Abstract: The technology disclosed provides a so-called “pointer sentinel mixture architecture” for neural network sequence models that has the ability to either reproduce a token from a recent context or produce a token from a predefined vocabulary. In one implementation, a pointer sentinel-LSTM architecture achieves state of the art language modeling performance of 70.9 perplexity on the Penn Treebank dataset, while using far fewer parameters than a standard softmax LSTM.

Abstract: Systems and methods are provided for efficient off-policy credit assignment (ECA) in reinforcement learning. ECA allows principled credit assignment for off-policy samples, and therefore improves sample efficiency and asymptotic performance. One aspect of ECA is to formulate the optimization of expected return as approximate inference, where policy is approximating a learned prior distribution, which leads to a principled way of utilizing off-policy samples. Other features are also provided.

Abstract: A method for providing a neural network system includes performing contrastive learning to the neural network system to generate a trained neural network system. The performing the contrastive learning includes performing first model augmentation to a first encoder of the neural network system to generate a first embedding of a sample, performing second model augmentation to the first encoder to generate a second embedding of the sample, and optimizing the first encoder using a contrastive loss based on the first embedding and the second embedding. The trained neural network system is provided to perform a task.

Abstract: Approaches for private and interpretable machine learning systems include a system for processing a query. The system includes one or more teacher modules for receiving a query and generating a respective output, one or more privacy sanitization modules for privacy sanitizing the respective output of each of the one or more teacher modules, and a student module for receiving a query and the privacy sanitized respective output of each of the one or more teacher modules and generating a result. Each of the one or more teacher modules is trained using a respective private data set. The student module is trained using a public data set. In some embodiments, human understandable interpretations of an output from the student module is provided to a model user.

Abstract: The disclosed technology reveals a hierarchical policy network, for use by a software agent, to accomplish an objective that requires execution of multiple tasks. A terminal policy learned by training the agent on a terminal task set, serves as a base task set of the intermediate task set. An intermediate policy learned by training the agent on an intermediate task set serves as a base policy of the top policy. A top policy learned by training the agent on a top task set serves as a base task set of the top task set. The agent is configurable to accomplish the objective by traversal of the hierarchical policy network. A current task in a current task set is executed by executing a previously-learned task selected from a corresponding base task set governed by a corresponding base policy, or performing a primitive action selected from a library of primitive actions.

Abstract: A machine learning based model generates a feature representation of a text sequence, for example, a natural language sentence or phrase. The system trains the machine learning based model by receiving an input text sequence and perturbing the input text sequence by masking a subset of tokens. The machine learning based model is used to predict the masked tokens. A predicted text sequence is generated based on the predictions of the masked tokens. The system processes the predicted text sequence using the machine learning based model to determine whether a token was predicted or an original token. The parameters of the machine learning based model are adjusted to minimize an aggregate loss based on prediction of the correct word for a masked token and a classification of a word as original or replaced.

Abstract: An online system allows user interactions using natural language expressions. The online system uses a machine learning based model to infer an intent represented by a user expression. The machine learning based model takes as input a user expression and an example expression to compute a score indicating whether the user expression matches the example expression. Based on the scores, the intent inference module determines a most applicable intent for the expression. The online system determines a confidence threshold such that user expressions indicating a high confidence are assigned the most applicable intent and user expressions indicating a low confidence are assigned an out-of-scope intent. The online system encodes the example expressions using the machine learning based model. The online system may compare an encoded user expression with encoded example expressions to identify a subset of example expressions used to determine the most applicable intent.

Abstract: Approaches for the translation of structured text include an embedding module for encoding and embedding source text in a first language, an encoder for encoding output of the embedding module, a decoder for iteratively decoding output of the encoder based on generated tokens in translated text from previous iterations, a beam module for constraining output of the decoder with respect to possible embedded tags to include in the translated text for a current iteration using a beam search, and a layer for selecting a token to be included in the translated text for the current iteration. The translated text is in a second language different from the first language. In some embodiments, the approach further includes scoring and pointer modules for selecting the token based on the output of the beam module or copied from the source text or reference text from a training pair best matching the source text.

Abstract: An embodiment proposed herein uses sparsification techniques to train the neural network with a high feature dimension that may yield desirable in-domain detection accuracy but may prune away dimensions in the output that are less important. Specifically, a sparsification vector is generated based on Gaussian distribution (or other probabilistic distribution) and is used to multiply with the higher dimension output to reduce the number of feature dimensions. The pruned output may be then used for the neural network to learn the sparsification vector. In this way, out-of-distribution detection accuracy can be improved.

Abstract: A system performs conversations with users using chatbots customized for performing a set of tasks. The system may be a multi-tenant system that allows customization of the chatbots for each tenant. The system processes sentences that may include negation or coreferences. The system determines a confidence score for an input sentence using an intent detection model, for example, a neural network. The system modifies the sentence to generate a modified sentence, for example, by removing a negation or by replacing a pronoun with an entity. The system generates a confidence score for the modified sentence using the intent detection model. The system determines the intent of the sentence based on the confidence scores of the sentence and the modified sentence. The system performs tasks based on the determined intent and performs conversations with users based on the tasks.

Abstract: A computing system uses neural networks to translate natural language queries to database queries. The computing system uses a plurality of machine learning based models, each machine learning model for generating a portion of the database query. The machine learning models use an input representation generated based on terms of the input natural language query, a set of columns of the database schema, and the vocabulary of a database query language, for example, structured query language SQL. The plurality of machine learning based models may include an aggregation classifier model for determining an aggregation operator in the database query, a result column predictor model for determining the result columns of the database query, and a condition clause predictor model for determining the condition clause of the database query. The condition clause predictor is based on reinforcement learning.

Abstract: Embodiments are directed to pre-training a transformer model using more parameters for sophisticated patterns (PSP++). The transformer model is divided into a held-out model and a main model. A forward pass and a backward pass are performed on the held-out model, where the forward pass determines self-attention hidden states of the held-out model and the backward pass determines loss of the held-out model. A forward pass on the main model is performed to determine a self-attention hidden states of the main model. The self-attention hidden states of the main model are concatenated with the self-attention hidden states of the held-out model. A backward pass is performed on the main model to determine a loss of the main model. The parameters of the held-out model are updated to reflect the loss of the held-out model and parameters of the main model are updated to reflect the loss of the main model.

Abstract: A system and corresponding method are provided for generating responses for a dialogue between a user and a computer. The system includes a memory storing information for a dialogue history and a knowledge base. An encoder may receive a new utterance from the user and generate a global memory pointer used for filtering the knowledge base information in the memory. A decoder may generate at least one local memory pointer and a sketch response for the new utterance. The sketch response includes at least one sketch tag to be replaced by knowledge base information from the memory. The system generates the dialogue computer response using the local memory pointer to select a word from the filtered knowledge base information to replace the at least one sketch tag in the sketch response.

Abstract: Embodiments described herein provide a dense hierarchical retrieval for open-domain question and answering for a corpus of documents using a document-level and passage-level dense retrieval model. Specifically, each document is viewed as a structural collection that has sections, subsections and paragraphs. Each document may be split into short length passages, where a document-level retrieval model and a passage-level retrieval model may be applied to return a smaller set of filtered texts. Top documents may be identified after encoding the question and the documents and determining document relevance scores to the encoded question. Thereafter, a set of top passages are further identified based on encoding of the passages and determining passage relevance scores to the encoded question. The document and passage relevance scores may be used in combination to determine a final retrieval ranking for the documents having the set of top passages.

Abstract: Some embodiments of the current disclosure disclose methods and systems for training for training a natural language processing intent classification model to perform few-shot classification tasks. In some embodiments, a pair of an utterance and a first semantic label labeling the utterance may be generated and a neural network that is configured to perform natural language inference tasks may be utilized to determine the existence of an entailment relationship between the utterance and the semantic label. The semantic label may be predicted as the intent class of the utterance based on the entailment relationship and the pair may be used to train the natural language processing intent classification model to perform few-shot classification tasks.

Abstract: Approaches for multitask learning as question answering include a method for training that includes receiving a plurality of training samples including training samples from a plurality of task types, presenting the training samples to a neural model to generate an answer, determining an error between the generated answer and the natural language ground truth answer for each training sample presented, and adjusting parameters of the neural model based on the error. Each of the training samples includes a natural language context, question, and ground truth answer. An order in which the training samples are presented to the neural model includes initially selecting the training samples according to a first training strategy and switching to selecting the training samples according to a second training strategy. In some embodiments the first training strategy is a sequential training strategy and the second training strategy is a joint training strategy.

Abstract: The system has a form analysis module that receives an image of a form into which values have been filled for the possible fields of information on the form, such as first name, address, age, and the like. By using a library of form templates, a form analysis module allows both flexibility of form processing and simplicity for the user. That is, the techniques used by the form analysis module allow the processing of any form image for which the library has a form template example. The form image need not precisely match any form template, but rather may be scaled or shifted relative to a corresponding template. Additionally, the user need only provide the form image itself, without providing any additional exemplars, metadata for training, or the like.

Abstract: Embodiments described herein provide a provide a fully unsupervised model for text compression. Specifically, the unsupervised model is configured to identify an optimal deletion path for each input sequence of texts (e.g., a sentence) and words from the input sequence are gradually deleted along the deletion path. To identify the optimal deletion path, the unsupervised model may adopt a pretrained bidirectional language model (BERT) to score each candidate deletion based on the average perplexity of the resulting sentence and performs a simple greedy look-ahead tree search to select the best deletion for each step.

Abstract: An embodiment provided herein preprocesses the input samples to the classification neural network, e.g., by adding Gaussian noise to word/sentence representations to make the function of the neural network satisfy Lipschitz property such that a small change in the input does not cause much change to the output if the input sample is in-distribution. Method to induce properties in the feature representation of neural network such that for out-of-distribution examples the feature representation magnitude is either close to zero or the feature representation is orthogonal to all class representations. Method to generate examples that are structurally similar to in-domain and semantically out-of domain for use in out-of-domain classification training. Method to prune feature representation dimension to mitigate long tail error of unused dimension in out-of-domain classification. Using these techniques, the accuracy of both in-domain and out-of-distribution identification can be improved.

Abstract: A system uses a neural network to detect anomalies in time series data. The system trains the neural network for a fixed number of iterations using data from a time window of the time series. The system uses the loss value at the end of the fixed number of iterations for identifying anomalies in the time series data. For a time window, the system initializes the neural network to random values and trains the neural network for a fixed number of iterations using the data of the time window. After the fixed number of iterations, the system compares the loss values for various data points to a threshold value. Data points having loss value exceeding a threshold are identified as anomalous data points.