1. Extracting Key Terms From Noisy and Multi-theme Documents Maria Grineva, Maxim Grinev and Dmitry Lizorkin Institute for System Programming of RAS

2. Outline 1.Key terms extraction: traditional approaches and applications 2.Using Wikipedia as a knowledge base for Natural Language Processing 3.Main techniques of our approach: Wikipedia-based semantic relatedness Network analysis algorithm to detect community structure in networks 4.Our method 5.Experimental evaluation

3. Key Terms Extraction Basic step for various NLP tasks: –document classification –document clustering –text summarization –inferring a more general topic of a text document Core task of Internet content-based advertising systems, such as Google AdSense and Yahoo! Contextual Match –Web pages are typically noisy (side bars/menus, comments, future announces, etc.) –Dealing with multi-theme Web pages (portal home pages, etc.)

4. Approaches to Key Terms Extraction Based on statistical learning: –use for example: frequency criterion (TFxIDF model), keyphrase-frequency, distance between terms normalized by the number of words in the document (KEA) –compute statistical features over Wikipedia corpus (Wikify! ) –require training set Based on analyzing syntactic or semantic term relatedness within a document –compute semantic relatedness between terms (using, for example, Wikipedia) –modeling document as a semantic graph of terms and applying graph analysis techniques to it (TextRank) –no training set required

5. Using Wikipedia as a Knowledge Base for Natural Language Processing Wikipedia (www.wikipedia.org) – free open encyclopedia –Today Wikipedia is the biggest encyclopedia (more than 2.7 million articles in English Wikipedia) –It is always up-to-date thanks to millions of editors over the world –Has huge network of cross-references between articles, large number of categories, redirect pages, disambiguation pages => rich resource for bootstrapping NLP and IR tasks

6. Basic Techniques of Our Method: Semantic Relatedness of Terms Semantic relatedness assigns a score for a pair of terms that represents the strength of relatedness between the terms We use Wikipedia compute terms semantic relatedness We use semantic relatedness to model document as a graph of terms

7. Wikipedia-based semantic relatedness for the two terms can be computed using: –the links found within their corresponding Wikipedia articles –Wikipedia categories structure –the article’s textual content Using Dice-measure for Wikipedia-based semantic relatedness Basic Techniques of Our Method: Semantic Relatedness of Terms

8. Basic Techniques of Our Method: Detecting Community Structure in Networks We discover terms communities in a document graph Community – densely interconnected group of nodes in a network Girvan-Newman algorithm for detection community structure in networks: betweenness – how much is edge “in between” different communities modularity - partition is a good one, if there are many edges within communities and only a few between them

9. Our Method 1.Candidate terms extraction 2.Word sense disambiguation 3.Building semantic graph 4.Discovering community structure of the semantic graph 5.Selecting valuable communities

10. Our Method: Candidate Terms Extraction Goal: extract all terms from the document and for each term prepare a set of Wikipedia articles that can describe its meaning Parse the input document and extract all possible n- grams For each n-gram (+ its morphological variations) provide a set of Wikipedia article titles –“drinks”, “drinking”, “drink” => [Wikipedia:] Drink; Drinking

11. Our Method: Word Sense Disambiguation Goal: choose the most appropriate Wikipedia article from the set of candidate articles for each ambiguous term extracted on the previous step Use of Wikipedia disambiguation and redirect pages to obtain candidate meanings of ambiguous terms Denis Turdakov, Pavel Velikhov “Semantic Relatedness Metric for Wikipedia Concepts Based on Link Analysis and its Application to Word Sense Disambiguation” SYRCoDIS, 2008

12. Our Method: Building Semantic Graph Goal: building document semantic graph using semantic relatedness between terms Semantic graph built from a news article "Apple to Make ITunes More Accessible For the Blind"

13. Our Method: Detecting Community Structure of the Semantic Graph

14. Our Method: Selecting Valuable Communities Goal: rank term communities in a way that: –the highest ranked communities contain key terms –the lowest ranked communities contain not important terms, and possible disambiguation mistakes Use: –density of community – sum of inner edges of community divided by the number of vertices in this community –informativeness – sum of keyphraseness measure (Wikipedia-based TFxIDF analogue) of community terms Community rank: density*informativeness

15. Our Method: Selecting Valuable Communities In 73% of web pages decline in communities scores separates key-terms communities from non-important ones

16. Advantages of the Method No training. Instead of training the system with hand- created examples, we use semantic information derived from Wikipedia Noise and multi-theme stability. Good at filtering out noise and discover topics in Web pages Thematically grouped key terms. Significantly improve further inferring of document topics using, for example, spreading activation over Wikipedia categories graph High accuracy. Evaluated using human judgments (further in this presentation)

17. Experimental Evaluation on Noise-free dataset Classical – TFxIDF, Yahoo! Terms Extractor Wikipedia-based – Wikify!, TextRank Evaluation on noise-free dataset (blog posts) using human judgment

18. Comparison to other methods Experimental Evaluation on Web Pages Performance of our method on different kinds of Web pages

19. Multi-theme stability evaluated on compound Web pages (popular news site, portal homepages, etc.) Experimental Evaluation on Web Pages

20. Thank You! Any Questions? Email upa@grinev.net maxim@grinev.net lizorkin@ispras.ru