Apparatus for automatic theme detection from unstructured data

• US 10,372,741 B2
• Filed: 03/01/2013
• Issued: 08/06/2019
• Est. Priority Date: 03/02/2012
• Status: Active Grant

• Current Assignee: Clarabridge Incorporated
• Original Assignee: Clarabridge Incorporated
• Inventors: Mathew, Thomas, Voorhees, Kenneth Robert
• Primary Examiner(s): Featherstone, Mark D
• Assistant Examiner(s): McQuitery, Diedra
• Application Number:
  US13/783,063
• Publication Number:
  US 20130268534A1
• Time in Patent Office: 2,349 Days
• Field of Search: 704 9, 705 729, 706 14, 706 46, 706 11, 706 12, 706 47, 706 48, 706 52, 706 55, 706 20, 706 2021, 707E17014, 707E17108, 707E17044, 707737, 707705, 707E17005, 707E17059, 707E17089, 707706, 707709, 707736, 707769, 707E17002, 707E17008, 707E17033, 707E17058, 707E17069, 707E17075, 707E17084, 707E17099, 707722, 707731, 707740, 707741, 707748, 707749, 707738, 707739, 715221, 715222, 715255, 715719, 715764, 725 13
• CPC Class Codes: G06F 16/35 Clustering; Classification
• CPC Class Codes: G06F 40/237 Lexical tools
• 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 40/40 Processing or translation of natural language natural language analysis G06F40/20; semantic analysis G06F40/30

Abstract

This apparatus provides a system and method of determining significant repeating themes in a collection of documents. The apparatus operates unsupervised and leverages a natural language processing mechanism supported with lexicon, synonym and taxonomy dictionaries to determine themes and establish their relevance using a two-level hierarchical structure. The apparatus also assigns meaningful names to identified themes and determines a set of rules that describe the theme such that it can be applied to categorize other documents outside of the collection as well.

Claims

1. A system comprising:

a repository of unstructured documents stored in a computing system;

a natural language processor configured to perform language processing; and

a non-transitory computer-readable storage medium comprising instructions that, when executed, enable a computing system to detect themes within the unstructured documents by;

removing noise words from the unstructured documents, to yield clean documents;

initiate a sentiment computation component configured to determine sentiment of each word in the clean documents by;

assigning to each word in the clean documents at least one of a positive sentiment, a negative sentiment, and neutral sentiment, to yield assigned sentiments;

determining a sentiment probability of a section of the unstructured data based on the assigned sentiments of words in the section, to yield assigned sectional sentiment; and

determining an overall sentiment probability distribution for the unstructured documents based on the assigned sectional sentiment of multiple sections of the unstructured documents;

initiate a theme detection component configured to;

discover themes based on topics with neutral sentiment when the topics are located in a section of the unstructured documents with a sentiment probability that is greater than an overall sentiment probability distribution;

assign labels to each discovered theme;

identify patterns that describe each theme;

identify instances of the themes within individual documents of the unstructured documents based on a presence of the patterns in the individual documents; and

organize the themes in a hierarchy using the instances of the themes; and

initiate a user interface configured to;

allow an operator to initiate theme detection by the theme detection component; and

allow an operator to view and interact with results of the theme detection, wherein the results comprise at least one of the assigned labels, the patterns, and the hierarchy.

2. The system of claim 1, further comprising:

a component configured to;

acquire additional unstructured documents; and

add the additional unstructured documents to the repository of unstructured documents.

3. The system of claim 1, further comprising:

a natural language processing component configured to extract features from unstructured text,

wherein the theme detection component uses the extracted features in the theme detection process.

4. The system of claim 1, wherein the section of the unstructured data is a sentence.

5. The system of claim 4, wherein each sentence in the unstructured documents has an assigned sentiment based on sentiments of individual, non-noise words contained within each respective sentence.

6. The system of claim 1,

wherein the theme detection component uses the determined sentiment to weight the significance of a discovered theme.

7. The system of claim 1, wherein the themes organized into the hierarchy comprise top level themes and second level themes,

wherein the second level themes are based on terms that appear in the unstructured documents at a frequency above a predetermined threshold, and

wherein a second level theme is associated with a top-level theme if terms of the second level theme appear within a predetermined distance of the top-level theme.

8. The system of claim 1, wherein the themes organized into the hierarchy comprise top level themes and second level themes, and

wherein a second level theme is based on a term that co-occurs with a top level theme for at least a predetermined percentage of total occurrences of the term.

9. The system of claim 8, wherein the second level theme is also based on the co-occurrence of the term with the top level theme above a predetermined percentage of total occurrences of the top level theme.

10. The system of claim 1, wherein the hierarchy includes top level themes and second level themes, the second level themes being based on terms that co-occur with the top level themes and being created using parts of speech of terms, and

wherein the second level themes comprise frequently co-occurring nouns and adjectives if the top level theme is a noun, and wherein the second level themes comprise frequently co-occurring nouns if the top level theme is a verb.

11. A method of determining themes from a collection of unstructured text documents, the method comprising:

receiving a set of unstructured text documents to process;

removing noise words from the unstructured text documents, to yield clean documents;

initiate a language processing component configured to process unstructured data collected from the unstructured text documents;

determining, by a computer system configured to perform natural language processing, sentiment associated with each word in the clean documents, by;

assigning to each word in the clean documents at least one of a positive sentiment, a negative sentiment, and a neutral sentiment, to yield assigned sentiments;

determining a sentiment probability of a section of the unstructured data based on the assigned sentiments of words in the section; and

determining an overall sentiment probability distribution for the unstructured documents based on the assigned sectional sentiment of multiple sections of the unstructured text documents;

determining, by a computing system, a first set of topics within the unstructured text documents of the set and determining a second set of topics based on frequently occurring terms within the set of unstructured text documents;

determining, by the computing system, a label for each term in the frequently occurring terms;

determining, by the computing system, one or more text patterns, wherein the one or more text patterns are used to identify if a term is contained within a document; and

creating, by the computing system, a category model to organize the identified terms as themes in a hierarchical structure that includes top level themes,

wherein a topic having neutral sentiment is identified as a theme when the topic is located in a section of the unstructured text documents with a sentiment probability that is greater than an overall sentiment probability distribution.

12. The method of claim 11, wherein the section of the unstructured data is a sentence.

13. The method of claim 11, further comprising:

creating second level themes based on terms that co-occur with the top level discovered themes; and

incorporating the second level themes into the category model.

14. The method of claim 11, further comprising:

selecting nouns and verbs as top level themes using parts of speech of terms.

15. The method of claim 14, further comprising:

selecting a noun as top level themes if the noun demonstrates a tendency to appear more in sentences expressing neutral sentiment, and

selecting a noun as a second level theme when the noun demonstrates a positive or negative sentiment.

16. The method of claim 14, further comprising:

using verbs as a top level theme only if the verb is associated with neutral sentiment.

17. The method of claim 13, wherein the creating the second level themes based on terms that co-occur with the top level discovered themes further comprises:

creating the second level of themes using parts of speech of terms, wherein the second level themes comprise frequently co-occurring nouns and adjectives if the top level theme is a noun, and wherein the second level themes comprise frequently co-occurring nouns if the top level theme is a verb.

18. The method of claim 11, wherein the set of unstructured text documents comprises categorized documents and uncategorized documents;

and wherein the receiving the set of unstructured text documents to process further comprises;

removing the categorized documents from the set of unstructured text documents prior to identifying themes.

19. The method of claim 11, wherein the set of unstructured text documents comprises uncategorized documents, and wherein the receiving the set of unstructured text documents to process further comprises:

isolating the uncategorized documents to process for theme detection from a current unstructured document harvest, wherein only uncategorized documents from a document categorization step are used in the current unstructured document harvest.

20. The method of claim 11, further comprising:

grouping the top level themes using an ontology to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the ontology.

21. The method of claim 11, further comprising:

grouping top level themes using a synonym dictionary to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the synonym dictionary.

22. The method of claim 13, further comprising:

grouping second level themes using ontology to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the ontology.

23. The method of claim 13, further comprising:

grouping second level themes using a synonym dictionary to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the synonym dictionary.

24. The method of claim 11, wherein the hierarchical structure further includes second level themes, and

wherein a second level theme is based on a term that co-occurs with a top level theme for at least a predetermined percentage of total occurrences of the term.

25. The method of claim 24, wherein the second level theme is also based on the co-occurrence of the term with the top level theme above a predetermined percentage of total occurrences of the top level theme.

26. A non-transitory computer-readable storage medium comprising instructions that when executed enable a natural language processing computing system to:

receiving a set of unstructured text documents to process;

removing noise words from the unstructured text documents, to yield clean documents;

initiate a language processing component configured to process unstructured data collected from the unstructured text documents;

determining sentiment associated with each word in the clean documents, by;

assigning to each word in the clean documents at least one of a positive sentiment, a negative sentiment, and a neutral sentiment, to yield assigned sentiments;

determining a sentiment probability of a section of the unstructured data based on the assigned sentiments of words in the section; and

determining an overall sentiment probability distribution for the unstructured documents based on the assigned sectional sentiment of multiple sections of the unstructured text documents;

determining a first set of topics within the unstructured text documents of the set and determining a second set of topics based on frequently occurring terms within the set of unstructured text documents;

determining a label for each term in the frequently occurring terms;

determining one or more text patterns, wherein the one or more text patterns are used to identify if a term is contained within a document; and

creating a category model to organize the identified terms as themes in a hierarchical structure that includes top level themes,

wherein a topic having neutral sentiment is identified as a theme when the topic is located in a section of the unstructured text documents with a sentiment probability that is greater than an overall sentiment probability distribution.

27. The non-transitory computer-readable storage medium of claim 26,

wherein the section of the unstructured data is a sentence.

28. The non-transitory computer-readable storage medium of claim 26, further comprising instructions that when executed enable the computing system to:

create second level themes based on terms that co-occur with the top level discovered themes; and

incorporate the second level themes into the category model.

29. The non-transitory computer-readable storage medium of claim 26, further comprising instructions that when executed enable the computing system to:

select nouns and verbs as top level themes using parts of speech of terms.

30. The non-transitory computer-readable storage medium of claim 29, further comprising instructions that when executed enable the computing system to:

select a noun as top level themes if the noun demonstrates a tendency to appear more in sentences expressing sentiment.

31. The non-transitory computer-readable storage medium of claim 29, further comprising instructions that when executed enable the computing system to:

use verbs as a top level theme only if the verb is associated with sentiment.

32. The non-transitory computer-readable storage medium of claim 28, wherein the instructions for creating second level themes based on terms that co-occur with the top level discovered themes further comprise instructions that when executed enable the computing system to:

create the second level of themes using parts of speech of terms, wherein the second level themes comprise frequently co-occurring nouns and adjectives if the top level theme is a noun, and wherein the second level themes comprise frequently co-occurring nouns if the top level theme is a verb.

33. The non-transitory computer-readable storage medium of claim 26, wherein the set of unstructured text documents comprises categorized documents and uncategorized documents;

and wherein the instructions for receiving the set of unstructured text documents to process further comprise instructions that when executed enable the computing system to;

remove the categorized documents from the set of unstructured text documents prior to identifying themes.

34. The non-transitory computer-readable storage medium of claim 26, wherein the set of unstructured text documents comprises categorized documents and uncategorized documents;

and wherein the instructions for receiving the set of unstructured text documents to process further comprise instructions that when executed enable the computing system to;

isolate uncategorized documents to process for theme detection from a current unstructured document harvest, wherein only uncategorized documents from a document categorization step are used in the current unstructured document harvest.

35. The non-transitory computer-readable storage medium of claim 26, further comprising instructions that when executed enable the computing system to:

group the top level themes using an ontology to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the ontology.

36. The non-transitory computer-readable storage medium of claim 26, further comprising instructions that when executed enable the computing system to:

group top level themes using a synonym dictionary to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the synonym dictionary.

37. The non-transitory computer-readable storage medium of claim 28, further comprising instructions that when executed enable the computing system to:

group second level themes using ontology to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the ontology.

38. The non-transitory computer-readable storage medium of claim 28, further comprising instructions that when executed enable the computing system to:

group second level themes using a synonym dictionary to produce one or more groups, wherein each of the one or more groups contains themes that are related according to the synonym dictionary.

39. The non-transitory computer-readable storage medium of claim 26, wherein the hierarchical structure further includes second level themes, and

wherein a second level theme is based on a term that co-occurs with a top level theme for at least a predetermined percentage of total occurrences of the term.

40. The non-transitory computer-readable storage medium of claim 39, wherein the second level theme is also based on the co-occurrence of the term with the top level theme above a predetermined percentage of total occurrences of the top level theme.

41. The non-transitory computer-readable storage medium of claim 26, wherein the hierarchical structure also include second level themes based on terms that co-occur with the top level themes, the second level themes being created using parts of speech of terms,

wherein the second level themes being incorporated into the category model, and

wherein the second level themes comprise frequently co-occurring nouns and adjectives if the top level theme is a noun, and wherein the second level themes comprise frequently co-occurring nouns if the top level theme is a verb.