Abstract: The technology disclosed provides a quasi-recurrent neural network (QRNN) encoder-decoder model that alternates convolutional layers, which apply in parallel across timesteps, and minimalist recurrent pooling layers that apply in parallel across feature dimensions.

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: The technology disclosed relates to an end-to-end neural network for question answering, referred to herein as “dynamic coattention network (DCN)”. Roughly described, the DCN includes an encoder neural network and a coattentive encoder that capture the interactions between a question and a document in a so-called “coattention encoding”. The DCN also includes a decoder neural network and highway maxout networks that process the coattention encoding to estimate start and end positions of a phrase in the document that responds to the question.

Abstract: Systems and methods are provided for near-zero-cost (NZC) query framework or approach for differentially private deep learning. To protect the privacy of training data during learning, the near-zero-cost query framework transfers knowledge from an ensemble of teacher models trained on partitions of the data to a student model. Privacy guarantees may be understood intuitively and expressed rigorously in terms of differential privacy. Other features are also provided.

Abstract: Systems and methods for dense captioning of a video include a multi-layer encoder stack configured to receive information extracted from a plurality of video frames, a proposal decoder coupled to the encoder stack and configured to receive one or more outputs from the encoder stack, a masking unit configured to mask the one or more outputs from the encoder stack according to one or more outputs from the proposal decoder, and a decoder stack coupled to the masking unit and configured to receive the masked one or more outputs from the encoder stack. Generating the dense captioning based on one or more outputs of the decoder stack. In some embodiments, the one or more outputs from the proposal decoder include a differentiable mask. In some embodiments, during training, error in the dense captioning is back propagated to the decoder stack, the encoder stack, and the proposal decoder.

Abstract: A system is provided for natural language processing. In some embodiments, the system includes an encoder for generating context-specific word vectors for at least one input sequence of words. The encoder is pre-trained using training data for performing a first natural language processing task. A neural network performs a second natural language processing task on the at least one input sequence of words using the context-specific word vectors. The first natural language process task is different from the second natural language processing task and the neural network is separately trained from the encoder.

Abstract: A method for maintaining a dialogue state associated with a dialogue between a user and a digital system includes receiving, by a dialogue state tracker associated with the digital system, a representation of a user communication, updating, by the dialogue state tracker, the dialogue state and providing a system response based on the updated dialogue state. The dialogue state is updated by evaluating, based on the representation of the user communication, a plurality of member scores corresponding to a plurality of ontology members of an ontology set, and selecting, based on the plurality of member scores, zero or more of the plurality of ontology members to add to or remove from the dialogue state.

Abstract: The technology disclosed provides a so-called “joint many-task neural network model” to solve a variety of increasingly complex natural language processing (NLP) tasks using growing depth of layers in a single end-to-end model. The model is successively trained by considering linguistic hierarchies, directly connecting word representations to all model layers, explicitly using predictions in lower tasks, and applying a so-called “successive regularization” technique to prevent catastrophic forgetting. Three examples of lower level model layers are part-of-speech (POS) tagging layer, chunking layer, and dependency parsing layer. Two examples of higher level model layers are semantic relatedness layer and textual entailment layer. The model achieves the state-of-the-art results on chunking, dependency parsing, semantic relatedness and textual entailment.

Abstract: A system is disclosed for providing an abstractive summary of a source textual document. The system includes an encoder, a decoder, and a fusion layer. The encoder is capable of generating an encoding for the source textual document. The decoder is separated into a contextual model and a language model. The contextual model is capable of extracting words from the source textual document using the encoding. The language model is capable of generating vectors paraphrasing the source textual document based on pre-training with a training dataset. The fusion layer is capable of generating the abstractive summary of the source textual document from the extracted words and the generated vectors for paraphrasing. In some embodiments, the system utilizes a novelty metric to encourage the generation of novel phrases for inclusion in the abstractive summary.

Abstract: Embodiments described herein provide a two-stage online detection of action start system including a classification module and a localization module. The classification module generates a set of action scores corresponding to a first video frame from the video, based on the first video frame and video frames before the first video frames in the video. Each action score indicating a respective probability that the first video frame contains a respective action class. The localization module is coupled to the classification module for receiving the set of action scores from the classification module and generating an action-agnostic start probability that the first video frame contains an action start.

Abstract: Approaches for multitask learning as question answering include an input layer for encoding a context and a question, a self-attention based transformer including an encoder and a decoder, a first bi-directional long-term short-term memory (biLSTM) for further encoding an output of the encoder, a long-term short-term memory (LSTM) for generating a context-adjusted hidden state from the output of the decoder and a hidden state, an attention network for generating first attention weights based on an output of the first biLSTM and an output of the LSTM, a vocabulary layer for generating a distribution over a vocabulary, a context layer for generating a distribution over the context, and a switch for generating a weighting between the distributions over the vocabulary and the context, generating a composite distribution based on the weighting, and selecting a word of an answer using the composite distribution.

Abstract: Verification of discriminative models includes receiving an input; receiving a prediction from a discriminative model for the input; encoding, using an encoder, a latent variable based on the input; decoding, using a decoder, a reconstructed input based on the prediction and the latent variable; and determining, using an anomaly detection module, whether the prediction is reliable based on the input, the reconstructed input, and the latent variable. The encoder and the decoder are jointly trained to maximize an evidence lower bound of the encoder and the decoder. In some embodiments, the encoder and the decoder are further trained using a disentanglement constraint between the prediction and the latent variable. In some embodiments, the encoder and the decoder are further trained without using inputs that are out of a distribution of inputs used to train the discriminative model or that are adversarial to the discriminative model.

Abstract: Embodiments described herein provide a pipelined natural language question answering system that improves a BERT-based system. Specifically, the natural language question answering system uses a pipeline of neural networks each trained to perform a particular task. The context selection network identifies premium context from context for the question. The question type network identifies the natural language question as a yes, no, or span question and a yes or no answer to the natural language question when the question is a yes or no question. The span extraction model determines an answer span to the natural language question when the question is a span question.

Abstract: Systems and methods are provided for weakly supervised natural language localization (WSNLL), for example, as implemented in a neural network or model. The WSNLL network is trained with long, untrimmed videos, i.e., videos that have not been temporally segmented or annotated. The WSNLL network or model defines or generates a video-sentence pair, which corresponds to a pairing of an untrimmed video with an input text sentence. According to some embodiments, the WSNLL network or model is implemented with a two-branch architecture, where one branch performs segment sentence alignment and the other one conducts segment selection.

Abstract: A method for evaluating robustness of one or more target neural network models using natural typos. The method includes receiving one or more natural typo generation rules associated with a first task associated with a first input document type, receiving a first target neural network model, and receiving a first document and corresponding its ground truth labels. The method further includes generating one or more natural typos for the first document based on the one or more natural typo generation rules, and providing, to the first target neural network model, a test document generated based on the first document and the one or more natural typos as an input document to generate a first output. A robustness evaluation result of the first target neural network model is generated based on a comparison between the output and the ground truth labels.

Abstract: The technology disclosed presents a novel spatial attention model that uses current hidden state information of a decoder long short-term memory (LSTM) to guide attention and to extract spatial image features for use in image captioning. The technology disclosed also presents a novel adaptive attention model for image captioning that mixes visual information from a convolutional neural network (CNN) and linguistic information from an LSTM. At each timestep, the adaptive attention model automatically decides how heavily to rely on the image, as opposed to the linguistic model, to emit the next caption word. The technology disclosed further adds a new auxiliary sentinel gate to an LSTM architecture and produces a sentinel LSTM (Sn-LSTM). The sentinel gate produces a visual sentinel at each timestep, which is an additional representation, derived from the LSTM's memory, of long and short term visual and linguistic information.

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: The technology disclosed provides a so-called “joint many-task neural network model” to solve a variety of increasingly complex natural language processing (NLP) tasks using growing depth of layers in a single end-to-end model. The model is successively trained by considering linguistic hierarchies, directly connecting word representations to all model layers, explicitly using predictions in lower tasks, and applying a so-called “successive regularization” technique to prevent catastrophic forgetting. Three examples of lower level model layers are part-of-speech (POS) tagging layer, chunking layer, and dependency parsing layer. Two examples of higher level model layers are semantic relatedness layer and textual entailment layer. The model achieves the state-of-the-art results on chunking, dependency parsing, semantic relatedness and textual entailment.

Abstract: A system is provided for natural language processing. In some embodiments, the system includes an encoder for generating context-specific word vectors for at least one input sequence of words. The encoder is pre-trained using training data for performing a first natural language processing task. A neural network performs a second natural language processing task on the at least one input sequence of words using the context-specific word vectors. The first natural language process task is different from the second natural language processing task and the neural network is separately trained from the encoder. In some embodiments, the first natural processing task can be machine translation, and the second natural processing task can be one of sentiment analysis, question classification, entailment classification, and question answering.

Abstract: Systems and methods for unifying question answering and text classification via span extraction include a preprocessor for preparing a source text and an auxiliary text based on a task type of a natural language processing task, an encoder for receiving the source text and the auxiliary text from the preprocessor and generating an encoded representation of a combination of the source text and the auxiliary text, and a span-extractive decoder for receiving the encoded representation and identifying a span of text within the source text that is a result of the NLP task. The task type is one of entailment, classification, or regression. In some embodiments, the source text includes one or more of text received as input when the task type is entailment, a list of classifications when the task type is entailment or classification, or a list of similarity options when the task type is regression.