Albanian Part-of-Speech Tagging: Gold Standard and Evaluation

Albanian Part-of-Speech Tagging: Gold Standard and Evaluation Besim Kabashi, Thomas Proisl Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg Professur f¨ur Korpuslinguistik Bismarckstr. 6, 91054 Erlangen, Germany {besim.kabashi, thomas.proisl}@fau.de Abstract In this paper, we present a gold standard corpus for Albanian part-of-speech tagging and perform evaluation experiments with different statistical taggers. The corpus consists of more than 31,000 tokens and has been manually annotated with a medium-sized tagset that can adequately represent the syntagmatic aspects of the language. We provide mappings from the full tagset to both the original Google Universal Part-of-Speech Tags and the variant used in the Universal Dependencies project. We perform experiments with different taggers on the full tagset as well as on the coarser tagsets and achieve accuracies of up to 95.10%. Keywords: Albanian, part-of-speech tagging, gold standard 1. Introduction Albanian is an Indo-European language that is spoken by ap- proximately 5.4 million people. Within the Indo-European family of languages, Albanian constitutes a subgroup of its own; it is a distinct branch on the same level as, for example, the Hellenic, Romance, Slavic or Germanic subgroups. The language has a diverse vocabulary with many loan words due to language contact with Greek, Latin/Italian, Slavic languages and Turkish. Albanian has a rich morphological system and a relatively free word order, similar to, for example, German. Particu- larly challenging, from a pos tagging point of view, are the many multi-word units. An interesting and frequent phe- nomenon are multi-word units that have articles or particles as their first part. These combinations are borderline cases at the lexis-grammar interface. They are certainly more grammatical in nature and written as two separate graphical tokens. Here is one example from standard Albanian: i mir¨e “(the) good one” (masc.) vs. e mir¨e “(the) good one” (fem.). Presence or absence of the preceding article or particle can change the meaning of a word or its part of speech. Albanian is one of the Balkan languages with the least re- sources available. In particular, there is no part-of-speech tagged corpus available,1 let alone one that covers the multi- word phenomena mentioned above. This paper is a step to improve that situation. In the next section, we briefly discuss previous work in the areas of part-of-speech tagging and morphological analysis. In Section 3., we present a new manually annotated gold standard corpus for part-of-speech tagging. The corpus is annotated with a medium-sized tagset that can adequately represent the multi-word phenomena mentioned above. To improve interoperability with existing multilingual tools and resources, we provide mappings to the popular Universal Part-of-Speech Tags, both the Google 1The Albanian National Corpus (http://web-corpora. net/AlbanianCorpus/search/), which consists of roughly 16.7 million tokens, is only annotated with morphological analyses that have not been disambiguated (and only for words known to the morphological analyzer). The AlCo corpus (Kabashi, 2017), which consists of roughly 100 million tokens, is part-of-speech tagged, but is not yet publicly available. and the Universal Dependencies variant (Section 4.). In Sec- tion 5., we perform evaluation experiments with different part-of-speech taggers on the full tagset and …

Content truncated to 500 words