Paradigm classification in supervised learning of morphology

Human Language Technologies: The 2015 Annual Conference of the North American Chapter of the ACL, pages 1024–1029, Denver, Colorado, May 31 – June 5, 2015. c⃝2015 Association for Computational Linguistics Paradigm classification in supervised learning of morphology Malin Ahlberg Spr˚akbanken University of Gothenburg malin.ahlberg@gu.se Markus Forsberg Spr˚akbanken University of Gothenburg markus.forsberg@gu.se Mans Hulden Department of Linguistics University of Colorado Boulder mans.hulden@colorado.edu Abstract Supervised morphological paradigm learning by identifying and aligning the longest com- mon subsequence found in inflection tables has recently been proposed as a simple yet competitive way to induce morphological pat- terns. We combine this non-probabilistic strat- egy of inflection table generalization with a discriminative classifier to permit the recon- struction of complete inflection tables of un- seen words. Our system learns morphological paradigms from labeled examples of inflection patterns (inflection tables) and then produces inflection tables from unseen lemmas or base forms. We evaluate the approach on datasets covering 11 different languages and show that this approach results in consistently higher ac- curacies vis-`a-vis other methods on the same task, thus indicating that the general method is a viable approach to quickly creating high- accuracy morphological resources. 1 Introduction Use of detailed and sophisticated morphological fea- tures has been found to be crucial for many down- stream NLP tasks, including part-of-speech tag- ging and parsing (Tseng et al., 2005; Spoustov´a et al., 2007). However, creating an accurate wide- coverage morphological analyzer for a new lan- guage that can be used in tandem with other higher- level analyses is an arduous task. Learning word inflection patterns by organizing related word-forms into morphological paradigms based on the longest common subsequence (LCS) found in an inflection table has recently been proposed as a method for supervised and semi- supervised induction of morphological processing tools from labeled data (Ahlberg et al., 2014). Also, the argument that the LCS shared by different in- flected forms of a word—even if discontinuous within a word—corresponds strongly to a cross- linguistic notion of a ‘stem’ has later been advanced independently on grounds of descriptive economy and minimum description length (Lee and Gold- smith, 2014). We used this idea in (Ahlberg et al., 2014) to create a relatively simple-to-implement system that learns paradigms from example inflection tables and is then able to reconstruct inflection tables for un- seen words by comparing suffixes of new base forms to base forms seen during training. The system performs well on available datasets and results in human-readable and editable output. The longest common subsequence strategy itself shows little bias toward any specific morphological process such as prefixation, suffixation, or infixation. Using the model, we argued, a selection of ready-inflected ta- bles could be quickly provided by a linguist, allow- ing rapid development of morphological resources for languages for which few such resources exist. Potentially, however, the model’s commitment to a simple suffix-based learner is a weakness. To assess this, we evaluate a similar LCS-based gen- eralization system with a more refined discrimina- tive classifier that takes advantage of substrings in …

Content truncated to 500 words