TLT-CRF: A Lexicon-supported Morphological Tagger for Latin Based on Conditional Random Fields

TLT-CRF: A Lexicon-supported Morphological Tagger for Latin Based on Conditional Random Fields Tim vor der Br¨uck, Alexander Mehler Distributed Secure Software Systems, Text Technology Lab Lucerne University of Applied Sciences and Arts, Goethe-Universit¨at Frankfurt am Main tim.vorderbrueck@hslu.ch, amehler@em.uni-frankfurt.de Abstract We present a morphological tagger for Latin, called TTLab Latin Tagger based on Conditional Random Fields (TLT-CRF) which uses a large Latin lexicon. Beyond Part of Speech (PoS), TLT-CRF tags eight inflectional categories of verbs, adjectives or nouns. It utilizes a statistical model based on CRFs together with a rule interpreter that addresses scenarios of sparse training data. We present results of evaluating TLT-CRF to answer the question what can be learnt following the paradigm of 1st order CRFs in conjunction with a large lexical resource and a rule interpreter. Furthermore, we investigate the contigency of representational features and targeted parts of speech to learn about selective features. Keywords: morphological tagging, rules, CRFs, Latin lexicon 1. Introduction A tagger that determines morphological, inflectional infor- mation for each word beyond its PoS is called a morpho- logical tagger. We describe a morphological tagger (the tagset being employed is a subset of the Stuttgart-T¨ubingen TagSet (STTS)) called TTLab Latin Tagger based on Con- ditional Random Fields (TLT-CRF). It determines the PoS as well as inflectional categories such as number, case, gender, comparison degree, mood, tense and voice. TLT- CRF is based on a CRF model together with a large Latin lexicon. It explores several lexical features (e.g., for de- termining agreement) and, thus, is distinguished from re- lated approaches. To the best of our knowledge, it is the only hybrid morphological tagger that integrates a rule in- terpreter and a statistical model in the area of Latin texts. Rules are useful in cases of data sparseness which do not allow for defining sufficient training data. The source code of the tagger can be obtained from http://prepro. hucompute.org. 2. Related Work There is a multitude of statistical approaches for PoS tag- ging. Almost all of them rely on a statistical model that is trained on annotated examples and then applied to pre- viously unseen texts. These models include decision trees (Schmid, 1994), Hidden Markov Models (Brants, 2000), Maximum Entropy Models (Toutanova et al., 2003), neu- ral networks and structured SVMs. An exception is de- scribed by (Bellegarda, 2010) who employs learning per analogy together with latent semantic analysis (Landauer et al., 1998) (called latent analogy). While there exists a lot of freely available PoS taggers, so-called morpholog- ical taggers are still rare. An exception is (M¨uller and Sch¨utze, 2015) who provide a morphological tagger for several languages that is based on 3rd order CRFs (Lafferty et al., 2001). A hybrid morphological tagger for Czech is introduced by (Spoustov´a, 2008). The employed statisti- cal model is a combination of HMMs, of a maximum en- tropy model and of perceptrons, where the first two models are now seen to be outdated due to problems of including arbitrary features (HMM) and due to the label bias prob- …

Content truncated to 500 words