Apparatus and method for extracting semantic topic

• US 10,423,723 B2
• Filed: 06/03/2015
• Issued: 09/24/2019
• Est. Priority Date: 12/06/2012
• Status: Active Grant

• Current Assignee: Korea University Research and Business Foundation
• Original Assignee: Korea University Research and Business Foundation
• Inventors: Lee, Sang Keun, Alam, Md Hijbul
• Primary Examiner(s): Willis, Amanda L
• Application Number:
  US14/729,306
• Publication Number:
  US 20150268930A1
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/36 Creation of semantic tools, e.g. ontology or thesauri
• CPC Class Codes: G06F 16/9535 Search customisation based on user profiles and personalisation
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06F 7/24 Sorting, i.e. extracting data from one or more carriers, rearranging the data in numerical or other ordered sequence, and rerecording the sorted data on the original carrier or on a different carrier or set of carriers sorting methods in generalG06F7/36 takes precedence
• CPC Class Codes: G06N 7/01 Probabilistic graphical models, e.g. probabilistic networks
• CPC Class Codes: G06Q 10/10 Office automation; Time management

Abstract

In accordance with a first exemplary embodiment, there is provided a method for extracting semantic topics from document sets in which opinions about an object are described using an apparatus capable of calculating a probability distribution. The method include (a) extracting word distributions about sentiment global topics and sentiment local topics; (b) extracting a global topic distribution, a local topic distribution and sentiment distributions about the global and local topics from the document sets; (c) performing statistical inference about each of the distributions extracted in the step (a) and the step (b); (d) extracting a global or local topic and a sentiment relevant to the global or local topic from the distributions of the inference performed in the step (c); and (e) extracting a word from the word distributions about sentiment global topics or sentiment local topics on the basis of the topic and sentiment extracted in the step (d).

Claims

1. A method for automatically extracting one or more semantic topics from one or more electronic document sets in which user'"'"'s opinions of an object are described using an apparatus capable of calculating a probability distribution, the method comprising:

(a) extracting, by a processor, a global word distribution from a plurality of global topic-sentiment pairs and a local word distribution from a plurality of local topic-sentiment pairs;

(b) extracting, by the processor, a global topic distribution corresponding to words constituting a global topic, a global sentiment distribution corresponding to words constituting a first sentiment about the global topic, a local topic distribution corresponding to words constituting a local topic, and a local sentiment distribution corresponding to words constituting a second sentiment about the local topic with respect to each document of the electronic document sets;

(c) performing, by the processor, statistical inference about each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution extracted in the step (a) and step (b);

(d) with respect to each document of the electronic document sets, extracting, by the processor, a first global topic from the global topic distribution and the global sentiment distribution, and a first local topic from the local topic distribution and the local sentiment distribution, and extracting, by the processor, a third sentiment relevant to the first global topic and a fourth sentiment relevant to the first local topic; and

(e) extracting, by the processor, one or more words from the global word distribution and the local word distribution on the basis of the first global topic, the first local topic, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic,

wherein each of the plurality of the global topic-sentiment pairs includes a first word expressing the global topic and a second word expressing the sentiment about the global topic, and each of the plurality of the local topic-sentiment pairs includes a third word expressing the local topic and a fourth word expressing the sentiment about the local topic,

wherein the step (b) includes;

extracting, by the processor, the global topic distribution and the global sentiment distribution about the global topic from the each document;

shifting, by the processor, one or more sliding windows overlapped with each other in each document;

extracting, by the processor, a categorical distribution of the sliding window; and

extracting, by the processor, the local topic distribution, the local sentiment distribution about the local topic and a topic context distribution based on words extracted from a sentence in the one or more sliding window;

wherein a size of the one or more sliding windows is set for the third word expressing the local topic and the fourth word expressing the sentiment about the local topic to be extracted together,

wherein the step (d) includes;

selecting a first sliding window and a first topic context with respect to each document;

if the first topic context is global, selecting the first global topic and the third sentiment relevant to the first global topic from the global topic distribution and the global sentiment distribution; and

,

if the first topic context is local, selecting the first local topic and the fourth sentiment relevant to the first local topic from the local topic distribution and the local sentiment distribution;

wherein the global topic includes a first group of aspects that are used to only distinguish the object from another object among aspects of the object, and

the local topic includes a second group of aspects that are sentiment-oriented and are used to calculate a rating of the object,

wherein the one or more words extracted in the step (e) includes the first global topic of the object, the first local topic of the object, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic, and

wherein the first global topic, the first local topic, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic are used to calculate a rating of the object and a rating of the aspect of the object.

2. The method of claim 1,

wherein each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution is based on a multinomial distribution or a categorical distribution.

3. The method of claim 2,

wherein each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution is constructed on basis of a Dirichlet prior probability.

4. The method of claim 1,

wherein the statistical inference uses a Gibbs sampling algorithm.

5. An apparatus for automatically extracting one or more semantic topics, comprising:

a storage memory medium configured to store steps of a semantic topic extraction program;

a document storage that stores one or more electronic document sets in which user'"'"'s opinions of an object are described; and

a processor configured to execute the steps of the semantic topic extraction program and that extracts a global topic including a first group of aspects that are used to only distinguish the object from another object among aspects of the object, a local topic including a second group of aspects that are sentiment-oriented and are used to calculate a rating of the object, and a sentiment about the global topic or the local topic from the electronic document sets stored in the document storage,

wherein the processor extracts a global word distribution from a plurality of global topic-sentiment pairs and a local word distribution from a plurality of local topic-sentiment pairs,

wherein each of the plurality of the global topic-sentiment pairs includes a first word expressing the global topic and a second word expressing the sentiment about the global topic, and each of the plurality of the local topic-sentiment pairs includes a third word expressing the local topic and a fourth word expressing the sentiment about the local topic,

extracts a global topic distribution corresponding to words constituting the global topic, a global sentiment distribution corresponding to words constituting a first sentiment about the global topic, a local topic distribution corresponding to words constituting the local topic, and a local sentiment distribution corresponding to words constituting a second sentiment about the local topic with respect to each document of the electronic document sets,

performs statistical inference about each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution extracted,

extracts a first global topic from the global topic distribution and the global sentiment distribution, and a first local topic from the local topic distribution and the local sentiment distribution, and extracts a third sentiment relevant to the first global topic and a fourth sentiment relevant to the first local topic, with respect to each document of the electronic document sets, and

extracts one or more words from the global word distribution and the local word distribution on the basis of the first global topic, the first local topic, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic,

wherein the processor extracts the global topic distribution, the global sentiment distribution, by extracting the global topic distribution and the global sentiment distribution about the global topic from the each document, shifting one or more sliding windows overlapped with each other in each document, extracting a categorical distribution of the sliding window, and extracting the local topic distribution, the local sentiment distribution about the local topic and a topic context distribution based on words extracted from a sentence in the one or more sliding window, wherein a size of the one or more sliding windows is set for the third word expressing the local topic and the fourth word expressing the sentiment about the local topic to be extracted together,

wherein the processor extracts the first global topic, the first local topic, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic, by selecting a first sliding window and a first topic context with respect to each document, selecting, if the first topic context is global, the first global topic and the third sentiment relevant to the first global topic from the global topic distribution and the global sentiment distribution, and selecting, if the first topic context is local, the first local topic and the fourth sentiment relevant to the first local topic from the local topic distribution and the local sentiment distribution,

wherein the one or more words extracted from the global word distribution and the local word distribution includes the first global topic of the object, the first local topic of the object, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic, and

wherein the first global topic, the first local topic, the third sentiment relevant to the first global topic and the fourth sentiment relevant to the first local topic are used to calculate a rating of the object and a rating of the aspect of the object.

6. The apparatus of claim 5,

wherein each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution is constructed on basis of a Dirichlet prior probability.

7. The apparatus of claim 5,

wherein the processor uses Gibbs sampling in statistical inference of each of the global word distribution, the local word distribution, the global topic distribution, the global sentiment distribution, the local topic distribution and the local sentiment distribution.