Loose term-centric representation for term classification in aspect-based sentiment analysis

Patent metadata

Field	Value
Publication number	US-9633007-B1
Application number	US-201615079883-A
Country	US
Kind code	B1
Filing date	Mar 24, 2016
Priority date	Mar 24, 2016
Publication date	Apr 25, 2017
Grant date	Apr 25, 2017

Abstract

A method for aspect categorization includes receiving an input text sequence and identifying aspect terms and sentiment phrases in the input text sequence, where present. For an identified aspect term, identifying sentiment dependencies in which the aspect term is in a syntactic dependency with one of the identified sentiment phrases, and identifying pseudo-dependencies from a dependency graph of the input text sequence. The dependency graph includes a sequence of nodes. In a pseudo-dependency, a node representing the aspect term precedes or follows a node representing a semantic anchor in the dependency graph without an intervening other aspect term. Features for the aspect term are extracted from at least one of identified sentiment dependencies and identified pseudo-dependencies. With a classifier trained to output at least one of category labels and polarity labels for aspect terms, classifying the identified aspect term based on the extracted features.

First claim

What is claimed is:

1. A method for aspect categorization comprising:
receiving an input text sequence;
providing for identifying aspect terms in the input text sequence;
providing for identifying sentiment phrases in the input text sequence;
for an identified aspect term:
providing for identifying sentiment dependencies in which the aspect term is in a syntactic dependency with one of the identified sentiment phrases, and
from a dependency graph of the input text sequence, the dependency graph comprising a sequence of nodes, providing for identifying pseudo-dependencies in which a node representing the aspect term precedes or follows a node representing a semantic anchor in the dependency graph without an intervening other aspect term;

extracting features from at least one of identified sentiment dependencies and identified pseudo-dependencies;
with a classifier trained to output at least one of category labels and polarity labels for aspect terms, classifying the identified aspect term based on the extracted features; and
outputting information based on the classification.

2. The method of claim 1 , wherein at least one of the providing for identifying aspect terms and sentiment phrases, providing for identifying sentiment dependencies, providing for identifying pseudo-dependencies, extracting features, and classifying the identified aspect term based on the extracted features is performed with a processor.

3. The method of claim 1 , wherein the semantic anchor nodes are selected from nodes with lexical semantic information about categories, nodes with polarity information, and combinations thereof.

4. The method of claim 1 , wherein the providing identifying pseudo-dependencies excludes from consideration semantic anchor nodes representing semantic anchors which are identified as being in a sentiment dependency with another aspect term represented as a node in the dependency graph.

5. The method of claim 4 , wherein the providing identifying pseudo-dependencies excludes from consideration backward links from the aspect term node to semantic anchor nodes representing semantic anchors where the aspect term node has been identified as being in a sentiment dependency with another aspect term represented as a node in the dependency graph.

6. The method of claim 1 , wherein the extraction of features takes into account syntactic negation associated with an identified semantic anchor.

7. The method of claim 1 , wherein each aspect term is a word or sequence of words which names a particular aspect of a target entity in a domain of interest.

8. The method of claim 1 , wherein the providing for identifying aspect terms in the input text sequence comprises at least one of rule-based detection of aspect terms and CRF-based detection of aspect terms.

9. The method of claim 1 , wherein the providing for identifying sentiment phrases in the input text sequence comprises accessing a polar vocabulary which includes a set of sentiment phrases.

10. The method of claim 1 , wherein the classifier includes a first classifier for outputting category labels for aspect terms and a second classifier for outputting polarity labels for aspect terms.

11. The method of claim 10 , wherein the second classifier is trained with generic features in which aspect terms are replaced by categories.

12. The method of claim 1 , wherein the category labels include at least four category labels.

13. The method of claim 1 , wherein the classifier includes one of a model learned by logistic regression and a model leaned using singular value decomposition of the set of features and a one-versus-all Elastic Net regression model to infer category and polarity.

14. The method of claim 1 , further comprising training a first classifier for predicting category labels for aspect terms and a second classifier for predicting polarity labels for aspect terms based on features extracted from training samples labeled with category and polarity information.

15. The method of claim 14 , wherein the training includes extracting sentiment dependencies and pseudo-dependencies from the training samples and generating feature-based representations of aspect terms in the training samples based on the respective extracted sentiment dependencies and pseudo-dependencies.

16. A computer program product comprising a non-transitory recording medium storing instructions, which when executed on a computer, causes the computer to perform the method of claim 1 .

17. A system comprising memory which stores instructions for performing the method of claim 1 and a processor in communication with the memory which executes the instructions.

18. A system for aspect categorization comprising:
an aspect term detection component which detects aspect terms providing for identifying aspect terms in an input text sequence;
a sentiment phrase detection component which detects sentiment phrases in the input text sequence;
a sentiment dependency extraction component which identifies sentiment dependencies in which an identified aspect term is in a syntactic dependency with an identified sentiment phrase;
a pseudo-dependency extraction component which identifies pseudo-dependencies from a dependency graph of the input text sequence, the pseudo-dependencies including an aspect term and a semantic anchor preceding or following the aspect term in the dependency graph without an intervening other aspect term;
a feature extraction component which extracts features from at least one of identified sentiment dependencies and identified pseudo-dependencies;
a classifier which classifies an identified aspect term based on the extracted features according to at least one of categories and polarity; and
a processor which implements the components.

19. The system of claim 18 , wherein the classifier includes a first classifier trained to predict category labels for aspect terms and a second classifier trained to predict polarity labels for aspect terms.

20. A method for training classifiers for aspect categorization comprising:
receiving a set of training samples in which aspect terms are labeled with a category selected from a plurality of predefined categories and a polarity selected from a plurality of polarities;
extracting sentiment dependencies and pseudo-dependencies from the training samples, each sentiment dependency including an aspect term in a syntactic dependency with a sentiment phrase, each pseudo-dependency including an aspect term and a semantic anchor which precedes or follows the aspect term in a dependency graph of the input text sequence without any intervening aspect terms;
extracting features for each of a plurality of the aspect terms based on any sentiment dependencies and pseudo-dependencies in which the aspect term participates;
and learning at least one classifier model for predicting category labels and polarity labels for an input text sequence based on the extracted features,
wherein at least one of the extraction of sentiment dependencies and pseudo-dependencies and the extraction of features is performed with a processor.

Assignees

• Xerox Corp

Inventors

• Brun Caroline
• Perez Julien
• Roux Claude

Classifications

• G06F40/30 Primary

  Semantic analysis · CPC title

• G06F16/9024

  Graphs; Linked lists (G06F16/9027 takes precedence) · CPC title

• G06F40/284 Primary

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F17/2785 Primary

  Physics · mapped topic

• G06F17/2715

  Physics · mapped topic