Abstract: Machine classifiers in accordance with embodiments of the invention capture long-term temporal dependencies in the dialogue data better than the existing recurrent neural network-based architectures. Additionally, machine classifiers may model the joint distribution of the context and response as opposed to the conditional distribution of the response given the context as employed in sequence-to-sequence frameworks. Further, input data may be bidirectionally encoded using both forward and backward separators. The forward and backward representations of the input data may be used to train the machine classifiers using a single generative model and/or shared parameters between the encoder and decoder of the machine classifier. During inference, the backward model may be used to reevaluate previously generated output sequences and the forward model may be used to generate an output sequence based on the previously generated output sequences.

Abstract: Systems described herein may use transformer-based machine classifiers to perform a variety of natural language understanding tasks including, but not limited to sentence classification, named entity recognition, sentence similarity, and question answering. The exceptional performance of transformer-based language models is due to their ability to capture long-term temporal dependencies in input sequences. Machine classifiers may be trained using training data sets for multiple tasks, such as but not limited to sentence classification tasks and sequence labeling tasks. Loss masking may be employed in the machine classifier to jointly train the machine classifier on multiple tasks simultaneously. The user of transformer encoders in the machine classifiers, which treat each output sequence independently of other output sequences, in accordance with aspects of the invention do not require joint labeling to model tasks.

Abstract: Machine classifiers in accordance with embodiments of the invention capture long-term temporal dependencies in particular tasks, such as turn-based dialogues. Machine classifiers may be used to help users to perform tasks indicated by the user. When a user utterance is received, natural language processing techniques may be used to understand the user's intent. Templates may be determined based on the user's intent in the generation of responses to solicit information from the user. A variety of persona attributes may be determined for a user. The persona attributes may be determined based on the user's utterances and/or provided as metadata included with the user's utterances. A response persona may be used to generate responses to the user's utterances such that the generated responses match a tone appropriate to the task. A response persona may be used to generate templates to solicit additional information and/or generate responses appropriate to the task.

Abstract: Aspects described herein may relate to the determination of data that is indicative of a greater range of speech properties than input text data. The determined data may be used as input to one or more speech processing tasks, such as model training, model validation, model testing, or classification. For example, after a model is trained based on the determined data, the model's performance may exhibit more resilience to a wider range of speech properties. The determined data may include one or more modified versions of the input text data. The one or more modified versions may be associated with the one or more speakers or accents and/or may be associated with one or more levels of semantic similarity in relation to the input text data. The one or more modified versions may be determined based on one or more machine learning algorithms.

Abstract: The disclosed technology involves autonomously identifying goals and sub-goals from a user utterance and generating responses to the user based on the goals and sub-goals.

Abstract: Systems and methods are provided herein for utilizing a knowledge base to improve online automated dialogue responses based on machine learning models. Contextual customer data stored in external memory may be used for retraining a machine learning model to incorporate new observations into the model and to reduce bias and/or improve fairness in associated automated responses without having to retrain an entire memory architecture. The disclosed technology may improve the accuracy of machine learning models by using potentially private contextual customer data to inform the model while eliminating the ability of an intruder to access such data when the model is utilized in cloud-based services.

Abstract: In a variety of embodiments, machine classifiers may model multi-turn dialogue as a one-to-many prediction task. The machine classifier may be trained using adversarial bootstrapping between a generator and a discriminator with multi-turn capabilities. The machine classifiers may be trained in both auto-regressive and traditional teacher-forcing modes, with the generator including a hierarchical recurrent encoder-decoder network and the discriminator including a bi-directional recurrent neural network. The discriminator input may include a mixture of ground truth labels, the teacher-forcing outputs of the generator, and/or noise data. This mixture of input data may allow for richer feedback on the autoregressive outputs of the generator. The outputs can be ranked based on the discriminator feedback and a response selected from the ranked outputs.

Abstract: In a variety of embodiments, machine classifiers may model multi-turn dialogue as a one-to-many prediction task. The machine classifier may be trained using adversarial bootstrapping between a generator and a discriminator with multi-turn capabilities. The machine classifiers may be trained in both auto-regressive and traditional teacher-forcing modes, with the generator including a hierarchical recurrent encoder-decoder network and the discriminator including a bi-directional recurrent neural network. The discriminator input may include a mixture of ground truth labels, the teacher-forcing outputs of the generator, and/or noise data. This mixture of input data may allow for richer feedback on the autoregressive outputs of the generator. The outputs can be ranked based on the discriminator feedback and a response selected from the ranked outputs.

Abstract: Systems described herein may use transformer-based machine classifiers to perform a variety of natural language understanding tasks including, but not limited to sentence classification, named entity recognition, sentence similarity, and question answering. The exceptional performance of transformer-based language models is due to their ability to capture long-term temporal dependencies in input sequences. Machine classifiers may be trained using training data sets for multiple tasks, such as but not limited to sentence classification tasks and sequence labeling tasks. Loss masking may be employed in the machine classifier to jointly train the machine classifier on multiple tasks simultaneously. The user of transformer encoders in the machine classifiers, which treat each output sequence independently of other output sequences, in accordance with aspects of the invention do not require joint labeling to model tasks.

Abstract: A system for using hash keys to preserve privacy across multiple tasks is disclosed. The system may provide training batch(es) of input observations each having a customer request and stored task to an encoder, and assign a hash key(s) to each of the stored tasks. The system may provide a new batch of input observations with a new customer request and new task to the encoder. The encoder may generate a new hash key assigned to the new customer request and determine whether any existing hash key corresponds with the new hash key. If so, the system may associate the new batch of input observations with the corresponding hash key and update the corresponding hash key such that it is also configured to provide access to the new batch of input observations. If not, the system may generate a new stored task and assign the new hash key to it.

Abstract: A method, computer system, and computer program product for decision support is provided. The present invention may include receiving a problem case information and generating a query based on the problem case information. The present invention may also include generating a plurality of answers for the query using the question-answering module. The present invention may also include calculating numerical values for multiple evidence dimensions from evidence sources for each of the answers using the question-answering module and may further include calculating a corresponding confidence value for each of the answers based on the numerical value of each evidence dimension using the question-answering module. The present invention may also include outputting the generated answers, the corresponding confidence values, and the numerical values of each evidence dimension for one or more selected answers using the input/output module.

Abstract: Systems and methods for generating responses to user input such as dialogues, and images are discussed. The system may generate, by a response generation module of at least one server, an optimal generated response to the user communication by applying an generative adversarial network. In some embodiments, the generative adversarial network may include a hierarchical recurrent encoder decoder generative adversarial network including a generator and a discriminator component.

Abstract: Memory augmented neural networks may use one or more neural encoders to transform input data into distributed representations and a memory module to store the representations with individual addresses. Memory augmented neural networks allow for few-shot learning capabilities because latent representations are persistent between training examples and gradient-based updates affect only certain memory locations via content-based lookups. When a query vector is not found in memory and the memory is full, existing memories that are positively associated with a particular representation may be identified, redundant memories may be aged, and updated memories may be generated. These updated memories retain relevant information acquired during training and reduce redundancy in the memories stored using the memory module, thereby improving the efficiency of data storage and reducing overfitting of data typically encountered with existing neural networks using memory modules.

Abstract: A system for using hash keys to preserve privacy across multiple tasks is disclosed. The system may provide training batch(es) of input observations each having a customer request and stored task to an encoder, and assign a hash key(s) to each of the stored tasks. The system may provide a new batch of input observations with a new customer request and new task to the encoder. The encoder may generate a new hash key assigned to the new customer request and determine whether any existing hash key corresponds with the new hash key. If so, the system may associate the new batch of input observations with the corresponding hash key and update the corresponding hash key such that it is also configured to provide access to the new batch of input observations. If not, the system may generate a new stored task and assign the new hash key to it.

Abstract: Aspects described herein may allow for the application of stochastic gradient boosting techniques to the training of deep neural networks by disallowing gradient back propagation from examples that are correctly classified by the neural network model while still keeping correctly classified examples in the gradient averaging. Removing the gradient contribution from correctly classified examples may regularize the deep neural network and prevent the model from overfitting. Further aspects described herein may provide for scheduled boosting during the training of the deep neural network model conditioned on a mini-batch accuracy and/or a number of training iterations. The model training process may start un-boosted, using maximum likelihood objectives or another first loss function.

Abstract: Memory augmented neural networks may use one or more neural encoders to transform input data into distributed representations and a memory module to store the representations with individual addresses. Memory augmented neural networks allow for few-shot learning capabilities because latent representations are persistent between training examples and gradient-based updates affect only certain memory locations via content-based lookups. When a query vector is not found in memory and the memory is full, existing memories that are positively associated with a particular representation may be identified, redundant memories may be aged, and updated memories may be generated. These updated memories retain relevant information acquired during training and reduce redundancy in the memories stored using the memory module, thereby improving the efficiency of data storage and reducing overfitting of data typically encountered with existing neural networks using memory modules.

Abstract: Systems and methods for generating responses to user input such as dialogues, and images are discussed. The system may generate, by a response generation module of at least one server, an optimal generated response to the user communication by applying an generative adversarial network. In some embodiments, the generative adversarial network may include a hierarchical recurrent encoder decoder generative adversarial network including a generator and a discriminator component.

Abstract: A method for determining machine learning training parameters is disclosed. The method can include a processor receiving a first input. The processor may receive a first response to the first input, determine a first intent, and identify a first action. The processor can then determine first trainable parameter(s) and determine whether the first trainable parameter(s) is negative or positive. Further, the processor can update a training algorithm based on the first trainable parameter(s). The processor can then receive a second input and determine a second intent for the second input. The processor can also determine a second action for the second intent and transmit the second action to a user. The processor can then determine second trainable parameter(s) and determine whether the second trainable parameter(s) is positive or negative. Finally, the processor can further update the training algorithm based on the second trainable parameter(s).

Abstract: Systems and methods are provided herein for autonomously determining and resolving a customer's perceived discrepancy during a customer service interaction. The method can include receiving an incoming communication from a customer; extracting a perceived state and an expected state (possibly of a product or service) based on the incoming communication; determining a perceived discrepancy between the perceived and expected states of the customer; retrieving customer information; extracting a current state of the customer from the retrieved customer information, verifying, by a rule-based platform, the discrepancy; generating a response based on the discrepancy after comparing the perceived stated with the current state, where the response may include a confirmation or a correction related to the discrepancy and a personalized explanation describing the current state of the customer; and outputting, for presentation to the customer, the response.

Abstract: Aspects described herein may allow for the application of stochastic gradient boosting techniques to the training of deep neural networks by disallowing gradient back propagation from examples that are correctly classified by the neural network model while still keeping correctly classified examples in the gradient averaging. Removing the gradient contribution from correctly classified examples may regularize the deep neural network and prevent the model from overfitting. Further aspects described herein may provide for scheduled boosting during the training of the deep neural network model conditioned on a mini-batch accuracy and/or a number of training iterations. The model training process may start un-boosted, using maximum likelihood objectives or another first loss function.