Recognizing Sentence-level Logical Document Structures with the Help of Context-free Grammars

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5282–5290 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5282 Recognizing Sentence-level Logical Document Structures with the Help of Context-free Grammars Jonathan Hildebrand1, Wahed Hemati2, Alexander Mehler3 1,2,3Text Technology Lab 1,2,3Goethe-Universit¨at Frankfurt 1mail@jonathanhildebrand.de, 2hemati@em.uni-frankfurt.de, 3mehler@em.uni-frankfurt.de Abstract Current sentence boundary detectors, such as (Azzi et al., 2019; Gonz´alez-Gallardo and Torres-Moreno, 2018), split documents into sequentially ordered sentences by detecting their beginnings and ends. Sentences, however, are more deeply structured even on this side of constituent and dependency structure: they can consist of a main sentence and several subordinate clauses as well as further segments (e.g. inserts in parentheses); they can even recursively embed whole sentences and then contain multiple sentence beginnings and ends. In this paper, we introduce a tool that segments sentences into tree structures to detect this type of recursive structure. To this end, we retrain different constituency parsers with the help of modified training data to transform them into sentence segmenters. With these segmenters, documents are mapped to sequences of sentence-related “logical document structures” (cf. (Power et al., 2003)). The resulting segmenters aim to improve downstream tasks by providing additional structural information. In this context, we experiment with German dependency parsing. We show that for certain sentence categories, which can be determined automatically, improvements in German dependency parsing can be achieved using our segmenter for preprocessing. The assumption suggests that improvements in other languages and tasks can be achieved. Keywords: sentence segmentation, logical document structure, constituency parsing, dependency parsing, CoNLL shared task 1. Introduction Tools for sentence structure analysis, i.e. constituency and dependency parsers (Chen and Manning, 2014; Dyer et al., 2016a; Kiperwasser and Goldberg, 2016; Straka et al., 2016; Nguyen and Verspoor, 2018), operate on sentence units, the boundaries of which are recognized during pre- processing. However, current sentence boundary detectors (such as (Azzi et al., 2019; Gonz´alez-Gallardo and Torres- Moreno, 2018)) operate sequentially; they detect sequences of sentence beginnings and ends without their dependen- cies. In contrast to this, natural language sentences are of- ten more deeply structured: they can, for example, consist of a main clause and several subordinate clauses or other segments (e.g. inserts in parentheses). They can even em- bed entire sentences recursively and then contain several sentence beginnings and ends. This happens, for example, when embedding direct speech into indirect speech, where embedded sentences can even be interrupted and continued later within the embedding sentence. This is illustrated by the following sentence from Kafka’s The Metamorphosis, in which the sentence beginning with Gregor is interrupted, continued later, and followed by a second embedded sen- tence: “Gregor,” it cried – it was the mother –, “it is three-quarter of seven. Weren’t you going some- where?” Even under the assumption that a boundary detector recog- nizes this as a single sentence without terminating a sup- posed first sentence with the point after seven, it would not sufficiently recognize its inner structure, according to which a matrix …

Content truncated to 500 words