Shallow Discourse Parsing for Under-Resourced Languages: Combining Machine Translation and Annotation Projection

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 1044–1050 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 1044 Shallow Discourse Parsing for Under-Resourced Languages: Combining Machine Translation and Annotation Projection Henny Sluyter-G¨athje, Peter Bourgonje, Manfred Stede Universit¨at Potsdam, Applied Computational Linguistics Potsdam, Germany firstname.lastname@uni-potsdam.de Abstract Shallow Discourse Parsing (SDP), the identification of coherence relations between text spans, relies on large amounts of training data, which so far exists only for English - any other language is in this respect an under-resourced one. For those languages where machine translation from English is available with reasonable quality, MT in conjunction with annotation projection can be an option for producing an SDP resource. In our study, we translate the English Penn Discourse TreeBank into German and experiment with various methods of annotation projection to arrive at the German counterpart of the PDTB. We describe the key characteristics of the corpus as well as some typical sources of errors encountered during its creation. Then we evaluate the GermanPDTB by training components for selected sub-tasks of discourse parsing on this silver data and compare performance to the same components when trained on the gold, original PDTB corpus. Keywords: machine translation, annotation projection, discourse parsing 1. Introduction Texts are not a random collection of sentences: they are texts because they convey a certain sense of coherence. The uncovering of the coherence relations holding a text to- gether is referred to as the task of discourse parsing. Like many other tasks based on automatically parsing input text in the larger field of Natural Language Processing, proce- dures often rely on the availability of training data anno- tated for the type of information to be extracted. In the case of coherence relations, such annotations are notoriously difficult and time-consuming to obtain, and inter-annotator agreement rates are lower than for many other tasks. As a result, the amount of available training data is compar- atively small, especially for languages other than English (see Section 2.). In this paper, we present a corpus annotated for dis- course relations obtained through automatically translating an existing English corpus (the Penn Discourse TreeBank, henceforth: PDTB, (Prasad et al., 2008)), and using word alignment to project the English annotations on the German target text. The result is the GermanPDTB, a German cor- pus annotated for shallow discourse relations in the (finan- cial) news domain. We provide details on the method used to create this corpus, sum up the key characteristics and use the GermanPDTB to enrich a pre-existing German connec- tive lexicon. In addition, we provide an extrinsic evalua- tion of the corpus using components of a German discourse parser and compare performance of selected (sub-)tasks on GermanPDTB to the original English PDTB. The rest of this paper is structured as follows: Section 2. lists similar corpora for German and other languages. Sec- tion 3. briefly describes the different coherence annotation types in the original PDTB and consequently the German- PDTB. Section 4. explains …

Content truncated to 500 words