Back to results
US 11,775,778 B2Granted
Machine translation of entities
AssigneeMicrosoft Technology Licensing, LLC
InventorsZhuliu Li, Xiao Yan, Yiming Wang, Jaewon Yang
Life of this patent
Estimating the term…
Abstract
Embodiments of the disclosed technologies incorporate taxonomy information into a cross-lingual entity graph and input the taxonomy-informed cross-lingual entity graph into a graph neural network. The graph neural network computes semantic alignment scores for node pairs. The semantic alignment scores are used to determine whether a node pair represents a valid machine translation.
Claims20
A method for performing machine translation, comprising:
- identifying an unlinked entity pair within a cross-lingual entity graph;
- the cross-lingual entity graph comprises a plurality of source entities, a plurality of target entities, and a plurality of links;
- at least two of the plurality of links comprise different link types;
- the unlinked entity pair comprises a source entity of the plurality of source entities and a target entity of the plurality of target entities;
- determining a sub-graph of the cross-lingual entity graph that includes the unlinked entity pair and at least one link of the plurality of links;
- inputting the sub-graph into a graph neural network (GNN) that has been trained using different link types;
- in response to inputting the sub-graph into the GNN, outputting, by the GNN, a value;
- determining that the value satisfies a criterion;
- creating and storing a link between the source entity and the target entity in the cross-lingual entity graph;
- the link indicates that the target entity is a valid translation of the source entity;
- and
- the cross-lingual entity graph is created by, before the sub-graph is input into the GNN, using taxonomy links to create a merged graph by merging a source language graph that does not comprise taxonomy links with a taxonomy graph and connecting the taxonomy graph with a target language graph.
9 more elements
A machine translation system, comprising:
- at least one processor;
- and
- computer memory operably coupled to the at least one processor;
- the computer memory comprises instructions that, when executed by the at least one processor, cause the at least one processor to:
- configure the computer memory according to a relational graph convolutional network (RGCN),
- the RGCN comprises a plurality of graph convolutional sub-networks trained using a cross-lingual entity graph;
- the cross-lingual entity graph comprises a first plurality of entities connected by a plurality of within language taxonomy links, a second plurality of entities connected by a plurality of within language embedding links, and a third plurality of entities connected by a plurality of between language links;
- the plurality of within language taxonomy links indicate at least two different link types obtained from a taxonomy;
- the plurality of within language embedding links indicate semantic similarities between connected entities;
- the plurality of between language links each indicate a valid machine translation;
- and
- before the cross-lingual entity graph is used to train the plurality of graph convolutional sub-networks, the cross-lingual entity graph is created by (i) merging a source language semantic graph that does not comprise taxonomy links with the taxonomy and (ii) connecting the taxonomy to a target language graph.
9 more elements
A method for performing machine translation, comprising:
- embedding a source entity in a cross-lingual entity graph stored in computer memory;
- identifying a target entity in the cross-lingual entity graph;
- the target entity comprises a valid translation of the source entity;
- the cross-lingual entity graph comprises a plurality of source entities, a plurality of target entities, a plurality of within language links that each indicate a different link type obtained from a taxonomy, and a plurality of between language links created using entity alignment scores computed by a graph convolutional network (GCN);
- and
- the cross-lingual entity graph is created by (i) merging a source language semantic graph that does not comprise taxonomy links with the taxonomy and (ii) connecting the taxonomy with a target language graph.
3 more elements
Dependencies are inferred from the claim text.