Salesforce, Inc.

Salesforce Com, Inc.

Nazneen Rajani

Wojciech Kryscinski

Prafulla Kumar Choubey

Wenpeng Yin

Caiming Xiong

Richard Socher

Wenhao Liu

Bryan McCann

Chien-Sheng Wu

Semih Yavuz

Tong Niu

Xi Victoria Lin

Jesse Vig

Alexander R Fabbri

Benjamin Newman

Jered McInerney

Qingyun Wang

Tanya Goyal

Embodiments described herein provide a document summarization framework that employs an ensemble of summarization models, each of which is a modified version of a base summarization model to control hallucination. For example, a base summarization model may first be trained on a full training data set. The trained base summarization model is then fine-tuned using a first filtered subset of the training data which contains noisy data, resulting in an 'anti-expert' model. The parameters of the anti-expert model are subtracted from the parameters of the trained base model to produce a final summarization model which yields robust factual performance.

Systems and methods for parameter ensembling for reducing hallucination in abstractive summarization

Embodiments described herein adopts a k nearest neighbor (kNN) mechanism over a model's hidden representations to identify training examples closest to a given test example. Specifically, a training set of sequences and a test sequence are received, each of which is mapped to a respective hidden representation vector using a base model. A set of indices for each sequence index that minimizes a distance between the respective hidden state vector and a test hidden state vector is then determined A weighted k-nearest neighbor probability score can then be computed from the set of indices to generate a probability distribution over labels for the test sequence.

Systems and methods for a k-nearest neighbor based mechanism of natural language processing models

Embodiments described herein provide a system and method for extracting factual information. The system transforms a query into a natural language prompt in a format of a query subject and a queried relation. The system encodes, via an embedding layer of a pre-trained language model, the natural language prompt into a first embedding. The system encodes, via the adapter model, the first embedding into a second embedding based on a probability that the second embedding returns the factual information when the second embedding is fed the first attention layer of the pre-trained language model. The system decodes, by the first attention layer of the pre-trained language mode, the second embedding into a response to the query. The system extracts the factual information from the decoded response to the query.

Systems and methods for factual extraction from language model

A method for using a neural network model for natural language processing (NLP) includes receiving training data associated with a source domain and a target domain; and generating one or more query batches. Each query batch includes one or more source tasks associated with the source domain and one or more target tasks associated with the target domain. For each query batch, class representations are generated for each class in the source domain and the target domain. A query batch loss for the query batch is generated based on the corresponding class representations. An optimization is performed on the neural network model by adjusting its network parameters based on the query batch loss. The optimized neural network model is used to perform one or more new NLP tasks.

System and method for natural language processing using neural network with cross-task training

The decoder network includes multiple decoders trained to generate different types of summaries. The lower layers of the multiple decoders are shared. The upper layers of the multiple decoders do not overlap. The multiple decoders generate probability distributions. A gating mechanism combines the probability distributions of the multiple decoders into a probability distribution of the decoder network. Words in the summary are selected based on the probability distribution of the decoder network.

Multi-attribute control for text summarization using multiple decoder heads

Embodiments described herein provide systems and methods for data-to-text generation. The embodiments receive input data that includes a resource description framework (RDF) triples in an RDF graph. A data-to-text generation system generates position aware embeddings, including position embeddings, triple role embeddings, and tree-level embeddings. Using the position aware embeddings and the RDF graph, the data-to-text generation system generates a textual description for the RDF graph.

Structured graph-to-text generation with two step fine-tuning

Embodiments described herein provide methods and systems for summarizing multiple documents. A system receives a plurality of documents and generates embeddings of the sentences from the plurality of documents. The embedded sentences are clustered in a representation space. Sentences from a reference summary are embedded and aligned with the closest cluster. Sentences from each cluster are summarized with the aligned reference sentences as a target. A loss is computed based on the summarized sentences and the aligned references, and the natural language processing model is updated based on the loss. Sentences may be masked from being used in the summarization by identifying sentences that are contradicted by other sentences within the plurality of documents.

Systems and methods for explainable and factual multi-document summarization

According to some embodiments, systems and methods are provided to develop or provide common sense auto-generated explanations (CAGE) for the reasoning used by an artificial intelligence, neural network, or deep learning model to make a prediction. In some embodiments, the systems and methods use supervised fine-tuning on a language model (LM) to generate such explanations. These explanations may then be used for downstream classification.

Growing community of inventors

Mountain View, CA, United States of America

Nazneen Rajani