Action Classification in Action Ontology Building Using Robot-Specific Texts
Keywords:Verb sense disambiguation, verb context classification, su-pervised machine learning.
Instructions written in human-language cause no perception problems for humans, but become a challenge when translating them into robot executable format. This complex translation process covers different phases, including instruction completion by adding obligatory information that is not explicitly given in human-oriented instructions. Robot action ontology is a common source of such additional information, and it is normally structured around a limited number of verbs, denoting robot specific action categories, each of them characterized by a certain action environment. Semi-manual action ontology building procedures are normally based on domain-specific human-language text mining, and one of the problems to be solved is the assignment of action categories for the obtained verbs. Verbs in English language are very polysemous, therefore action category, referring to different robot capabilities, can be determined only after comprehensive analysis of the verb’s context. The task we solve is formulated as the text classification task, where action categories are treated as classes, and appropriate verb context – as classification instances. Since all classes are clearly defined, supervised machine learning paradigm is the best selection to tackle this problem.
We experimentally investigated different context window widths; directions (context on the right, left, both sides of analyzed verb); and feature types (symbolic, lexical, morphological, aggregated). All statements were proved after exploration of two different datasets.The fact that all obtained results are above random and majority baselines allow us to claim that the proposed method can be used for predicting action categories. The best obtained results were achieved with Support Vector Machine method using window width of only 25 symbols on the right and bag-of-words as features. This exceeded random and majority baselines by more than 37% reaching 60% of accuracy.