Multi-lingual semantic parser based on transferred learning

• US 10,460,036 B2
• Filed: 04/23/2018
• Issued: 10/29/2019
• Est. Priority Date: 04/23/2017
• Status: Active Grant

• Current Assignee: Cerence Inc.
• Original Assignee: VoiceBox Technologies Corporation (Microsoft Corporation)
• Inventors: Duong, Long, Afshar, Hadi, Estival, Dominique, Pink, Glen, Cohen, Philip, Johnson, Mark Edward
• Primary Examiner(s): Nguyen, Khai N.
• Application Number:
  US15/959,833
• Publication Number:
  US 20180307679A1
• Time in Patent Office: 554 Days
• Field of Search: 704 2, 704 8, 704 9, 704277
• CPC Class Codes: G06F 40/205 Parsing
• CPC Class Codes: G06F 40/242 Dictionaries
• CPC Class Codes: G06F 40/284 Lexical analysis, e.g. tokenisation or collocates
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 40/45 Example-based machine translation; Alignment
• CPC Class Codes: G06F 40/47 Machine-assisted translation, e.g. using translation memory
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/082 modifying the architecture, e.g. adding, deleting or silencing nodes or connections

Abstract

The disclosure relates to transferred learning from a first language (e.g., a source language for which a semantic parser has been defined) to a second language (e.g., a target language for which a semantic parser has not been defined). A system may use knowledge from a trained model in one language to model another language. For example, the system may transfer knowledge of a semantic parser from a first (e.g., source) language to a second (e.g., target) language. Such transfer of knowledge may occur and be useful when the first language has sufficient training data but the second language has insufficient training data. The foregoing transfer of knowledge may extend the semantic parser for multiple languages (e.g., the first language and the second language).

Claims

1. A system of generating a multi-lingual semantic parser based on transferred learning from a first language having sufficient training data to a second language having insufficient training data, the system comprising:

a computer system programmed to;

obtain a corpus of words in the first language;

generate a plurality of cross-lingual word features that each predicts a target word in the second language that corresponds to a respective source word in the corpus of words based on contextual information in the corpus of words and a translation of the respective source word from the first language to the second language;

obtain at least a first predefined utterance in the first language;

encode, using at least a first encoder, the first predefined utterance as a plurality of first vectors, wherein each first vector is based on a lookup of the plurality of cross-lingual word features using a corresponding word in the first predefined utterance;

decode, using a decoder, the plurality of first vectors to generate one or more words in the second language corresponding to the first predefined utterance in the first language; and

adapt a semantic parser trained on the first language based at least on the decoded plurality of first vectors to generate the multi-lingual semantic parser that parses utterances having one or more words in the second language and utterances having one or more words in the first language.

2. The system of claim 1, wherein to generate the plurality of cross-lingual word features, the computer system is programmed to:

determine a context of a first word of the corpus of words based on one or more words adjacent to the first word in the corpus of words.

3. The system of claim 1, wherein the computer system is further programmed to:

obtain at least the first predefined utterance in the second language, wherein the first predefined utterance in the first language and the first predefined utterance in the second language is obtained from a predefined training dataset;

encode, using at least a second encoder, the first predefined utterance in the second language as a plurality of second vectors, wherein each second vector is based on a lookup of the plurality of cross-lingual word features using a corresponding word in the second predefined utterance;

decode, using a decoder, the plurality of second vectors to generate one or more words in the first language corresponding to the first predefined utterance in the second language; and

wherein the semantic parser is further adapted based on the decoded plurality of second vectors.

4. The system of claim 3, wherein the computer system is further programmed to:

define and use a logical form having a single format for both the first predefined utterance in the first language and the first predefined utterance in the second language.

5. The system of claim 1, wherein the translation of the respective source word from the first language to the second language is via a dictionary or machine translation.

6. The system of claim 1, wherein to decode the plurality of first vectors, the computer system is programmed to:

for at least a first vector among the plurality of vectors, attend to different parts of the plurality of vectors to decode the first vector.

7. The system of claim 1, wherein the computer system is further programmed to:

initialize the plurality of cross-lingual word features with pre-trained word embeddings trained on a second corpus of words in the first language.

8. The system of claim 1, wherein the computer system is further programmed to:

identify a low frequency word in the corpus of words;

replace the low frequency word with a symbol; and

target the symbol during training for disambiguate the low frequency word.

9. The system of claim 1, wherein the computer system is further programmed to:

identify a low frequency word in the corpus of words;

replace the low frequency word with a symbol; and

obtain an identification of an entity associated with the low frequency word based on named entity recognition input.

10. The system of claim 1, wherein the computer system is further programmed to:

receive a natural language utterance comprising one or more words in the second language; and

use the multi-lingual semantic parser to parse the natural language utterance.

11. The system of claim 10, wherein the natural language utterance further comprises one or more words in the first language, and wherein the multi-lingual semantic parser to parse the one or more words in the first language natural language utterance.

12. The system of claim 11, wherein the computer system is further programmed to:

identify a predominant one of the first language or the second language in the natural language utterance; and

generate a response to the natural language utterance in the predominant one of the first language or the second language in the natural language utterance.

13. The system of claim 11, wherein the computer system is further programmed to:

identify a requested language output from the natural language utterance; and

generate a response to the natural language utterance in the requested language.

14. The system of claim 1, wherein the corpus of words relates to a single domain of information.

15. A computer-implemented method of generating a multi-lingual semantic parser based on transferred learning from a first language having sufficient training data to a second language having insufficient training data, the method being implemented on a computer system, the method comprising:

obtaining a corpus of words in the first language;

generating a plurality of cross-lingual word features that each predicts a target word in the second language that corresponds to a respective source word in the corpus of words based on contextual information in the corpus of words and a translation of the respective source word from the first language to the second language;

obtaining at least a first predefined utterance in the first language;

encoding, using at least a first encoder, the first predefined utterance as a plurality of first vectors, wherein each first vector is based on a lookup of the plurality of cross-lingual word features using a corresponding word in the first predefined utterance;

decoding, using a decoder, the plurality of first vectors to generate one or more words in the second language corresponding to the first predefined utterance in the first language; and

adapting a semantic parser trained on the first language based at least on the decoded plurality of first vectors to generate the multi-lingual semantic parser that parses utterances having one or more words in the second language and utterances having one or more words in the first language.

16. The computer-implemented method of claim 15, wherein generating the plurality of cross-lingual word features comprises:

determining a context of a first word of the corpus of words based on one or more words adjacent to the first word in the corpus of words.