1. P2P Concept Search Fausto Giunchiglia Uladzimir Kharkevich S.R.H Noori April 21st, 2009, Madrid, Spain

2. Problems of syntactic approach Low precision. Caused by: ▫Polysemy,word or phrase with more senses:  Java -> Island, coffee, programming language?  Check -> bank check or Verification? ▫Complex concepts  Computer table -> A laptop computer is on a coffee table. Low recall. Caused by: ▫Synonymy, different words with similar meanings:  Student and Pupil ▫Related concepts:  Color -> Red, Blue  Car -> Volvo, FIAT, BMW Synonymy Polysemy

3. Scalability problem Current web is a huge repository of documents Number of documents keeps growing significantly ▫Making difficult to locate relevant documents Web is a highly dynamic system ▫Peers are continually joining and leaving the network All these makes the search problem complex.

4. 4 Concept Search Goal: To extend syntactic search (address problems) ▫address the ambiguity problem of NL ▫make uses of related complex concepts ▫should not be worse than syntactic search!!! IR_System = Moving from Syntactic IR to C-Search does not require the introduction of new data structures or retrieval models CSearch reuses retrieval models and data structures of syntactic search ▫words (W) are substituted with complex concepts (C) ▫syntactic matching WMatch is substituted with semantic matching When no semantic information is available, CSearch reduces to syntactic search Fausto Giunchiglia, Uladzimir Kharkevich, and Ilya Zaihrayeu. Concept search. In Proc. of ESWC'09, Lecture Notes in Computer Science. Springer, 2009.

5. Words To Complex Concepts: Extract phrases ▫Descriptive phrase : ▫E.g., A little dog or a huge cat Convert NL phrases to Complex formulas ▫Complex concepts are computed by analyzing meaning of the words and phrases. ▫Expressed in a propositional Description Logic (DL) ▫E.g., (little-4 ⊓ dog-1) ⊔ (huge-1 ⊓ cat-1) Lack of background knowledge: ▫Sometimes it is not possible to find a concept for a word. ▫=> word is used as the identifier for a concept

6. Query answer E.g., A(big-1 ⊓ animal-1, T) = D1 (huge-1 ⊓ white-1 ⊓ elephent-1) CSearch uses the following three methods to access the background knowledge T, stored on a single peer : ▫getConcepts(W) - returns a set of all the possible meanings (atomic concepts A) for word W. ▫getChildren(A) - returns a set of all the more specific atomic concepts of the given atomic concept A in T. ▫getParents(A) - returns a set of all the more general atomic concepts of the atomic concept A in T. Syntactic matching to semantic matching

7. P2P CSearch Main idea is to extending CSearch to address the scalability problem ▫Reasoning extended single BK T to the distributed BK TP2P ▫Centralized inverted index to distributed index build on top of DHT.

8. Distributed Background Knowledge(DBK) Atomic concepts are indexed by words using the DHT 'put' operation ▫e.g., put(canine, {canine-1, canine-2}). Every atomic concept is indexed by related atomic concepts+ their relations. DHT 'put' operation is modified to put(A, B, Rel), ▫e.g., put(canine-2, dog-1, ' ⊑ '), put(canine-2, carnivore-1, ' ⊒ '). Getting data from DBK ▫getConcepts(W), getChildren(A) and getParents(A) are implemented by using are implemented by using the DHT 'get' operation ▫we modified DHT 'get‘ operation get(A, Rel) ▫e.g., getChildren(A) = get(W, ‘ ⊑ '), getParents(A) = get(W, ‘ ⊒ ').

9. Indexing Indexing of documents is performed as follows. ▫Every peer computes a set of atomic concepts A which appear in the representations of peer's documents. ▫For every atomic concept A, the peer computes a set of documents d which contain A. ▫For every pair the peer computes a set S(d, A) of all the document complex concepts Cd in d, which contain A. ▫For every A, the peer sends document summaries corresponding to A, i.e., pairs, to a peer pA responsible for A in DBK. ▫The peer pA indexes these summaries using the local CSearch.

10. Retriveval Step 1 : A peer pI initiates the query process for query concept Cq and initialize the query answer QA. Step 2 : For every conjunctive component ⊓ Aq in Cq, pI selects concept A in ⊓ Aq with the smallest number of more specific atomic concepts. For every selected A, Cq is propagated to the peer pA responsible for A. Step 3 : pA receives the Cq and locally computes a set of documents which belong to the query answer. The results are sent directly to pI. On receiving new results, pI merges them with QA. Step 4 : pA computes a set Cms of all more specific atomic concepts B which are directly connected to the given atomic concept A in TP2P. Cms is computed by querying locally stored more specific concepts. Step 5 : pA propagates Cq to all the peers pB responsible for concepts B in Cms, i.e., Step 2 is repeated on all pB.

11. Example query answering

12. Conclusion & Future work P2P CSerarch addresses the scalability problem of CSerarch and the ambiguity problem of natural language in P2P syntactic search. Future work includes: ▫Development of techniques which can control the quality of a user input and in general to control the quality of DBK; ▫Development of document relevance metrics based on both syntactic and semantic similarity of query and document descriptions; ▫Evaluating the efficiency of the proposed solution.

13. 13 Thank You! To read more: Fausto Giunchiglia, Uladzimir Kharkevich, and Ilya Zaihrayeu. Concept Search. In Proc. of ESWC'09. Fausto Giunchiglia, Uladzimir Kharkevich, S.R.H Noori P2P Concept Search. Poster at SemSearch 2009 workshop.