Abstract: In implementations of systems for training language models and preserving privacy, a computing device implements a privacy system to predict a next word after a last word in a sequence of words by processing input data using a machine learning model trained on training data to predict next words after last words in sequences of words. The training data describes a corpus of text associated with clients and including sensitive samples and non-sensitive samples. The machine learning model is trained by sampling a client of the clients and using a subset of the sensitive samples associated with the client and a subset of the non-sensitive samples associated with the client to update parameters of the machine learning model. The privacy system generates an indication of the next word after the last word in the sequence of words for display in a user interface.

Abstract: One aspect of systems and methods for data correction includes identifying a false label from among predicted labels corresponding to different parts of an input sample, wherein the predicted labels are generated by a neural network trained based on a training set comprising training samples and training labels corresponding to parts of the training samples; computing an influence of each of the training labels on the false label by approximating a change in a conditional loss for the neural network corresponding to each of the training labels; identifying a part of a training sample of the training samples and a corresponding source label from among the training labels based on the computed influence; and modifying the training set based on the identified part of the training sample and the corresponding source label to obtain a corrected training set.

Abstract: The present disclosure describes a model for large scale color prediction of objects identified in images. Embodiments of the present disclosure include an object detection network, an attention network, and a color classification network. The object detection network generates object features for an object in an image and may include a convolutional neural network (CNN), region proposal network, or a ResNet. The attention network generates an attention vector for the object based on the object features, wherein the attention network takes a query vector based on the object features, and a plurality of key vector and a plurality of value vectors corresponding to a plurality of colors as input. The color classification network generates a color attribute vector based on the attention vector, wherein the color attribute vector indicates a probability of the object including each of the plurality of colors.

Abstract: Systems and methods for natural language processing are described. Embodiments of the inventive concept are configured to receive an input sequence and a plurality of candidate long forms for a short form contained in the input sequence, encode the input sequence to produce an input sequence representation, encode each of the plurality of candidate long forms to produce a plurality of candidate long form representations, wherein each of the candidate long form representations is based on a plurality of sample expressions and each of the sample expressions includes a candidate long form and contextual information, compute a plurality of similarity scores based on the candidate long form representations and the input sequence representation, and select a long form for the short form based on the plurality of similarity scores.

Abstract: The present disclosure describes systems and methods for dialog processing and information retrieval. Embodiments of the present disclosure provide a dialog system (e.g., a task-oriented dialog system) with adaptive recurrent hopping and dual context encoding to receive and understand a natural language query from a user, manage dialog based on natural language conversation, and generate natural language responses. For example, a memory network can employ a memory recurrent neural net layer and a decision meta network (e.g., a subnet) to determine an adaptive number of memory hops for obtaining readouts from a knowledge base. Further, in some embodiments, a memory network uses a dual context encoder to encode information from original context and canonical context using parallel encoding layers.

Abstract: The present disclosure provides systems and methods for relationship extraction. Embodiments of the present disclosure provide a relationship extraction network trained to identify relationships among entities in an input text. The relationship extraction network is used to generate a dependency path between entities in an input phrase. The dependency path includes a set of words that connect the entities, and is used to predict a relationship between the entities. In some cases, the dependency path is related to a syntax tree, but it may include additional words, and omit some words from a path extracted based on a syntax tree.

Abstract: Systems and methods for natural language processing are described. One or more embodiments of the present disclosure receive a document comprising a plurality of words organized into a plurality of sentences, the words comprising an event trigger word and an argument candidate word, generate word representation vectors for the words, generate a plurality of document structures including a semantic structure for the document based on the word representation vectors, a syntax structure representing dependency relationships between the words, and a discourse structure representing discourse information of the document based on the plurality of sentences, generate a relationship representation vector based on the document structures, and predict a relationship between the event trigger word and the argument candidate word based on the relationship representation vector.

Abstract: Systems and methods for color prediction are described. Embodiments of the present disclosure receive an image that includes an object including a color, generate a color vector based on the image using a color classification network, where the color vector includes a color value corresponding to each of a set of colors, generate a bias vector by comparing the color vector to teach of a set of center vectors, where each of the set of center vectors corresponds to a color of the set of colors, and generate an unbiased color vector based on the color vector and the bias vector, where the unbiased color vector indicates the color of the object.

Abstract: Systems and methods for natural language processing are described. One or more embodiments of the present disclosure generate a word embedding for each word of an input phrase, wherein the input phrase indicates a sentiment toward an aspect term, compute a gate vector based on the aspect term, identify a dependency tree representing relations between words of the input phrase, generate a representation vector based on the dependency tree and the word embedding using a graph convolution network, wherein the gate vector is applied to a layer of the graph convolution network, and generate a probability distribution over a plurality of sentiments based on the representation vector.

Abstract: Systems and methods for intent discovery and video summarization are described. Embodiments of the present disclosure receive a video and a transcript of the video, encode the video to obtain a sequence of video encodings, encode the transcript to obtain a sequence of text encodings, apply a visual gate to the sequence of text encodings based on the sequence of video encodings to obtain gated text encodings, and generate an intent label for the transcript based on the gated text encodings.

Abstract: Multitask machine-learning model training and training data augmentation techniques are described. In one example, training is performed for multiple tasks simultaneously as part of training a multitask machine-learning model using question pairs. Examples of the multiple tasks include question summarization and recognizing question entailment. Further, a loss function is described that incorporates a parameter sharing loss that is configured to adjust an amount that parameters are shared between corresponding layers trained for the first and second tasks, respectively. In an implementation, training data augmentation techniques are also employed by synthesizing question pairs, automatically and without user intervention, to improve accuracy in model training.

Abstract: Systems and methods for text processing are described. Embodiments of the present disclosure receive a query comprising a natural language expression; extract a plurality of mentions from the query; generate a relation vector between a pair of the plurality of mentions using a relation encoder network, wherein the relation encoder network is trained using a contrastive learning process where mention pairs from a same document are labeled as positive samples and mention pairs from different documents are labeled as negative samples; combine the plurality of mentions with the relation vector to obtain a virtual knowledge graph of the query; identify a document corresponding to the query by comparing the virtual knowledge graph of the query to a virtual knowledge graph of the document; and transmit a response to the query, wherein the response includes a reference to the document.

Abstract: Certain embodiments involve using a machine-learning tool to generate metadata identifying segments and topics for text within a document. For instance, in some embodiments, a text processing system obtains input text and applies a segmentation-and-labeling model to the input text. The segmentation-and-labeling model is trained to generate a predicted segment for the input text using a segmentation network. The segmentation-and-labeling model is also trained to generate a topic for the predicted segment using a pooling network of the model to the predicted segment. The output of the model is usable for generating metadata identifying the predicted segment and the associated topic.

Abstract: Systems and methods for video segmentation and summarization are described. Embodiments of the present disclosure receive a video and a transcript of the video; generate visual features representing frames of the video using an image encoder; generate language features representing the transcript using a text encoder, wherein the image encoder and the text encoder are trained based on a correlation between training visual features and training language features; and segment the video into a plurality of video segments based on the visual features and the language features.

Abstract: Systems and methods for natural language processing are described. Embodiments of the present disclosure receive text including an event trigger word indicating an occurrence of an event; classify the event trigger word to obtain an event type using a few-shot classification network, wherein the few-shot classification network is trained by storing first labeled samples during a first training iteration and using the first labeled samples for computing a loss function during a second training iteration that includes a support set with second labeled samples having a same ground-truth label as the first labeled samples; and transmit event detection information including the event trigger word and the event type.

Abstract: The present disclosure relates to systems, methods, and non-transitory computer readable media for accurately and flexibly generating topic divergence classifications for digital videos based on words from the digital videos and further based on a digital text corpus representing a target topic. Particularly, the disclosed systems utilize a topic-specific knowledge encoder neural network to generate a topic divergence classification for a digital video to indicate whether or not the digital video diverges from a target topic. In some embodiments, the disclosed systems determine topic divergence classifications contemporaneously in real time for livestream digital videos or for stored digital videos (e.g., digital video tutorials).

Abstract: Systems and methods for natural language processing are described. One or more embodiments of the disclosure provide an entity matching apparatus trained using machine learning techniques to determine whether a query name corresponds to a candidate name based on a similarity score. In some examples, the query name and the candidate name are encoded using a character encoder to produce a regularized input sequence and a regularized candidate sequence, respectively. The regularized input sequence and the regularized candidate sequence are formed from a regularized character set having fewer characters than a natural language character set.

Abstract: Systems, methods, and non-transitory computer-readable media can generate a natural language model that provides user-entity differential privacy. For example, in one or more embodiments, a system samples sensitive data points from a natural language dataset. Using the sampled sensitive data points, the system determines gradient values corresponding to the natural language model. Further, the system generates noise for the natural language model. The system generates parameters for the natural language model using the gradient values and the noise, facilitating simultaneous protection of the users and sensitive entities associated with the natural language dataset. In some implementations, the system generates the natural language model through an iterative process (e.g., by iteratively modifying the parameters).

Abstract: Certain embodiments involve using a machine-learning tool to generate metadata identifying segments and topics for text within a document. For instance, in some embodiments, a text processing system obtains input text and applies a segmentation-and-labeling model to the input text. The segmentation-and-labeling model is trained to generate a predicted segment for the input text using a segmentation network. The segmentation-and-labeling model is also trained to generate a topic for the predicted segment using a pooling network of the model to the predicted segment. The output of the model is usable for generating metadata identifying the predicted segment and the associated topic.

Abstract: The present disclosure describes systems and methods for information retrieval. Embodiments of the disclosure provide a retrieval network that leverages external knowledge to provide reformulated search query suggestions, enabling more efficient network searching and information retrieval. For example, a search query from a user (e.g., a query mention of a knowledge graph entity that is included in a search query from a user) may be added to a knowledge graph as a surrogate entity via entity linking. Embedding techniques are then invoked on the updated knowledge graph (e.g., the knowledge graph that includes additional edges between surrogate entities and other entities of the original knowledge graph), and entities neighboring the surrogate entity are retrieved based on the embedding (e.g., based on a computed distance between the surrogate entity and candidate entities in the embedding space). Search results can then be ranked and displayed based on relevance to the neighboring entity.