Automatic Phrase Recognition in Historical German

Automatic Phrase Recognition in Historical German Katrin Ortmann Department of Linguistics Fakultät für Philologie Ruhr-Universität Bochum ortmann@linguistics.rub.de Abstract Due to a lack of annotated data, theories of historical syntax are often based on very small, manually compiled data sets. To enable the empirical evaluation of existing hypotheses, the present study explores the automatic recogni- tion of phrases in historical German. Using modern and historical treebanks, training data for a neural sequence labeling tool and a proba- bilistic parser is created, and both methods are compared on a variety of data sets. The evalua- tion shows that the unlexicalized parser outper- forms the sequence labeling approach, achiev- ing F1-scores of 87%–91% on modern German and between 73% and 85% on different histori- cal corpora. An error analysis indicates that ac- curacy decreases especially for longer phrases, but most of the errors concern incorrect phrase boundaries, suggesting further potential for im- provement. 1 Introduction In recent years, the availability of ever-larger data sets and increasing computational power have led to major changes in the way language is analyzed. Today, NLP tools can automatically enrich large amounts of text quickly and accurately with linguis- tic annotations needed for commercial or research purposes. When it comes to non-standard data like historical language, though, the availability of mod- els and annotated corpora is still limited compared to modern language and hypotheses are often based on qualitative analyses of very small data sets. For example, Speyer (2011) investigates object order in the middle field of Early New High German sen- tences based on a total of 70 pairs of direct and indi- rect objects from three centuries. Similarly, Light (2012) grounds her study of extraposition, i.e. the movement of elements behind the clause-final verb, on 115 cases of extraposed subjects in one Early New High German bible translation, while Sapp (2014) analyzes 683 extraposed phrases spread over texts from five centuries. Although data-driven qual- itative analyses like these provide valuable insights for linguistic research, they require a lot of manual effort and cannot achieve the same statistical signif- icance as studies of modern language. Recently, there have been several attempts to ad- dress the lack of annotated historical data and pro- vide a basis for the empirical evaluation of exist- ing hypotheses by automatically identifying relevant syntactic units in historical text (e.g. Chiarcos et al., 2018; Ortmann, 2020, 2021). The present paper takes a similar approach and looks explicitly at the units targeted by the qualitative studies mentioned above, namely phrases. In the context of this study, phrases are under- stood as continuous, non-overlapping constituents from a sentence’s parse tree. Since the concrete definition of constituents may vary depending on the annotation scheme and not all constituents are equally relevant for linguistic studies like the ones mentioned above, this paper focuses on four main phrase types: noun phrases (NP), prepositional phrases (PP), adjective phrases (AP), and adverb phrases (ADVP). For each sentence, only the high- est non-terminal nodes of the given types are consid- ered, ignoring …

Content truncated to 500 words