Back to results
US 11,893,352 B2Granted
Dependency path reasoning for measurement extraction
AssigneeAdobe Inc.
InventorsAmir Pouran Ben Veyseh, Franck Dernoncourt
Life of this patent
Estimating the term…
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.
Claims20
A method comprising:
- generating a word representation for each word in an input phrase, wherein the input phrase includes a subject entity, an object entity, a candidate word, and one or more additional words;
- generating a path vector representing a transformation between the subject entity and the object entity using a path network, wherein the path vector is generated based on the word representation for the subject entity and the object entity, respectively;
- computing an exclusion vector corresponding to the candidate word, wherein the exclusion vector is based on the one or more additional words from the input phrase excluding the candidate word, the subject entity, and the object entity;
- computing a transfer vector corresponding to the candidate word based on the path vector and the exclusion vector;
- including the candidate word in a dependency path between the subject entity and the object entity based on a similarity between the transfer vector and the word representation for the candidate word;
- computing a path representation based on the dependency path;
- and
- identifying a relationship between the subject entity and the object entity based on the path representation.
5 more elements
An apparatus comprising:
- an input encoder configured to generate a word representation for each word of an input phrase, wherein the input phrase includes a subject entity, an object entity, a candidate word, and one or more additional words;
- a path network configured to
- generate a path vector representing a transformation between the subject entity and the object entity based on the word representation for the subject entity and the object entity, respectively,
- compute an exclusion vector corresponding to the candidate word, wherein the exclusion vector is based the on one or more additional words from the input phrase excluding the candidate word, the subject entity, and the object entity,
- compute a transfer vector corresponding to the candidate word based on the path vector and the exclusion vector,
- including the candidate word in a dependency path between the subject entity and the object entity based on a similarity between the transfer vector and the word representation for the candidate word, and
- compute a path representation based on the dependency path;
- and
- a decoder configured to identify a relationship between the subject entity and the object entity.
6 more elements
A method of training a neural network, the method comprising:
- receiving a training phrase including a subject entity, an object entity, a candidate word, one or more additional words, and a ground truth relationship between the subject entity and the object entity;
- generating a path vector representing a transformation between the subject entity and the object entity using a path network, wherein the path vector is generated based on a word representation for the subject entity and the object entity, respectively;
- computing an exclusion vector corresponding to the candidate word, wherein the exclusion vector is based on the one or more additional words from the training phrase excluding the candidate word, the subject entity, and the object entity;
- computing a transfer vector corresponding to the candidate word based on the path vector and the exclusion vector;
- including the candidate word in a dependency path between the subject entity and the object entity based on a similarity between the transfer vector and the word representation for the candidate word;
- generating a path representation based on the dependency path;
- predicting a relationship between the subject entity and the object entity based on the path representation using a decoder;
- comparing the ground truth relationship with the predicted relationship to produce a prediction loss;
- and
- training the path network based on the prediction loss.
7 more elements
Dependencies are inferred from the claim text.