recognition.

Background Art

[0002] Machine learning (ML) is the scientific study of algorithms and statistical models that a computer system uses to perform specific tasks by relying on patterns and inference instead of using explicit instructions. Machine learning is considered a subset of artificial intelligence. Machine learning algorithms build mathematical models based on sample data known as training data to make predictions or decisions without being explicitly programmed to perform a task. Machine learning algorithms are used in a wide range of applications where it is difficult or impossible to develop conventional algorithms for effectively performing tasks such as email filtering and computer vision.

[0003] In machine learning, hyperparameters are configurations that are external to the model and whose values cannot be estimated from the data. Hyperparameters are used in the process of estimating model parameters. Hyperparameters are set before the learning (e.g., training) process begins, while the values of other parameters are derived through training. Different model training algorithms require different hyperparameters, and some simple algorithms such as the least squares method require none. Given a set of hyperparameters, the training algorithm learns parameter values from the data. For example, the Least Absolute Shrinkage and Selection Operator (LASSO) is an algorithm that adds a regularization hyperparameter to the least squares method and needs to be set before estimating the parameters through the training algorithm. Similar machine learning models may require different hyperparameters (e.g., different constraints, weights, learning rates) to generalize different data patterns.

[0004] Deep learning is a field of machine learning based on a series of algorithms that use model architectures with complex structures or, in other ways, model architectures often composed of multiple non-linear transformations to model high-level abstractions within data. Deep learning is part of a broad family of machine learning techniques based on the learned representations of data. Observations (e.g., images) can be represented in many ways, such as vectors of intensity values per pixel, or in more abstract ways, such as sets of edges, regions of a particular shape, etc. In some representations, tasks can be easily learned from examples (e.g., face recognition or emotion recognition). Deep learning algorithms often use cascades of multiple layers of non-linear processing units for feature extraction and transformation. Each successive layer uses the output from the previous layer as input. The algorithms may or may not be supervised, and applications include pattern analysis (unsupervised) and classification (supervised). Deep learning models include artificial neural networks (ANNs) inspired by information processing in biological systems and decentralized communication nodes. ANNs differ in various ways from the living brain.

[0005] Neural networks (NNs) are computing systems inspired by biological neural networks. NNs are not just simple algorithms but a framework where various machine learning algorithms cooperate to process complex data inputs. Such systems generally "learn" to perform tasks by considering examples without being programmed with task-specific rules. For example, in image recognition, NNs analyze sample images correctly labeled as "cat" or "not a cat" and learn to identify images containing cats by using the results to distinguish cats in other images. For example, NNs achieve this without prior knowledge about cats, such as that cats have fur, tails, whiskers, and pointed ears. Instead, NNs automatically generate discriminative features from the learning materials. NNs are based on a collection of connected units or nodes called artificial neurons, which roughly model the neurons in a living brain. Each connection can transmit a signal from one artificial neuron to another, like a synapse in a biological brain. An artificial neuron that receives a signal can process that signal and transmit it to another artificial neuron.

[0006] In a typical NN implementation, the signals in the connections between artificial neurons are real numbers, and the output of each artificial neuron is calculated by a non-linear function of the sum of its inputs. The connections between artificial neurons are called "edges". Artificial neurons and edges have weights that are adjusted as learning progresses. The weights increase or decrease the strength of the signal in the connection. An artificial neuron can have a threshold such that a signal is transmitted only if the sum of the signals exceeds the threshold. Typically, artificial neurons are grouped into layers. Different layers can perform different types of transformations on the input. Signals move from the first layer (input layer) to the last layer (output layer). They may also move through multiple layers several times.

[0007] A convolutional neural network (CNN) is a type of neural network that is most commonly applied to the analysis of visual images. CNNs are a regularized version of multi-layer perceptrons (e.g., fully connected networks), where each neuron in one layer is connected to all neurons in the next layer. CNNs utilize hierarchical patterns in data and assemble more complex patterns using smaller and simpler patterns. CNNs divide an image into small patches (e.g., 5×5 pixel patches) and move across the image with a specified stride. Therefore, in terms of the scale of connectivity and complexity, CNNs are positioned lower. Since CNNs use little preprocessing compared to other image classification algorithms, the network can learn filters that were conventionally processed manually by traditional algorithms.

[0008] An artificial neural network (ANN) is a computing system inspired by biological neural networks. An ANN itself is not an algorithm but a framework for various machine learning algorithms to cooperate in processing complex data inputs. Such systems generally "learn" to perform tasks by considering examples without being programmed with task-specific rules. For example, in image recognition, ANNs learn to identify images containing cats by analyzing sample images correctly labeled as "cat" or "not a cat" and using the results to identify cats in other images. For example, ANNs achieve this without prior knowledge about cats, such as that cats have fur, tails, whiskers, and pointed ears. Instead, ANNs automatically generate discriminative features from the learning material. ANNs are based on a collection of connected units or nodes called artificial neurons, which roughly model neurons in the living brain. Each connection can transmit a signal from one artificial neuron to another, similar to synapses in the biological brain. An artificial neuron that receives a signal can process that signal and transmit it to another artificial neuron.

[0009] In a general ANN implementation, the signals in the connections between artificial neurons are real numbers, and the output of each artificial neuron is calculated by a non-linear function of the sum of its inputs. The connections between artificial neurons are called "edges". Artificial neurons and edges have weights that are adjusted as learning progresses. The weights increase or decrease the strength of the signal in the connection. An artificial neuron can have a threshold such that a signal is transmitted only if the sum of the signals exceeds the threshold. Typically, artificial neurons are grouped into layers. For different layers, different types of transformations can be performed on the input. The signal moves from the first layer (input layer) to the last layer (output layer). It may also move after moving through multiple layers several times.

[0010] A Recurrent Neural Network (RNN) is a class of ANN where the connections between nodes form a directed graph along a sequence, enabling the network to exhibit temporally dynamic behavior with respect to a time sequence. Different from feedforward neural networks, RNNs can use an internal state (memory) to process a sequence of inputs, which allows RNNs to be applied to tasks such as unsegmented connected handwriting recognition or speech recognition. Long Short-Term Memory (LSTM) units are alternative layer units of a Recurrent Neural Network (RNN). An RNN consisting of LSTM units is called an LSTM network. A common LSTM unit is composed of a cell, an input gate, an output gate, and a forget gate. The cell stores values over an arbitrary time interval, and the gates control the flow of information into and out of the cell. In LSTM, the learning rate following the network size is the most important hyperparameter. SUMMARY OF THE INVENTION PROBLEMS TO BE SOLVED BY THE INVENTION

[0011] Provided are a computer-implemented method, a computer program product, and a computer system that can train better when the amount of voice training data is limited.

Means for Solving the Problem

[0012] According to one aspect of the present invention, a computer-implemented method is provided. The method includes accessing one or more intents and related entities from a limited amount of speech-to-text training data in a single language. Using the accessed one or more intents and related entities to identify speech-to-text training data in one or more other languages different from the single language. Training a neural network based on the identification of the speech-to-text training data in one or more other languages and the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in one or more other languages.

Brief Description of the Drawings

[0013] Preferred embodiments of the present invention will be described with reference to the following drawings.

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Embodiments for Carrying Out the Invention

[0014] Embodiments of the present invention recognize that conventional speech language understanding systems typically consist of two parts. An automatic speech recognition (ASR) system that converts speech to text, followed by a natural language understanding module for intent recognition, entity extraction, etc. In current neural network-based architectures, it is possible to train a single end-to-end system that can directly extract intent and entity information from speech signals without generating intermediate text representations of the input. Embodiments of the present invention recognize the deficiencies of current neural network-based architectures, namely that the amount of task-specific training data (e.g., speech data with intent labels) for training these systems is often limited.

[0015] In these settings, embodiments of the present invention recognize that conventional systems have advantages. Since conventional systems are partially trained, it is possible to achieve good performance by training each component separately. The automatic speech recognition component can be trained with a large amount of transcribed data collected independently without intent labels. The subsequent intent classifier can be trained with a relatively small amount of data. Also, since the data for training the intent classifier is purely text-based, the data sparsity problem of speech data labeled with intents is often alleviated.

[0016] Embodiments of the present invention recognize the difficulty of training end-to-end voice intent systems due to the need to use very large amounts of voice data. Embodiments of the present invention recognize that the amount of voice training data is often limited. This problem is exacerbated when automatic speech recognition considers subtasks, as the voice data also needs to be transcribed.

[0017] Embodiments of the present invention provide a solution for limited training data. In other words, embodiments of the present invention provide a solution for alleviating the data training problem through multilingual training. For example, embodiments of the present invention can pool data from various languages and train a neural network model in multilingual training. This trained neural network model includes a set of shared layers common to the pool of languages. Next, embodiments of the present invention can use language-specific targets. As will be discussed in more detail later in this specification, the parameters of the network model can be better trained by pooling data from multiple languages.

[0018] FIG. 1 is a functional block diagram showing a generally designated computing environment 100 according to an embodiment of the present invention. FIG. 1 provides only an illustration of one embodiment and does not imply any limitation regarding the environment in which different embodiments may be implemented. Many modifications to the illustrated environment may be made by those skilled in the art without departing from the scope of the present invention as recited in the claims.

[0019] Computing environment 100 includes client computing device 102 and server computer 108, all of which are interconnected via network 106. Client computing device 102 and server computer 108 can be stand-alone computer devices, management servers, web servers, mobile computing devices, or any other electronic device or computing system capable of receiving, transmitting, and processing data. In other embodiments, client computing device 102 and server computer 108 can represent a server computing system that utilizes multiple computers, such as in a cloud computing environment. In another embodiment, client computing device 102 and server computer 108 can be a laptop computer, tablet computer, netbook computer, personal computer (PC), desktop computer, personal digital assistant (PDA), smartphone, or any programmable electronic device capable of communicating with various components within computing environment 100 and other computing devices (not shown). In another embodiment, client computing device 102 and server computer 108 each represent a computing system that utilizes clustered computers and components (e.g., database server computers, application server computers, etc.), which function as a single pool of seamless resources when accessed within computing environment 100. In some embodiments, client computing device 102 and server computer 108 are a single device. Client computing device 102 and server computer 108 may include internal and external hardware components capable of executing machine-readable program instructions, as shown and described in more detail with respect to FIG. 7.

[0020] In this embodiment, the client computing device 102 is a user device associated with a user and includes an application 104. The application 104 communicates with the server computer 108 to access the general-purpose intent recognizer 110 (e.g., using TCP / IP) and access content, user information, and database information. The application 104 can further communicate with the general-purpose intent recognizer 110 to send instructions for training a multi-language neural network intent classifier, as discussed in more detail with respect to FIGS. 2-6.

[0021] The network 106 can be, for example, a telecommunications network, a local area network (LAN), a wide area network (WAN) such as the Internet, or a combination of these three, and can include wired, wireless, or fiber optic connections. The network 106 can include one or more wired or wireless networks or both that can receive and transmit data, voice, or video signals (including multimedia signals that include voice, data, and video information) or combinations thereof. Generally, the network 106 can be any combination of connections and protocols that support communication between the client computing device 102 and the server computer 108 and other computing devices (not shown) within the computing environment 100.

[0022] Server computer 108 is a digital device that hosts general-purpose intent recognizer 110 and database 112. In this embodiment, general-purpose intent recognizer 110 resides on server computer 108. In other embodiments, general-purpose intent recognizer 110 can have an instance of a program (not shown) stored locally on client computer device 102. In other embodiments, general-purpose intent recognizer 110 can be a stand-alone program or system that trains a multi-language neural network intent classifier. In still other embodiments, general-purpose intent recognizer 110 can be stored on any number of computing devices.

[0023] General-purpose intent recognizer 110 trains a multi-language neural network intent classifier. That is, general-purpose intent recognizer 110 can identify the intent from the voice or text regardless of the language associated with the received content (such as voice or text). In this embodiment, general-purpose intent recognizer 110 includes intent classifier 114. Intent classifier 114 classifies the intent from the received content.

[0024] As used herein, "intent" refers to a mental or purposeful state. For example, "intent" refers to the purpose, will, or determination behind an action, thought, or utterance. As a specific example, in the sentence "I want to book a flight from New York to Boston", the intent is "flight booking". New York and Boston are the values corresponding to "departure city" and "arrival city". The set of entities and values within the sentence can also be regarded as part of the intent of the sentence. There are many other attributes that can be classified or referred to as "intent", such as part-of-speech tags, dialogue state tags, etc.

[0025] In this embodiment, the content refers to the received media. For example, the media can include one or more audio files that include voice. The media can also include received text or text files. In some embodiments, the media can further include video files that include audio (e.g., voice).

[0026] In this embodiment, the general-purpose intent recognizer 110 uses multi-lingual voice (e.g., known training data) as received input and pre-trains a neural network by pooling together shared parameters trained in all languages, as will be described in more detail with respect to FIG. 2.

[0027] In this embodiment, the term "shared parameters" is used to represent a set of common layers of a neural network trained using multi-lingual data. The layers of a neural network include a set of nodes. Each layer is connected to other layers via connections with weights. Nodes are associated with different types of non-linearity, bias terms, gates that control information flow, etc. There are several variations for nodes and network connections depending on the type of network that can be used. LSTMs, CNNs, RNNs, DNNs are examples of neural networks. For example, the shared parameters 204 (discussed and described in FIG. 2) are a representation of a set of common network layers that form part of a multi-lingual intent classifier.

[0028] In this embodiment, the term "language-specific parameter" is used to represent a set of layers of a neural network trained using language-specific data. The layers of the neural network include a set of nodes. Each layer is connected to other layers via connections with weights. Nodes are associated with different types of non-linearity, bias terms, gates that control information flow, etc. Depending on the type of network that can be used, there are several variations in nodes and network connections. LSTMs, CNNs, RNNs, DNNs are examples of neural networks. The language-specific parameters 206A-N (shown in FIG. 2) are representations of language-specific network layers that form part of a multi-language intent classifier. While processing multi-language data from N languages, component 204 (also called shared parameter 204) is trained on all the data. On the other hand, components 206A-N (also called language-specific parameters 206A-N) are trained on data specific to each language. Each component 206A-N is connected to a single shared component 204. All languages also have language-specific intent prediction layers represented by 208A-N (also called intent languages 208A-N).

[0029] In this embodiment, the intent classifier 114 can then analyze the received content into individual language-specific parameters. Here, consider a reservation system that processes flight reservations in Spanish and English. To train this system, a corpus consisting of Spanish and English voice utterances is used. The multilingual voice corpus is annotated for various purposes such as "flight reservation", "flight cancellation", "flight status check", "travel reservation change", etc. Each utterance is also annotated with entities. At test time, when an English utterance passes through this network, the recognized intent and entities become available in the English-specific output of the network. For a Spanish utterance, the output becomes available in the Spanish-specific output layer. Such a network is pre-trained with the network architecture shown in FIG. 2. The network has layers trained with both Spanish and English data. Next, the multilingual voice representations from these shared network layers are passed to the language-specific layers to generate the desired intent output.

[0030] In the Spanish training data, the training signal passes through the shared layer, through the Spanish-specific layer, and is verified with the Spanish-specific output. In the English data, the training signal similarly passes through the shared layer before being processed by the English-specific layer and the output is collected at the English-specific output layer.

[0031] In another embodiment, the available multilingual data may not have intent labels and may simply have transcripts. Even in this case, the multilingual network can be pre-trained. Similar to the aforementioned pre-training, the available multilingual data can be used to train the shared layer and the language-specific layers with the multilingual data.

[0032] In another embodiment, the available data may be in a single language, but can, for example, accommodate different domains such as banking, airlines, hospitality, etc. In such a setup, data from various domains are pooled together. Similar to the multilingual case, the shared layer is trained on all the available data. However, the language-specific layers correspond to domain-specific layers. There are sets of layers corresponding to the banking domain, different sets of layers corresponding to the airline domain, etc. All of these domain-specific layers are connected to the set of shared layers.

[0033] In another embodiment, the available data may be in a single language, but may be drawn from different data sets, for example, data sets collected to model an interactive state, data sets of voice commands, etc. In this setup, data from various data sets are pooled together again. Similar to the multilingual case, the shared layer is trained on all the available data. However, the language-specific layers correspond to data-set-specific layers. There are sets of layers corresponding to the interactive state data set, different sets of layers corresponding to voice commands, etc. All of these domain-specific layers are connected to the set of shared layers.

[0034] In another embodiment, the model can also handle cases involving language switching. The data can, for example, have sentences where a person speaks in English, switches to Spanish, and then back to English (also known as code-switching). In this case, the data is from the same domain, but different languages are pooled, so the system can utilize the commonality of intents and entities. When trained in such a setting, in this embodiment, the system can process the speech using code-switching when terms in two languages are used within the same utterance. This would also be useful for call center analysis where it is necessary to analyze client data in different languages together. For example, it can analyze how often people are booking flights to Houston this month compared to last month, regardless of which language is being spoken.

[0035] Next, the intent classifier 114 can identify the respective intents from each language identified or otherwise. Once the multi-language / multi-domain / multi-corpus network is pre-trained as described above, it can be immediately refined for the final intent classification task. Here, consider the above-mentioned multi-language flight reservation network. The network is trained with both English and Spanish data using general flight reservation data and labels. Now, this network can be adapted to a specific airline and its specific data. Depending on the nature of the data, the entire pre-trained network or a part of the network can be used.

[0036] Scenario 1: The new data received is in English and has the general labels used for training the data. The new data includes acoustic characteristics due to the nature of different demographic data collection. Initialize a new network using the shared pre-trained network layer and the English-specific layer. This new network is trained with the new data.

[0037] Scenario 2: New data is received in English but is labeled with a new set of intent labels. In this case, the new network trained on the received data is initialized with the shared multi - language layer and the English - specific layer, but the new intent output layer is used to cover the new set of intent labels. Once initialized, the network is fully trained.

[0038] Scenario 3: New data is received in German and is also labeled with a new set of intent labels. In this setup, the new network is initialized with only the shared multi - language parameters. Use cases like this can also be assumed for multi - domain / multi - corpus scenarios.

[0039] Next, the intent classifier 114 can receive in real - time unknown media, i.e., unconfirmed audio. Next, as will be described in more detail with respect to Figure 3, the intent classifier 114 can access the shared parameters trained in all languages, identify each language that matches the unknown media and the associated language - specific parameters, and identify the intent associated with the unknown media.

[0040] In other embodiments, the general - purpose intent recognizer 110 can be modified to predict intents that occur across domains and languages. For example, a multi - language corpus can represent two domains, such as airlines and hospitality, in two languages (e.g., English and Spanish).

[0041] In this embodiment, the general - purpose intent recognizer 110 can include an additional layer trained on top of the language - specific layer that has learned the language - specific configuration, as will be discussed in more detail with respect to Figure 4.

[0042] Now, consider the foregoing embodiments of the multilingual airline reservation system. The data that can be used for training this system can be utilized in two parts. In the first data cut, only the transcripts of English and Spanish voices are available. The second data cut is a much smaller portion of the corpus that has both transcripts and intent labels. With the first data cut, the network can be trained to identify the key parts of each utterance that convey meaning. For example, part-of-speech tags can be attached to the transcripts. Words marked as nouns are marked as values of entities such as "destination airport", and verb phrases such as "fly to" help identify the intent. However, these syntactic structures vary by language. As will be described in more detail with respect to FIG. 4, layers 406A - N in a language-independent configuration of 404 can be considered as layers for learning part-of-speech (POS) representations for each language that are pooled together by the layer of 404. The last layer of 408 maps various part-of-speech (POS) representations to intent labels.

[0043] Thus, embodiments of the present invention can provide solutions for parameter sharing / learning from different data sets (multilingual, multi-domain, multiple data sets). Front-end sharing (e.g., shared parameters 404 across all languages) is useful for general speech modeling, for example. However, there is also sharing that is useful for modeling common language structures such as nouns and verbs. If there are similar domains (e.g., travel) in multiple languages, practical "domain logic" can be shared. This type of sharing is captured by a shared layer after the language-specific parameter layers (the second block, currently also marked as 404, needs to be numbered differently because it has a different level of sharing).

[0044] Also, according to an embodiment of the present invention, even when domains are not similar, data can be shared. Through data sharing and pre-training of specific parameters, the embodiments of the present invention capture some universal features of spoken language understanding (SLU) and mitigate the lack of data in any specific domain / language combination.

[0045] The multiple embodiments discussed in this specification provide various levels of composition regarding how intent labels can be composed. There are cases where intent labels are standardized / shared and cases where labels are not standardized / shared. In both such cases, benefits should be obtained from the solutions provided by specific embodiments of the present invention because it enables shared parameters and language / domain-specific parameters at multiple levels.

[0046] For example, the general-purpose intent recognizer 110 can receive content including one or more languages, access shared parameters trained in all languages, identify one or more languages, and identify specific parameters associated with each identified language. Next, the intent classifier 114 can access the shared parameters again to identify language-independent intents. The various layers of the neural network can be considered as transformations applied to input signals to generate various representations at levels of abstraction. The shared layer discussed in FIG. 5 (e.g., the shared parameters 504 across all languages) serves to remove unnecessary channel and speaker variations. Next, the language expressions from the perspective of basic acoustic units and the like generated by the shared layer are transformed by layers 506A - N to model language-specific details for each language. The language-specific details include language-specific graphemes or phonetic representations. For each language, these refined representations can be used to extract actual intent and grapheme symbols and construct entity values modeled by layers such as 508A - B.

[0047] In yet other embodiments, the intent classifier 114 can be adapted to a specific domain within the same language or even to a new language without retraining from scratch using a multilingual model. In such cases, the final language-specific parameters of the existing model are replaced with those of the new domain and respective language-specific layers, while the shared layers are kept as they are. The new model is then completed by training with new data.

[0048] Domains and languages that match (e.g., within a certain threshold percentage) the domains and languages used for training the multilingual network are examples of settings where the trained model can be processed without retraining. Using the aforementioned multilingual travel reservation system, after training to create general travel intents, the multilingual travel reservation system can be deployed in English or Spanish. The final language-specific parameters are replaced using a new output layer corresponding to the new set of intent labels as described above (e.g., regarding how various parts of the network can be kept as they are or replaced in other ways).

[0049] In yet other embodiments, the intent classifier 114 can be adapted to train the neural network on other related tasks in addition to the first classification task. In this way, the first classification task can be improved. As previously described in the example of when an utterance is processed by the intent classifier, not only is an intent label generated, but often values corresponding to various entities need to be generated. To correctly recognize these values, the intent recognition system (e.g., the general intent recognizer 110) must be able to accurately generate the text transcription. If the first classification task is intent recognition (e.g., "flight reservation"), the related classification task can be entity recognition (where the departure and destination airports are also recognized).

[0050] For example, embodiments of the present invention can improve the training of a proposed network for intent recognition and multi-task training with other related tasks (e.g., speech / text recognition as discussed in more detail with respect to FIG. 5). Some embodiments of the present invention can train these networks by self-monitoring after the training data has been appropriately modified (e.g., addition of noise, modification of speed / tone).

[0051] Database 112 can represent one or more databases or publicly available databases that store the received information and provide authorized access to the general intent recognizer 110. In general, database 112 can be implemented using any non-volatile storage medium known in the art. For example, database 112 can be implemented using a tape library, an optical library, one or more independent hard disk drives, or multiple hard disk drives within a redundant array of independent disks (RAID). In this embodiment, database 112 is stored in server computer 108.

[0052] FIG. 2 shows an exemplary block diagram 200 for training an end-to-end intent classifier according to an embodiment of the present invention.

[0053] In this example, input 202 is supplied to intent classifier 114. Input 202 can include any combination of audio, text, and video. For example, input 202 can include multilingual speech identified within an audio file. In other embodiments, input 202 can be a live stream of audio. In other embodiments, input 202 can be in a single language. Next, intent classifier 114 can access shared parameter 204. Shared parameter 204 can be a set of pre-trained data.

[0054] Next, the intent classifier 114 can identify the language-specific parameters 206A, 206B to 206N, and can identify the output intent languages 208A, 208B, and 208N respectively. Generally, the intent languages 208A, 208B, and 208N are one or more respective languages and intents related to the received input 202.

[0055] FIG. 3 shows an example of a block diagram 300 of an end-to-end speech / intent classifier adapted to a specific language according to an embodiment of the present invention.

[0056] This example shows a model initialized from a multilingual model and adapted to a specific task of a specific language. Thus, component 304 (also referred to as shared parameter 304 trained on all languages) is initialized from component 204.

[0057] In this example, the input 302 is supplied to the intent classifier 114. Similar to the input 202 in FIG. 2, the input 302 can include any combination of audio, text, and video. In this embodiment, the input 302 can be an unknown or otherwise unprocessed audio by the intent classifier 114. In other words, the input 302 can be untrained data supplied via a pre-trained intent classifier. In other embodiments, the input 302 can be a live stream of audio. Next, the intent classifier 114 can access the shared parameter 304 trained on all languages.

[0058] Next, the intent classifier 114 can identify the language-specific parameter 306 specific to the received input 302, and thus can identify the intent 308 that is the intent of the received input 302.

[0059] FIG. 4 shows an example of a block diagram 400 of a multilingual end-to-end speech intent classifier for general intent recognition according to an embodiment of the present invention.

[0060] In this example, the input 402 is supplied to or otherwise accessed by the intent classifier 114. The input 402 can include any combination of audio, text, and video. In this example, the input 402 is multilingual speech. In other embodiments, the input 402 can be a live stream of audio of multilingual speech. Next, the intent classifier 114 can access the shared parameters 404 that are trained across all languages.

[0061] From there, the intent classifier 114 can identify the language-specific parameters 406A, 406B through 406N. Different from the intent recognition device (including the intent classifier 114) discussed in FIG. 2, the intent classifier 114 then makes a second access to the shared parameters 404 across all languages. The intent classifier 114 can use the shared parameters 404 across all languages to identify language-independent intents 408. As described above, the part-of-speech tags can be considered as language-independent intents. In FIG. 4, the layers 404, 406A through N of the constructs independent of the learning language can be considered as layers that learn the part-of-speech (POS) representations of each language that are pooled together for each layer 404. The various part-of-speech (POS) representations of the last layer 408 are mapped to intent labels.

[0062] In other embodiments, an intent recognizer having the intent classifier 114 can be adapted to a specific domain within the same language or even to a new language without retraining from scratch using a multilingual model. In such cases, the final language-specific parameters of the existing model are replaced with new domain and respective language-specific layers, and the shared layers are maintained as they are. The new model is completed by training with new data.

[0063] FIG. 5 shows an example of a block diagram 500 showing multi-task training of a multi-language end-to-end speech / intent classifier according to an embodiment of the present invention.

[0064] In this example, a general-purpose intent recognizer having an intent classifier 114 can be further improved (i.e., better trained) by assigning subtasks in addition to the first task. The first task used here means the first task of identifying an intent from given or accessed speech data. The subtask can be other related tasks (e.g., speech / text recognition). In still other embodiments, the general-purpose intent recognizer can train these networks by self-monitoring after the training data has been appropriately modified (e.g., adding noise, modifying speed / tone, etc.).

[0065] In this example, the input 502 is supplied to or otherwise accessed by the intent classifier 114. Similar to the input 202 of FIG. 2, the input 502 can include any combination of audio, text, and video. In this example, the input 502 is multi-language speech. In other embodiments, the input 202 can be a live stream of multi-language speech audio. Next, the intent classifier 114 can access shared parameters 504 trained on all languages.

[0066] From there, the intent classifier 114 can identify language-specific parameters 506A, 506B to 506N, and thus can identify primary outputs 508A, 510A, and 512A, and corresponding secondary outputs 508B, 510B, and 512B, respectively. As described above, the first task is intent recognition. The second task is to recognize graphemes / phonemes / words in the input, often called automatic speech recognition. The recognized sequence of graphemes / phonemes / words is used to construct entity values.

[0067] Figure 6 is a flowchart 600 showing the operational steps for training an end-to-end speech, multi-language intent classifier according to an embodiment of the present invention.

[0068] In step 602, the general-purpose intent recognizer 110 receives information. In this embodiment, the general-purpose intent recognizer 110 receives a request from the client computing device 102. In other embodiments, the general-purpose intent recognizer 110 can receive information from one or more other components of the computing environment 100.

[0069] The information received by the general-purpose intent recognizer 110 refers to voice or text information. The information received by or otherwise accessible to the general-purpose intent recognizer 110 can include any combination of audio, text, and video. For example, the information can include one or more languages. The information can include a limited amount of speech-to-text training data in a single language. For example, a limited amount of speech-to-text training data in a single language can include interpreting intents and related entities from speech in a single language. In certain embodiments, the information can be a live stream of audio of multi-language speech. For example, the general-purpose intent recognizer 110 can receive information from one or more connected IoT devices.

[0070] In step 604, the generic intent recognizer 110 determines an intent from the received information. In cases where the generic intent recognizer 110 is initially trained, the received information can be used to pre-train the generic intent recognizer 110. In this embodiment, the generic intent recognizer 110 can determine the intent from the received information using a combination of one or more machine learning and artificial intelligence algorithms. In some embodiments, the generic intent recognizer 110 can use an existing neural network-based architecture. The generic intent recognizer 110 can then store the determined intent as part of the pre-trained data in the database 112.

[0071] In particular embodiments, if the generic intent recognizer 110 has already been pre-trained, the generic intent recognizer 110 may have access to pre-trained data including one or more intents and associated entities from a limited amount of speech-to-text training data in a single language.

[0072] In step 606, the generic intent recognizer 110 trains a neural network based on the determined intent. For example, the generic intent recognizer 110 trains the neural network using pre-trained data. In this embodiment, the pre-trained data may include shared parameters for all languages. The shared parameters are layers of the network jointly trained for all languages.

[0073] Next, the general-purpose intent recognizer 110 uses pre-trained data (e.g., one or more accessed intents and related entities) to identify training data for speech-to-text in one or more other languages, and can identify training data for speech-to-text in one or more other languages different from a single language. Next, the general-purpose intent recognizer 110 determines language-specific parameters based on shared parameters accessed for all languages, and subsequently stores the identified language-specific parameters in a database (e.g., database 112).

[0074] In step 608, the general-purpose intent recognizer 110 trains a neural network for natural language processing. In this embodiment, the general-purpose intent recognizer 110 trains a neural network for natural language processing by accessing an updated database (e.g., database 112 updated in the previous step). Accordingly, the general-purpose intent recognizer 110 can utilize the trained intent classifier 114 to identify an intent from the received input. More specifically, the general-purpose intent recognizer 110 can receive a multilingual input and identify an intent using a smaller training data sample size regardless of the received language.

[0075] Accordingly, the general-purpose intent recognizer 110 can be trained by preparing multilingual / multidomain / multicorpus data having transcripts and intents. Next, the general-purpose intent recognizer 110 can be trained using a shared layer and a language-specific layer. In certain embodiments, the general-purpose intent recognizer 110 can prepare domain / language / corpus-specific data, or is otherwise prepared. Finally, the general-purpose intent recognizer 110 can adapt the general-purpose intent recognizer using data from the shared layer and the language-specific layer, and initialize as many layers as possible from a pre-trained network.

[0076] FIG. 7 shows a block diagram of components of a computing system within computing environment 100 of FIG. 1, according to one embodiment of the present invention. It should be understood that FIG. 7 provides only an illustration of one embodiment and does not imply limitations as to the environments in which different embodiments may be implemented. Many modifications may be made to the depicted environments.

[0077] The programs described herein are identified based on the applications in which they are implemented in particular embodiments of the present invention. However, it should be understood that the nomenclature of any particular program herein is used for convenience only and, thus, the present invention should not be limited to use in any particular applications identified or implied or both by such nomenclature.

[0078] The computer system 700 includes a communication fabric 702 that provides communication between a cache 716, a memory 706, a persistent storage 708, a communication unit 712, and an input / output (I / O) interface 714. The communication fabric 702 can be implemented in any architecture designed to pass data or control information or both between a processor (such as a microprocessor, communication processor, network processor, etc.), system memory, peripheral devices, and any other hardware component within the system. For example, the communication fabric 702 can be implemented using one or more buses or a crossbar switch.

[0079] Memory 706 and persistent storage device 708 are computer-readable storage media. In this embodiment, memory 706 includes random access memory (RAM). Generally, memory 706 can include any suitable volatile or non-volatile computer-readable storage media. Cache 716 is a high-speed memory that improves the performance of computer processor 704 by holding recently accessed data and data close to the accessed data from memory 706.

[0080] General-purpose intent recognizer 110 (not shown in FIG. 7) is stored in persistent storage device 708 and memory 706 and is executed by one or more of the respective computer processors 704 via cache 716. According to one embodiment, persistent storage device 708 includes a magnetic hard disk drive. Alternatively, or in addition to the magnetic hard disk drive, persistent storage device 708 can include a solid-state hard drive, a semiconductor storage device, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a flash memory, or any other computer-readable storage media capable of storing program instructions or digital information.

[0081] The media used by persistent storage device 708 may be removable. For example, a removable hard drive can be used for persistent storage device 708. Other examples include optical and magnetic disks, thumb drives, and smart cards that are inserted into a drive for transfer to another computer-readable storage media that is also part of persistent storage device 708.

[0082] In these examples, the communication unit 712 provides communication with other data processing systems or devices. In these examples, the communication unit 712 includes one or more network interface cards. The communication unit 712 can provide communication using either or both physical and wireless communication links. The general-purpose intent recognizer 110 can be downloaded to the persistent storage device 708 via the communication unit 712.

[0083] The I / O interface 714 enables the input and output of data with other devices that can be connected to a client computing device or a server computer or both. For example, the I / O interface 714 can provide a connection to an external device 720 such as a keyboard, keypad, touch screen, or other suitable input device or a combination thereof. The external device 720 can also include a portable computer-readable storage medium such as a thumb drive, portable optical disk or magnetic disk, and memory card. Software and data used to implement embodiments of the present invention, such as the general-purpose intent recognizer 110, can be stored on a portable computer-readable storage medium and loaded onto the persistent storage device 708 via the I / O interface 714. The I / O interface 714 is also connected to a display 722.

[0084] The display 722 provides a mechanism for displaying data to the user and can be, for example, a computer monitor.

[0085] The present invention can be a system, a method, or a computer program product or both. The computer program product can include a computer-readable storage medium storing computer-readable program instructions for causing a processor to execute aspects of the present invention.

[0086] A computer-readable storage medium can be a tangible device that holds and stores instructions used by an instruction execution device. A computer-readable storage medium can be, by way of example, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or a suitable combination thereof. As a more specific example of a computer-readable storage medium, there can be a portable computer diskette, a hard disk, a RAM, a ROM, an EPROM (or flash memory), an SRAM, a CD-ROM, a DVD, a memory stick, a floppy disk, a punched card, a mechanically encoded device with instructions recorded on a raised structure in a groove, and suitable combinations thereof. A computer-readable storage device as used herein should not be construed as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., an optical pulse passing through an optical fiber cable), or an electrical signal transmitted via a wire.

[0087] The computer-readable program instructions described herein are downloadable from a computer-readable storage medium to respective computer devices / processing devices. Alternatively, they are downloadable via a network (e.g., the Internet, a LAN, a WAN, or a wireless network, or a combination thereof) to an external computer or an external storage device. The network can comprise a copper transmission cable, an optical transmission fiber, a wireless transmission, a router, a firewall, a switch, a gateway computer, or an edge server, or a combination thereof. A network adapter card or network interface within each computer device / processing device receives the computer-readable program instructions from the network and transfers the computer-readable program instructions for storage in a computer-readable storage medium in each respective computer device / processing device.

[0088] The computer-readable program instructions for carrying out the operations of the present invention may be source code or object code written in any combination of one or more programming languages, including assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or object-oriented programming languages such as Smalltalk and C++, and conventional procedural programming languages such as the "C" programming language and similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer as a stand-alone software package, or partially on the user's computer. Alternatively, it may be executed partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer may be connected to the user's computer via any type of network, including a LAN or WAN, or to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, an electronic circuit, including a programmable logic circuit, a field programmable gate array (FPGA), and a programmable logic array (PLA), can execute the computer-readable program instructions by utilizing the state information of the computer-readable program instructions to customize the electronic circuit for the purpose of implementing aspects of the present invention.

[0089] Each aspect of the present invention is described herein with reference to flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. Each block in the flowchart and / or block diagram, and combinations of multiple blocks in the flowchart and / or block diagram, can be executed by computer-readable program instructions.

[0090] The above computer-readable program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine. Thereby, these instructions, executed via the processor of such computer or other programmable data processing apparatus, create means for performing the functions / operations specified in one or more blocks in a flowchart and / or a block diagram and / or both. The above computer-readable program instructions may further be stored in a computer-readable storage medium that can be instructed to function in a specific manner for a computer, programmable data processing apparatus, or other devices or combinations thereof. Thereby, the computer-readable storage medium in which the instructions are stored constitutes a product including instructions for performing the mode of the functions / operations specified in one or more blocks in a flowchart and / or a block diagram and / or both.

[0091] Also, a computer-executable process may be generated by loading computer-readable program instructions onto a computer, other programmable apparatus, or other devices and causing a series of operation steps to be executed on the computer, other programmable apparatus, or other devices. Thereby, the instructions executed on the computer, other programmable apparatus, or other devices perform the functions / operations specified in one or more blocks in a flowchart and / or a block diagram and / or both.

[0092] The flowcharts and block diagrams in the drawings of the present disclosure illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram can represent a module, segment, or portion of one or more executable instructions for performing a particular logical function. In some other implementations, the functions shown within a block may be executed in an order different from that shown in each figure. For example, two consecutive blocks shown may actually be accomplished as one step, depending on the related functions, may be executed simultaneously or substantially simultaneously, may be executed in a partially or wholly temporally overlapping manner, or may be executed in the reverse order in some cases. It should be noted that each block in a block diagram or flowchart or both, and combinations of multiple blocks in a block diagram or flowchart or both, can be executed by a dedicated hardware-based system for performing a particular function or operation, or by a combination of dedicated hardware and computer instructions.

[0093] The description of various embodiments of the present invention is presented for purposes of illustration but is not intended to be exhaustive or to limit the invention to the disclosed embodiments. It will be apparent to those skilled in the art that many modifications and variations are possible without departing from the scope and spirit of the invention. The terms used herein are selected to best explain the principles of the embodiments, the practical application to or technical improvement of technologies found in the marketplace, or to enable those skilled in the art to understand the embodiments described herein.

Claims

1. accessing one or more intents and associated entities from a limited amount of speech-to-text training data in a single language; using the accessed one or more intents and associated entities to identify speech-to-text training data in one or more other languages different from the single language; training a neural network based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages, a computer-implemented method comprising.

2. The computer-implemented method of claim 1, further comprising training the neural network for natural language processing based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages.

3. The computer-implemented method of claim 1, further comprising receiving a limited amount of speech-to-text training data for training the neural network.

4. The computer-implemented method of claim 1, wherein the limited amount of speech-to-text training data comprises a single language and intents and associated entities interpreted from the speech of the single language.

5. The computer-implemented method of claim 1, wherein the limited amount of speech-to-text training data comprises a single language drawn from different data sets including a data set collected to model a dialogue state and a data set of voice commands.

6. The computer-implemented method of claim 1, further comprising enabling language switching by accessing a common domain and pooling different language data sets having a commonality of intent and entity sharing.

7. The parameters of the neural network trained using speech-to-text training data represent a set of layers of the neural network including a set of nodes, and each layer is connected to other layers within the set of layers via respective weighted connections. The computer-implemented method according to claim 1.

8. Program instructions for accessing one or more intents and related entities from a limited amount of speech-to-text training data in a single language, Program instructions for identifying speech-to-text training data in one or more other languages different from the single language using the accessed one or more intents and related entities, Program instructions for training a neural network based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages. A program.

9. The program is The program according to claim 8, further comprising program instructions for training the neural network for natural language processing based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages.

10. The program is The program according to claim 8, further comprising program instructions for receiving a limited amount of speech-to-text training data for training the neural network.

11. The limited amount of speech-to-text training data according to claim 8 includes a single language and intents and related entities interpreted from the speech of the single language.

12. The limited amount of speech-to-text training data according to claim 8 includes a single language drawn from different data sets including a data set collected for modeling a dialogue state and a data set of voice commands.

13. The program is The program according to claim 8, further comprising program instructions that enable language switching by accessing a common domain and pooling different language datasets having commonality in the sharing of intents and entities.

14. The parameters of a neural network trained using speech-to-text training data represent a set of layers of a neural network including a set of nodes, and each layer is connected to other layers within the set of layers via respective weighted connections. The program according to claim 8.

15. One or more computer processors, One or more computer-readable storage media, Program instructions stored in the one or more computer-readable storage media for execution by at least one of the one or more computer processors, the program instructions comprising: Program instructions for accessing one or more intents and related entities from a limited amount of speech-to-text training data in a single language; Program instructions for identifying speech-to-text training data in one or more other languages different from the single language using the one or more accessed intents and related entities; A computer system comprising program instructions for training a neural network based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages.

16. The program instructions stored in the one or more computer-readable storage media The computer system according to claim 15, further comprising program instructions for training the neural network for natural language processing based on the limited amount of speech-to-text training data in the single language and the identified speech-to-text training data in the one or more other languages.

17. The program instructions stored in the one or more computer-readable storage media further include program instructions for receiving a limited amount of speech-to-text training data for training the neural network. The computer system according to claim 15, further comprising program instructions for receiving a limited amount of speech-to-text training data for training the neural network. **Claim 18** The computer system according to claim 15, wherein the limited amount of speech-to-text training data includes a single language, and intents and associated entities interpreted from the speech of the single language. **Claim 19** The computer system according to claim 15, wherein the limited amount of speech-to-text training data includes a single language drawn from different data sets including a data set collected for modeling an interactive state and a data set of voice commands. **Claim 20** The program instructions stored in the one or more computer-readable storage media further include program instructions for enabling language switching by accessing a common domain and pooling different language data sets having commonality in the sharing of intents and entities. The computer system according to claim 15, further comprising program instructions for enabling language switching by accessing a common domain and pooling different language data sets having commonality in the sharing of intents and entities.

Images