Parsers Know Best: German PP Attachment Revisited

Proceedings of the 28th International Conference on Computational Linguistics, pages 2049–2061 Barcelona, Spain (Online), December 8-13, 2020 2049 Parsers Know Best: German PP Attachment Revisited Bich-Ngoc Do Leibniz ScienceCampus Universit¨at Heidelberg Heidelberg, Germany do@cl.uni-heidelberg.de Ines Rehbein Data and Web Science Group Universit¨at Mannheim Mannheim, Germany ines@informatik.uni-mannheim.de Abstract In the paper, we revisit the PP attachment problem which has been identified as one of the major sources for parser errors and discuss shortcomings of recent work. In particular, we show that using gold information for the extraction of attachment candidates as well as a missing compar- ison of the system’s output to the output of a full syntactic parser leads to an overly optimistic assessment of the results. We address these issues by presenting a realistic evaluation of the po- tential of different PP attachment systems, using fully predicted information as system input. We compare our results against the output of a strong neural parser and show that the full parsing approach is superior to modeling PP attachment disambiguation as a separate task. 1 Introduction Prepositional phrase (PP) attachment disambiguation, the task of identifying the correct attachment site for each preposition in the syntax tree, has often been described as the canonical case of structural ambiguity in NLP, with crucial impact on semantic interpretation. What makes PP attachment such a challenging task is that morpho-syntactic information is often insufficient to resolve the ambiguity, in particular for (semi-)free word order languages, and additional semantic information or even world knowledge is needed. Even though the PP attachment problem has been studied since the nineties (Hindle and Rooth, 1993; Brill and Resnik, 1994; Ratnaparkhi et al., 1994), it is still one of the hardest problems for syntactic parsing. For constituency parsing of English, Kummerfeld et al. (2012) showed that PP attachment errors are the largest error category across all parsers included in their evaluation. This also holds for dependency parsing. In experiments with the biaffine parser of Dozat and Manning (2017) on German, we found that 21.8% of the unlabeled attachment errors are due to incorrect PP attachments. To identify the correct attachment site for each preposition, the parser either needs to see these head- PP pairs during training or has to be able to deduce the correct attachment based on seen examples. We can improve lexical coverage by adding more training data, however, treebanking is extremely time- consuming and requires linguistic expertise. In addition, most PPs are adjuncts, which means that they can choose their heads more freely (in comparison to arguments) and it is thus harder to achieve sufficient coverage. A different approach tries to improve PP attachment accuracy by modeling the problem as a separate task. This has the advantage that we do not need fully annotated parse trees as input, and that such a setup makes it easy to integrate a wide range of heterogeneous features. Many studies have tried to solve this task, however, only few have shown that their system is able to improve the output of a …

Content truncated to 500 words