Goal-oriented conversation with code-mixed language

Patent metadata

Field	Value
Publication number	US-11769004-B2
Application number	US-202016732708-A
Country	US
Kind code	B2
Filing date	Jan 2, 2020
Priority date	Jan 2, 2020
Publication date	Sep 26, 2023
Grant date	Sep 26, 2023

Abstract

A computer system may create a language model corpus including multilingual alignment for training a combined language model and train (or pre-train) the combined language model. The computer system may create an adverse medication reaction corpus to include adverse medication reaction utterances and label an N-gram of an utterance in the adverse medication reaction utterances as a response to query, for multiple N-grams. The computer system may generate a code-mixed utterance model to perform code-mixed utterances in a turn by turn dialogue, by at least adding additional output layer including at least a start vector, language vector, and a query vector including at least the labeled N-gram, which are additional to the combined language model's predicted next words.

First claim

What is claimed is:

1. A computer-implemented system comprising:
a hardware processor;
a memory device coupled with the hardware processor:
the hardware processor configured to at least:
create a language model corpus including multilingual alignment for training a combined language model, the language model corpus created from data sourced from a computer network and train the combined language model;
create an adverse medication reaction corpus by analyzing data from online sources which include adverse medication reaction utterances and labeling an N-gram of an utterance in the adverse medication reaction utterances as a response to query;
and generate a code-mixed utterance model to perform code-mixed utterances in a turn by turn dialogue, by at least adding additional output layer including at least a start vector, language vector, and a query vector including at least the labeled N-gram, which are additional to the combined language model's predicted next words, wherein the start vector encodes what word should start a response to an input utterance, and the language vector encodes what language out of multiple languages used in the code-mixed utterances should a next word of the response be in, wherein the code-mixed utterances represent utterances performed using multiple languages in a mixed manner in a conversation;
and wherein the hardware processor is configured to align text from multiple languages by adversarial learning of a linear mapping between the multiple languages in an embedding space and refining alignments using geometric transformation.

2. The system of claim 1 , wherein the hardware processor is further configured to augment the adverse drag reaction corpus with at least one paraphrase of at least one of the utterances, the paraphrase being in multiple languages corresponding to the multiple languages.

3. The system of claim 1 , wherein the code-mixed utterance model includes a loss function including a 2-class softmax, in which a probability of a next utterance being a logical utterance with respect to a previous utterance is maximized.

4. The system of claim 1 , wherein the hardware processor is further configured to perform a fuzzification of the n-gram.

5. The system of claim 1 , wherein the combined language model includes a recurrent neural network.

6. The system of claim 1 , wherein the code-mixed utterance model includes a recurrent neural network.

7. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a device to cause the device to:
create a language model corpus including multilingual alignment for training a combined language model, the language model corpus created from data sourced from a computer network and train the combined language model;
create an adverse medication reaction corpus by analyzing data from online sources which include adverse medication reaction utterances and labeling an N-gram of an utterance in the adverse medication reaction utterances as a response to query;
and generate a code-mixed utterance model to perform code-mixed utterances in a turn by turn dialogue, by at least adding additional output layer including at least a start vector, language vector, and a query vector including at least the labeled N-gram, which are additional to the combined language model's predicted next words, wherein the start vector encodes what word should start a response to an input utterance, and the language vector encodes what language out of multiple languages used in the code-mixed utterances should a next word of the response be in, wherein the code-mixed utterances represent utterances performed using multiple languages in a mixed manner in a conversation;
and wherein the device is caused to align text from multiple languages by adversarial learning of a linear mapping between the multiple languages in an embedding space and refining alignments using geometric transformation.

8. The computer program product of claim 7 , wherein the device is further caused to augment the adverse drag reaction corpus with at least one paraphrase of at least one of the utterances, the paraphrase being in multiple languages corresponding to the multiple languages.

9. The computer program product of claim 7 , wherein the code-mixed utterance model includes a loss function including a 2-class softmax, in which a probability of a next utterance being a logical utterance with respect to a previous utterance is maximized.

10. The computer program product of claim 7 , wherein the device is further caused to perform a fuzzification of the n-gram.

11. The computer program product of claim 7 , wherein the combined language model includes a recurrent neural network.

12. The computer program product of claim 7 , wherein the code-mixed utterance model includes a recurrent neural network.

13. A computer-implemented method comprising:
creating a language model corpus including multilingual alignment for training a combined language model, the language model corpus created from data sourced from a computer network and train the combined language model;
creating an adverse medication reaction corpus by analyzing data from online sources which include adverse medication reaction utterances and labeling an N-gram of an utterance in the adverse medication reaction utterances as a response to query;
and generating a code-mixed utterance model to perform code-mixed utterances in a turn by turn dialogue, by at least adding additional output layer including at least a start vector, language vector, and a query vector including at least the labeled N-gram, which are additional to the combined language model's predicted next words, wherein the start vector encodes what word should start a response to an input utterance, and the language vector encodes what language out of multiple languages used in the code-mixed utterances should a next word of the response be in, wherein the code-mixed utterances represent utterances performed using multiple languages in a mixed manner in a conversation;
and aligning text from multiple languages by adversarial learning of a linear mapping between the multiple languages in an embedding space and refining alignments using geometric transformation.

14. The method of claim 13 , wherein further including augmenting the adverse drag reaction corpus with at least one paraphrase of at least one of the utterances, the paraphrase being in multiple languages corresponding to the multiple languages.

15. The method of claim 13 , wherein the code-mixed utterance model includes a loss function including a 2-class softmax, in which a probability of a next utterance being a logical utterance with respect to a previous utterance is maximized.

16. The method of claim 13 , further including performing a fuzzification of the n-gram.

17. The method of claim 13 , wherein generating the code-mixed utterance model includes training a recurrent neural network.

Assignees

• IBM

Inventors

• Akinwande Victor Abayomi
• Cintas Celia
• Walcott Aisha
• Ogallo William
• Remy Sekou Lionel

Classifications

• G06N3/0442

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

• G06N3/09

  Supervised learning · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06F40/20 Primary

  Natural language analysis (semantic analysis of natural language G06F40/30) · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title