Machine learning for translation to structured computer readable representation

Inventors

Shankar, Natarajan • Graham-Lengrand, Stephane • Elenius, Daniel • Yeh, Chih-Hung

Assignees

SRI International Inc

Abstract

In general, the disclosure describes techniques for machine learning for translation to structured computer readable representation. An example method to generate a training set for a natural language translation model includes receiving, by a computing system, a grammar comprising rules, one or more of the rules being associated with random biases; generating, by the computing system, at least one of random trees or random graphs based on the random biases in the grammar; for each of the random trees or random graphs, by the computing system, generating a natural language sample; and generating, by the computing system, the training set with the random trees or random graphs and the corresponding natural language samples.

Core Innovation

The invention relates to machine learning-based translation of natural-language standards text into domain-specific symbolic structured representations using a grammar and probabilistic generation of random trees and/or random graphs. A grammar comprises rules, and one or more rules are associated with respective random biases, where each random bias specifies a probability that a corresponding node for a corresponding rule is included when generating a random tree or a random graph.

Based on the random biases, at least one random tree or random graph is generated, and for each generated random tree or random graph, a corresponding natural language sample is generated. A training set is formed comprising training pairs, each training pair including one of the generated random trees or random graphs and the corresponding natural language sample, and the machine learning model is trained using the training set.

When applying the trained machine learning model, the system receives a natural language statement and applies the model to generate a translation tree. The translation tree maps the natural language statement to a symbolic language output, where the output is a structured statement defined by the structures produced from the grammar and supports translation into symbolic language outputs.

Claims Coverage

The document provides three independent claims, centered on generating training pairs from grammar-defined random trees or random graphs driven by random biases, and using a trained machine learning model to output a translation tree mapping a natural language statement to symbolic language output. The inventive features are the grammar with random biases for node inclusion, probabilistic random tree or random graph generation, generation of corresponding natural language samples, formation of training pairs, and applying the trained model to produce a translation tree for symbolic language output.

The core claim coverage is for training and using a machine learning model where training pairs are generated from grammar-defined random trees or random graphs produced using random biases that specify probabilities for node inclusion, and where applying the trained model outputs a translation tree that maps a natural language statement to symbolic language output.

Stated Advantages

Documented Applications