Method and apparatus for informative training repository building in sentiment analysis model learning and customization

Patent metadata

Field	Value
Publication number	US-10824812-B2
Application number	US-201615175808-A
Country	US
Kind code	B2
Filing date	Jun 7, 2016
Priority date	Jun 7, 2016
Publication date	Nov 3, 2020
Grant date	Nov 3, 2020

Abstract

The methods, systems, and computer program products described herein provide ways to generate an informative training corpus of samples for use in machine training a high-quality sentiment analysis computer model. In some aspects, a method is disclosed including receiving a plurality of training samples, extracting semantic and sentiment elements of one or more of the training samples, generalizing the semantic and sentiment elements of the one or more of the training samples, generating an informative ranking score for one or more of the training samples based on the generalized semantic and sentiment elements, selecting informative training samples from the plurality of training samples based at least in part on the generated informative ranking scores, and adding the selected informative training samples to an informative training corpus.

First claim

What is claimed is:

1. A method, comprising:
receiving, by at least one processor, a plurality of training samples;
extracting, by at least one processor, semantic and sentiment elements of one or more of the training samples;
generalizing, by at least one processor, the extracted semantic and sentiment elements of the one or more of the training samples;
generating, by at least one processor, an informative ranking score for one or more of the training samples based on the generalized semantic and sentiment elements, the informative ranking score generated as a function of a weighted element density associated with said one or more of the training samples and a weighted semantic diversity associated with said one or more of the training samples, wherein the weighted semantic diversity is determined as a sum of, a weighted ratio between aspect categories of a given training sample and all training samples being considered, and a weighted ratio between extracted elements of the given training sample and said all training samples being considered;
selecting, by at least one processor, informative training samples from the plurality of training samples based at least in part on the generated informative ranking scores;
adding, by at least one processor, the selected informative training samples to an informative training corpus; and
machine training a semantic analysis computer model using the informative training corpus.

2. The method of claim 1 , wherein extracting semantic and sentiment elements from one or more of the training samples includes, for each of the one or more of the training samples:
analyzing the training sample to identify an entity that is the subject of the training sample;
analyzing the training sample to identify one or more feature terms of the training sample; and
analyzing the training sample to identify one or more opinion words of the training sample.

3. The method of claim 2 , wherein generalizing the extracted semantic and sentiment elements of the one or more of the training samples includes, for each of the one or more training samples:
categorizing the identified entity and feature terms into one or more aspect categories;
grouping the identified opinion words into one or more opinion word groups; and
extracting a latent semantic association context structure for one or more of the aspect categories.

4. The method of claim 3 , wherein generating the informative ranking score for the one or more of the training samples includes, for each of the one or more training samples:
constructing a sentiment topic feature vector for the training sample based on at least one of the identified entity, the identified feature terms, the identified opinion words, the one or more aspect categories, the one or more opinion word groups, and the latent semantic association context for the one or more aspect categories; and
generating an informative ranking score for the training sample based at least in part on the generated sentiment topic feature vector.

5. The method of claim 3 , wherein the latent semantic association context structure comprises one or more context pairs, each pair comprising:
at least one of the identified entity and one of the feature terms; and
an opinion word.

6. The method of claim 1 , wherein the extracting, generalizing, generating, selecting, and adding are performed autonomously.

7. A system, comprising:
at least one hardware processor programmed to:
receive a plurality of training samples;
extract semantic and sentiment elements of one or more of the training samples;
generalize the extracted semantic and sentiment elements of the one or more of the training samples;
generate an informative ranking score for one or more of the training samples based on the generalized semantic and sentiment elements, the informative ranking score generated as a function of a weighted element density associated with said one or more of the training samples and a weighted semantic diversity associated with said one or more of the training samples, wherein the weighted semantic diversity is determined as a sum of, a weighted ratio between aspect categories of a given training sample and all training samples being considered, and a weighted ratio between extracted elements of the given training sample and said all training samples being considered;
select informative training samples from the plurality of training samples based at least in part on the generated informative ranking scores;
add the selected informative training samples to an informative training corpus; and

a storage device coupled to the at least one hardware processor,
wherein the at least one hardware processor is further programmed for machine training a semantic analysis computer model using the informative training corpus.

8. The system of claim 7 , wherein to extract the semantic and sentiment elements from one or more of the training samples, for each of the one or more of the training samples, the at least one hardware processor:
analyzes the training sample to identify an entity that is the subject of the training sample;
analyzes the training sample to identify one or more feature terms of the training sample; and
analyzes the training sample to identify one or more opinion words of the training sample.

9. The system of claim 8 , wherein to generalize the extracted semantic and sentiment elements of the one or more of the training samples, for each of the one or more training samples, the at least one hardware processor:
categorizes the identified entity and feature terms into one or more aspect categories;
groups the identified opinion words into one or more opinion word groups; and
extracts a latent semantic association context structure for one or more of the aspect categories.

10. The system of claim 9 , wherein to generate the informative ranking score for the one or more of the training samples, for each of the one or more training samples, the at least one hardware processor:
constructs a sentiment topic feature vector for the training sample based on at least one of the identified entity, the identified feature terms, the identified opinion words, the one or more aspect categories, the one or more opinion word groups, and the latent semantic association context for the one or more aspect categories; and
generates an informative ranking score for the training sample based at least in part on the generated sentiment topic feature vector.

11. The system of claim 9 , wherein the latent semantic association context structure comprises one or more context pairs, each context pair comprising:
at least one of the identified entity and one of the feature terms; and
an opinion word.

12. The system of claim 7 , wherein the at least one hardware processor performs extracting, generalizing, generating, selecting, and adding autonomously.

13. A computer program product comprising a computer readable storage medium having program instructions embodied therewith, wherein the computer readable storage medium is not a transitory signal per se, the program instructions readable by a processor to cause the processor to perform a method comprising:
receiving a plurality of training samples;
extracting semantic and sentiment elements of one or more of the training samples;
generalizing the extracted semantic and sentiment elements of the one or more of the training samples;
generating an informative ranking score for one or more of the training samples based on the generalized semantic and sentiment elements, the informative ranking score generated as a function of a weighted element density associated with said one or

Assignees

• IBM

Inventors

• Cai Keke
• Guo Honglei
• Jiang Jian Min
• Su Zhong
• Sun Changhua
• Tang Guoyu

Classifications

• G06Q30/02 Primary

  Marketing; Price estimation or determination; Fundraising · CPC title

• G06F40/30 Primary

  Semantic analysis · CPC title