An Experiment Using Markov Logic Networks to Extract Ontology Concepts from Text Lucas Drumond and Rosario Girardi Federal University of Maranh˜ ao, Computer Science Department, Av. dos Portugueses, s/n 65085-580 S˜ ao Lu´ ıs - MA, Brazil ldrumond@gmail.com,rgirardi@deinf.ufma.br http://gesec.deinf.ufma.br Abstract. Since ontology development is currently an error prone, time consuming and expensive task, ontology based systems suffer from the so called knowledge acquisition bottleneck. One approach for this prob- lem is to provide automatic or semi-automatic support for ontology con- struction. This field of research is known as ontology learning. Many techniques for learning ontologies from text have been proposed, most of them based on statistical learning and natural language processing methods. This work presents an approach for extracting ontology con- cepts from text that combines both ideas through statistical relational learning. A Markov Logic Network as been developed for this task and is described here. Key words: Ontology Learning; Knowledge Acquisition; Statistical Re- lational Learning 1 Introduction Though ontologies hold a great importance in modern knowledge based systems, their development is currently an error prone, time consuming and expensive task. An approach for this problem is the automatic or semi-automatic con- struction of ontologies, a field of research that is usually referred to as ontology learning [1][2]. According to Buitelaar [1] one of the sub-phases of ontology learning is con- cept extraction. Many techniques for concept extraction rely either on statistical analysis [3][4][5] or on linguistic patterns [6][7]. Statistical methods make use of the bag-of-words approach, which assumes that the terms are not correlated, i.e. they do not consider relation between words, represented by their syntac- tic dependencies. Such relations can be used by relational learning techniques by representing them through knowledge representation formalisms such as first order logic. However, relational learning techniques are not able to deal with the noise that arises from polysemy and ambiguity present in natural language texts. This work presents the Probabilistic Relational Concept Extraction (PRECE) technique and investigates the suitability of statistical learning techniques for on- tology learning tasks. Statistical relational learning [8] combines the expressive