Prototype-based Active Learning for Lemmatization

International Conference RANLP 2009 - Borovets, Bulgaria, pages 65–70 Prototype-based Active Learning for Lemmatization Walter Daelemans Centre for Dutch Language and Speech (CNTS) University of Antwerp Antwerp, Belgium walter.daelemans@ua.ac.be Hendrik J. Groenewald Centre for Text Technology (CTexT) North-West University Potchefstroom, South Africa handre.groenewald@nwu.ac.za Gerhard B. van Huyssteen Centre for Text Technology (CTexT) North-West University Potchefstroom, South Africa gvhuyssteen@csir.co.za Abstract Annotation of training data for machine learning is often a laborious and costly process. In Active Learning (AL), crite- ria are investigated that allow ordering the unannotated data in such a way that those instances potentially contributing most to the speed of learning can be annotated first. Within this context we explore a new approach that focuses on pro- totypicality as a criterion for the selection of instances to act as training data in order to optimize prediction accuracy. In parallel with the prototype-based active classification (PBAC) approach of Cebron & Berthold (2009), we investi- gate whether the basic PBAC assumption rings true for lin- guistic data. The NLP task we address is lemmatization, the reduction of inflected word forms to their base-form. We operationalize prototypicality as features (i.e. word frequen- cy and word length) of the already available training data items, and combine this with a measure of uncertainty (en- tropy). We show that the selection of less prototypical in- stances first, provides performance that is better than when data is randomly selected or when state of the art AL me- thods are used. We argue that this improvement is possible due to the fact that language processing tasks have highly disjunctive instance spaces, as there are often few regulari- ties and many irregularities. Keywords Active Learning; Prototype Theory; Lemmatization; Afrikaans. 1. Introduction Supervised machine learning techniques are still superior to unsupervised machine learning techniques for many NLP tasks. However, annotation of training data is often a laborious and costly process. In Active Learning (AL) [1,2] criteria are investigated that allow ordering the un- annotated data in such a way that those instances poten- tially contributing most to the speed of learning can be annotated first. Rather than relying on random samples to act as instances in training data, AL entails the selection of instances to act as training data in order to optimize prediction accuracy. Ideally, AL leads to the creation of a supervised learning classifier at a fraction of the annota- tion effort needed when selecting new training items to be annotated randomly, and without loss in accuracy. Also in domain adaptation, AL has been proposed [3] as a feasible approach. Research on AL centers around the develop- ment and comparison of different approaches that could be used to order the available unannotated examples in such a way that those selected first are the ones most in- formative for learning. Another research area is the design of suitable stopping criteria. In this paper, we explore a new approach that focuses on prototypicality (i.e. the degree to which some examples are better, more representative examples of a category than …

Content truncated to 500 words