EmpiriST Corpus 2.0: Adding Manual Normalization, Lemmatization and Semantic Tagging to a German Web and CMC Corpus

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6142–6148 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6142 EmpiriST Corpus 2.0: Adding Manual Normalization, Lemmatization and Semantic Tagging to a German Web and CMC Corpus Thomas Proisl, Natalie Dykes, Philipp Heinrich, Besim Kabashi, Andreas Blombach, Stefan Evert Computational Corpus Linguistics Group Friedrich-Alexander-Universit¨at Erlangen-N¨urnberg Bismarckstr. 6, 91054 Erlangen, Germany {thomas.proisl, natalie.mary.dykes, philipp.heinrich, besim.kabashi, andreas.blombach, stefan.evert}@fau.de Abstract The EmpiriST corpus (Beißwenger et al., 2016) is a manually tokenized and part-of-speech tagged corpus of approximately 23,000 tokens of German Web and CMC (computer-mediated communication) data. We extend the corpus with manually created annotation layers for word form normalization, lemmatization and lexical semantics. All annotations have been independently performed by multiple human annotators. We report inter-annotator agreements and results of baseline systems and state-of-the-art off-the-shelf tools. Keywords: CMC, annotation, resources, normalization, lemmatization, semantic tagging 1. Introduction Manually annotated data are crucial for training and evalu- ating statistical tools such as POS taggers and lemmatizers. The creation of these “gold standards” for corpus annota- tion layers is thus an inevitable (though labour-intensive and tedious) endeavour to make language data accessible. While there is a comparatively large amount of manually annotated data available for English, especially for standard (newspaper) texts, other languages and registers (such as computer-mediated communication, CMC) do not enjoy such great popularity. A notable exception are English Twit- ter data, for which both manually annotated corpora and designated tools have been developed (Ritter et al., 2011; Owoputi et al., 2013; Kong et al., 2014). Our focus is on German web and CMC data. Off-the-shelf natural language processing (NLP) tools trained on newspa- per corpora typically show a relatively poor performance on this kind of out-of-domain data (Giesbrecht and Evert, 2009; Neunerdt et al., 2013). As has been noted before, there are major linguistic differences between CMC and standard Ger- man (Haase et al., 1997; Runkehl et al., 1998; Dietterle et al., 2017; Beißwenger and Pappert, 2018): Computer-mediated communication, and chat communication in particular, has often been described as being “conceptually oral”, i. e. ex- hibiting phenomena typically associated with oral commu- nication. Examples include colloquial or dialectal word forms and constructions and utterances that are not syntac- tically well-formed. Another well-known phenomenon are substitutes for some of the non-verbal signals of oral com- munication, e. g. emoticons or action words (*freu*, from (sich) freuen, ‘to rejoice’). Substitutes for stress and prosody include character repetitions, all caps or simple mark-up (surrounding a word or phrase with asterisks, slashes, un- derscores, etc.). A higher rate of spelling errors (sometimes due to production speed), deliberate creative spellings and the use of CMC-specific acronyms (LOL, ROFL, IMHO) are also often associated with CMC data. In this paper, we describe our additions to the EmpiriST cor- pus (cf. Section 2.), a manually tokenized and part-of-speech CMC Web Total Training 5,109 4,944 10,053 Test 5,237 7,568 12,805 Total 10,346 12,512 22,858 Table 1: Sizes of the EmpiriST corpus and …

Content truncated to 500 words