Linguistically rich cross-lingual text event embeddings

Patent metadata

Field	Value
Publication number	US-11227128-B2
Application number	US-201916434710-A
Country	US
Kind code	B2
Filing date	Jun 7, 2019
Priority date	Jun 7, 2019
Publication date	Jan 18, 2022
Grant date	Jan 18, 2022

Abstract

A machine accesses a preexisting set of natural language text documents in multiple natural languages. Each natural language text document in at least a portion of the preexisting set is associated with an event. The machine trains, using the preexisting set of natural language text documents and the associated events, an event encoder to learn associations between texts and event annotations. The event encoder leverages a parser in each of the two or more natural languages. The machine generates, using the event encoder, new event annotations for texts. The machine trains, using the preexisting set of natural language text documents and the new event annotations for the texts generated by the event encoder, an event extraction engine to extract events from natural language texts in the two or more natural languages. The event extraction engine leverages the parser in each of the two or more natural languages.

First claim

What is claimed is:

1. An event extraction training apparatus, the apparatus comprising:
processing circuitry and memory; the processing circuitry to:
access a preexisting set of natural language text documents in two or more natural languages, wherein each natural language text document in at least a portion of the preexisting set is associated with an event;
train, using the preexisting set of natural language text documents and the associated events, an event encoder to learn associations between texts and event annotations, wherein the event encoder leverages a parser in each of the two or more natural languages, wherein the preexisting set of natural language text documents comprises more than a training adequacy threshold number of texts annotated with a first event type in a first natural language and fewer than a training inadequacy threshold number of texts annotated with the first event type in a second natural language, wherein the training inadequacy threshold number is less than the training adequacy threshold number;
generate, using the event encoder, new event annotations for texts;
train, using the preexisting set of natural language text documents and the new event annotations for the texts generated by the event encoder, an event extraction engine to extract events from natural language texts in the two or more natural languages, wherein the event extraction engine leverages the parser in each of the two or more natural languages; and
provide an output representing the trained event extraction engine, wherein the trained event extraction engine is trained to extract events of the first event type from texts in the first natural language and texts in the second natural language.

2. The event extraction training apparatus of claim 1 , wherein each event comprises one or more trigger words and one or more arguments.

3. The event extraction training apparatus of claim 2 , wherein the one or more arguments comprise one or more of: an agent/subject of the event, a patient/object of the event, a geographic location of the event, a time of the event, and an artifact of the event.

4. The event extraction training apparatus of claim 2 , wherein the one or more trigger words comprise one or more verbs representing an action of the event.

5. The event extraction training apparatus of claim 2 , wherein each event is represented as a numeric vector representing the one or more trigger words and the one or more arguments.

6. The event extraction training apparatus of claim 1 , wherein the processing circuitry is further to:
receive a new natural language text in one of the two or more natural languages;
identify, using the event extraction engine, a new event in the new natural language text; and
provide an output representing the new event.

7. The event extraction training apparatus of claim 1 , wherein the parser comprises one or more of: a grammatical parser and a semantic parser.

8. An event extraction inferencing, apparatus, the apparatus comprising:
processing circuitry and memory; the processing circuitry to:
receive a new natural language text;
identify, using an event extraction engine, a new event in the new natural language text; and
provide an output representing the new event, wherein the event extraction engine is trained by:
accessing, at a training apparatus, a preexisting set of natural language text documents in two or more natural languages, wherein each natural language text document in at least a portion of the preexisting set is associated with an event, and wherein the new natural language text is in one of the two or more natural languages;
training, using the preexisting set of natural language text documents and the associated events, an event encoder to learn associations between texts and event annotations, wherein the event encoder leverages a parser in each of the two or more natural languages, wherein the preexisting set of natural language text documents comprises more than a training adequacy threshold number of texts annotated with a first event type in a first natural language and fewer than a training inadequacy threshold number of texts annotated with the first event type in a second natural language, wherein the training inadequacy threshold number is less than the training adequacy threshold number;
generating, using the event encoder, new event annotations for texts; and
training, using the preexisting set of natural language text documents and the new event annotations for the texts generated by the event encoder, the event extraction engine to extract events from natural language texts in the two or more natural languages, wherein the event extraction engine leverages the parser in each of the two or more natural languages, wherein the trained event extraction engine is trained to extract events of the first event type from texts in the first natural language and texts in the second natural language.

9. The event extraction inferencing apparatus of claim 8 , wherein each event comprises one or more trigger words and one or more arguments.

10. The event extraction inferencing apparatus of claim 8 , wherein the parser comprises one or more of: a grammatical parser and a semantic parser.

11. A non-transitory machine-readable medium storing instructions which, when executed by processing circuitry of one or more machines, cause the processing circuitry to:
access a preexisting set of natural language text documents in two or more natural languages, wherein each natural language text document in at least a portion of the preexisting set is associated with an event;
train, using the preexisting set of natural language text documents and the associated events, an event encoder to learn associations between texts and event annotations, wherein the event encoder leverages a parser in each of the two or more natural languages wherein the preexisting set of natural language text documents comprises more than a training adequacy threshold number of texts annotated with a first event type in a first natural language and fewer than a training inadequacy threshold number of texts annotated with the first event type in a second natural language, wherein the training inadequacy threshold number is less than the training adequacy threshold number;
generate; using the event encoder; new event annotations for texts;
train, using the preexisting set of natural language text documents and the new event annotations for the texts generated by the event encoder, an event extraction engine to extract events from natural language texts in the two or more natural languages, wherein the event extraction engine leverages the parser in each of the two or more natural languages; and
provide an output representing the trained event extraction engine, wherein the trained event extraction engine is trained to extract events of the first event type from texts in the first natural language and texts in the second natural language.

12. The machine-readable medium of claim 11 , wherein each event comprises one or more trigger words and one or more arguments.

13. The machine-readable medium of claim 12 , wherein the one or more arguments comprise one or more of: an agent/subject of the event, a patient/object of the event, a geographic location of the event, a time of the event, and an artifact of the event.

14. The machine-readable medium of claim 12 , wherein the one or more trigger words comprise one or more verbs representing an action of the event.

15. The machine-readable medium of claim 12 , wherein each event is represented as a numeric vector representing the one or more trigger words an

Assignees

• Raytheon Bbn Technologies Corp

Inventors

• Min Bonan
• Chan Yee Seng
• Heintz Ilana

Classifications

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title