Background Knowledge for Ontology Construction Bla Fortuna Marko

Background Knowledge for Ontology Construction Blaž Fortuna, Marko Grobelnik, Dunja Mladenić, Institute Jožef Stefan, Slovenia

Bag-of-words Computers are used in increasingly diverse ways in Mathematics and the Physical and Life Sciences. This workshop aims to bring together researchers in Mathematics, Computer Science, and Sciences to explore the links between their disciplines and to encourage new collaborations. • There exist various ways of selecting word weights. In our paper we propose a method to learn them! Word Weigts • Documents are encoded as vectors • Each element of vector corresponds to frequency of one word • Each word can also be weighted corresponding to the importance of the word computer 0. 9 mathematics 0. 8 are 0. 01 and 0. 01 science 0. 9 … … Important Noise computer 2 computer 1. 8 mathematics 2 mathematics 1. 6 are 1 are 0. 01 and 4 and 0. 04 science 3 science 2. 7 … …

SVM Feature selection Input: • Set documents • Set of categories • Each document is assigned a subset of categories Output: • Ranking of words according to importance Intuition: • Word is important if it discriminates documents according to categories. Basic algorithm: • Learn linear SVM classifier for each of the categories. • Word is important if it is important for classification into any of the categories. Reference: • Brank J. , Grobelnik M. , Milic. Frayling N. & Mladenic D. Feature selection using support vector machines.

Word weight learning • The word weight learning method is based on SVM feature selection. • Besides ranking the words it also assigns them weights based on SVM classifier. Notation: • N – number of documents • {x 1, …, x. N} – documents • C(xi) – set of categories for document xi • n – number of words • {w 1, …, wn} – word weights • {nj 1, …, njn} – SVM normal vector for j-th category Algorithm: 1. Calculate linear SVM classifier for each category 2. Calculate word weights for each category from SVM normal vectors. Weight for i-th word and j-th category is: 3. Final word weights are calculated separately for each document:

Onto. Gen system • System for semi-automatic ontology construction – Why semi-automatic? The system only gives suggestions to the user, the user always makes the final decision. • • • The system is data-driven and can scale to large collections of documents. Current version focused on construction of Topic Ontologies, next version will be able to deal with more general ontologies. Can import/export RDF. • There is a big divide between unsupervised and fully supervised construction tools. • Both approaches have weak points: – it is difficult to obtain desired results using unsupervised methods, e. g. limited background knowledge – manual tools (e. g. Protégé, Onto. Studio) are time consuming, user needs to know the entire domain. • We combined these two approaches in order to eliminate these weaknesses: – the user guides the construction process, – the system helps the user with suggestions based on the document collection. http: //kt. ijs. si/blazf/examples/ontogen. html

How does Onte. Gen help? By identifying the topics and relations between them: … using k-means clustering: • cluster of documents => topic • documents are assigned to clusters => ‘subject-of’ relation • We can repeat clustering on a subset of documents assigned to a specific topic => identifies subtopics and ‘subtopic-of’ relation Context Topic By naming the topics: … using centroid vector: • A centroid vector of a given topic is the average document from this topic (normalised sum of topic’s documents) • Most descriptive keywords for a given topic are the words with the highest weights in the centroid vector. … using linear SVM classifier: • SVM classifier is trained to seperate documents of the given topic from the other document in the context • Words that are found most mportant for the classification are selected as keywords for the topic

Topic ontology visualization Selected topic Suggestions of subtopics Topic Keywords All documents Outlier detection Topic document

Topic ontology of Yahoo! Finances

Background knowledge in Onto. Gen • All of the methods in Onto. Gen are based on bag -of-words representation. • By using a different word weights we can tune these methods according to the user’s needs. • The user needs to group the documents into categories. This can be done efficiently using active learning. http: //kt. ijs. si/blazf/examples/ontogen. html

Influence of background knowledge • Data: Reuters news articles • Each news is assigned two different sets of tags: Topics view – Topics – Countries • Each set of tags offers a different view on the data Countries view Documents

Links • Onto. Gen: http: //ontogen. ijs. si/ • Text Garden: http: //www. textmining. net/