INDEXING CONTENT AT SEMANTIC LEVEL

• US 20110196670A1
• Filed: 02/01/2011
• Published: 08/11/2011
• Est. Priority Date: 02/09/2010
• Status: Active Grant

• Current Assignee: Siemens AG
• Original Assignee: Siemens Corp. (Siemens AG)
• Inventors: Dang, Jiangbo, Toklu, Candemir, Hampel, Kenneth, Kalender, Murat
• Granted Patent: US 8,751,218 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

Systems and methods are disclosed that perform automated semantic tagging. Automated semantic tagging produces semantically linked tags for a given text content. Embodiments provide ontology mapping algorithms and concept weighting algorithms that create accurate semantic tags that can be used to improve enterprise content management, and search for better knowledge management and collaboration.

Claims

1. A semantic tagging method that outputs semantically linked tags for text content comprising:

inputting the text content;

extracting nouns and noun phrases from the text content;

mapping the extracted nouns and noun phrases to terms of an ontology;

mapping the extracted nouns and noun phrases to a correct sense of the ontology terms using lexical chaining Word Sense Disambiguation (WSD) algorithms;

weighting the significance of the extracted nouns'"'"' and noun phrases'"'"' concepts from their ontological and statistical features;

extracting key-phrases from the weighted concepts; and

outputting the key-phrases as semantic tags.

2. The method according to claim 1 wherein the ontology for domain-independent applications is UNIpedia.

3. The method according to claim 1 wherein the ontology for domain-dependent applications is a domain-specific ontology.

4. The method according to claim 1 wherein extracting nouns and noun phrases from the text content further comprises using a Natural Language Processing (NLP) Application Programming Interface (API).

5. The method according to claim 4 wherein extracting nouns and noun phrases from the text content further comprises:

detecting sentences from the text content;

detecting tokens (words) from the detected sentences;

labeling the tokens; and

extracting consecutive noun tokens and noting their frequency of use.

6. The method according to claim 5 further comprising preprocessing the text content by removing all characters other than alphabetical characters and grammatical punctuation marks.

7. The method according to claim 6 further comprising removing numerical expressions.

8. The method according to claim 5 wherein detecting sentences further comprises splitting the text content into sentences.

9. The method according to claim 5 wherein detecting tokens from the sentences further comprises:

splitting the sentences into tokens; and

splitting the tokens that are contractions.

10. The method according to claim 5 wherein labeling the tokens further comprises labeling the tokens as part-of-speech such as noun, verb, adverb, etc.

11. The method according to claim 5 wherein extracting consecutive noun tokens further comprises parsing noun phrases.

12. The method according to claim 1 wherein mapping the extracted nouns and noun phrases to terms of an ontology further comprises:

accessing the ontology;

comparing the nouns and noun phrases with the ontology terms; and

if there is an exact match, retrieving the ontology term, and if there is not an exact match, performing phrase mapping using a rule-based algorithm.

13. The method according to claim 12 wherein the phrase mapping rule-based algorithm further comprises:

generating two sub-phrases by removing the first and last word from the noun phrases and replacing the original noun phrases with the two sub-phrases; and

repeating the phrase mapping until there is a match or there are no more noun phrases remaining.

14. The method according to claim 1 wherein mapping the extracted nouns and noun phrases to a correct sense of the ontology terms further comprises:

creating a hypernym lexical chain with WSD scores for senses of the input nouns and noun phrases;

creating a domain lexical chain with WSD scores for senses of the input nouns and noun phrases;

creating a holonym lexical chain with WSD scores for senses of the input nouns and noun phrases;

creating a category lexical chain with WSD scores for senses of the input nouns and noun phrases;

normalizing the WSD scores in each of the hypernym, domain, category and holonym lexical chains;

generating a WSD score using a sense rank algorithm;

assigning weights to each algorithm and summing the weighted WSD scores for the hypernym, domain, category, holonym lexical chains, and the sense rank algorithm;

evaluating the hypernym, domain, category and holonym lexical chains, the sense rank algorithm and optimizing their weights for a given target data; and

mapping the nouns and noun phrases to a highest ranked sense.

15. The method according to claim 14 wherein creating a hypernym lexical chain further comprises receiving a set of nouns and noun phrases and corresponding senses from the ontology and creating a lexical chain.

16. The method according to claim 15 wherein creating a hypernym lexical chain further comprises:

initializing each nouns'"'"' and noun phrases'"'"' senses WSD score to zero;

connecting one sense to another sense having a common ancestor; and

increasing the WSD scores of a sense'"'"'s ancestor until reaching a predetermined threshold of ancestor distance, wherein a sense that has a higher number of connected senses in the lexical chain receives a higher WSD score, and these senses form a noun and noun phrase context.

17. The method according to claim 16 wherein creating a hypernym lexical chain further comprises:

incrementing a sense score by adding the WSD scores of their ancestor senses; and

selecting extracted nouns and noun phrases based on senses that receive the highest WSD scores among the nouns'"'"' and noun phrases'"'"' senses.

18. The method according to claim 17 wherein creating a hypernym lexical chain for senses derived from the same nouns and noun phrases that have different parent senses and these parent senses may have common ancestor further comprises:

subtracting scores added by a sense itself and senses that have common sense names; and

adding an extra score for a sense based on the number of its synonyms.

19. The method according to claim 14 wherein creating a domain lexical chain further comprises:

connecting senses defined in a same domain via a domain concept node to form a domain lexical chain;

scoring each sense according to the number of concepts within the domain lexical chain;

if a sense itself is a domain concept, its WSD score is equal to the number of concepts within its lexical chain; and

mapping a noun or noun phrase to a sense that receives the highest WSD score among all senses of the noun or noun phrase.

20. The method according to claim 14 wherein creating a holonym lexical chain further comprises:

connecting two senses with each other if there is a holonym relationship between them or if they have a common holonym property;

considering only the first degree holonym relationships when connecting senses to create a holonym lexical chain;

scoring each sense based on the number of concepts in its connected holonym lexical chains; and

mapping a noun and noun phrase to a sense that receives a higher WSD score than the noun'"'"'s and noun phrase'"'"'s other senses.

21. The method according to claim 14 wherein creating a category lexical chain further comprise:

connecting senses defined in a same category via a category concept node to create a category lexical chain;

scoring a sense according to the number of concepts within the category lexical chain; and

mapping a noun and noun phrase to a sense that receives the highest WSD score among all senses of the noun and noun phrase.

22. The method according to claim 14 wherein summing the weighted WSD scores for the hypernym, domain, category, holonym lexical chains, and the sense rank algorithm further comprises:

calculating weights based on the performance of the hypernym lexical chain, domain lexical chain, holonym lexical chain, category lexical chain, and the sense rank algorithm; and

summing weighted WSD scores from the hypernym lexical chain, domain lexical chain, holonym lexical chain, category lexical chain, and the sense rank algorithm.

23. The method according to claim 14 wherein evaluating the hypernym, domain, category, holonym lexical chains, and the sense rank algorithm and optimizing their weights for a given target data further comprises:

selecting a subset from the target data source as a training data set;

performing the hypernym lexical chain, domain lexical chain, holonym lexical chain and category lexical chain algorithms;

calculating the hypernym lexical chain, domain lexical chain, holonym lexical chain, category lexical chain Precision, Recall, and F-measure;

including WSD scores from the sense rank algorithm;

calculating the sense rank algorithm Precision, Recall, and F-measure; and

calculating weights of scores from the hypernym lexical chain, domain lexical chain, holonym lexical chain, category lexical chain and sense rank algorithms based on the performance of the hypernym lexical chain, domain lexical chain, holonym lexical chain and category lexical chain and sense rank.

24. The method according to claim 1 wherein weighting the significance of the hypernym, domain, category and holonym lexical chains further comprises:

computing semantic Term Significance (TS) scores using lexical chaining WSD scores derived from previous phrase, Depth and Information Content (IC) values;

computing statistical TS scores using Term Frequency (TF) and Inverse Google Popularity (IGP);

computing concept TS scores by summing the semantic and statistical scores;

sorting concepts by their concept TS scores in descending order and extracting top key-phrases;

diversifying key-phrase distribution over different categories using heuristics; and

clustering key-phrases according to categories such as people, event, location, time, etc. to describe the text content.

25. The method according to claim 1 wherein extracting key-phrases from the weighted concepts further comprises:

filtering all key-phrases (redundant senses) except the top ranked key-phrase of a sense; and

filtering key-phrases that have a lower frequency than all occurrences of its children concepts.