From Lexical Functional Grammar to Enhanced Universal Dependencies: Linguistically informed treebanks of Polish

Syntactically annotated corpora, or ‘treebanks’, belong to the most heterogeneous kinds of linguistic resources. ey differ not only in the general kind of approach they adopt (constituency or dependency), but also in the number of representation levels they assume (often one, but sometimes two or more) and in the extent to which they follow an established linguistic theory (if at all). Also, even within one kind of approach, the representation of a particular phenomenon may differ widely between treebanks (see, e.g., Popel et al. 2013 for the treatment of coordination in various dependency treebanks).In treebank development, there is a clear tension between theoretical accuracy within a treebank and utilitarian consistency between treebanks of the same or different languages. On the one hand, utterances should be annotated with linguistically accurate and precise descriptions, and one way to achieve this is by following a specific linguistic theory, one with a welldefined terminology, good formal background and a body of carefully justified analyses of many phenomena of typologically diverse languages. An example of such a theory is Lexical Functional Grammar (LFG; Bresnan 1982; Dalrymple 2001; Bresnan et al. 2015; Dalrymple et al. 2018). However, LFG is not the only theory of this kind, and even within one theory, similar phenomena may receive very different representations, reflecting different traditions or different weights assigned to pieces of evidence supporting one or another analysis. So this theoretically-oriented approach to treebank development inevitably leads to the creation of treebanks with very diverse annotation schemes, which are often comprehensible only to a limited number of followers of a given linguistic theory.On the other hand, especially in the context of multilingual natural language processing (NLP), treebanks should ideally follow a common annotation scheme, one that is intelligible to a much broader group of treebank consumers than professional linguists working within a given theory. Moreover, similar phenomena and constructions should receive analogous representations, even if there are subtle – from the point of view of practical applications – differences suggesting dissimilar analyses. A recent attempt at such a comprehensive syntactic annotation scheme is Universal Dependencies (UD; http://universaldependencies.org/, Nivre et al. 2016). As a practical solution, UD aims at providing a maximally simple syntactic representation, one that is useful for various NLP applications, even if at the cost of linguistic precision.This monograph presents two treebanks of Polish which follow the two approaches, as well as the procedure of converting one to the other. Part I describes an LFG treebank, which – given that each utterance is annotated not only with a constituency tree but also with a nonarboreous functional structure – is called ‘structure bank’ below. Both structures adhere to the principles of Lexical Functional Grammar, but many aspects of the two representations are specific to Polish and to the LFG grammar which underlies the treebank (see Chapter 1); the role of particular attributes occurring in functional structures is described and illustrated in Chapter 2, while the role of different labels of syntactic nodes in constituency structures is explained in Chapter 3.Part II describes the procedure of converting this LFG structure bank to a UD treebank. The input to the conversion, an intermediate representation, and the output are presented in Chapter 4. The following Chapter 5 discusses some differences in tokenisation between the two resources. Further, Chapter 6 is devoted to the differences between the morphosyntactic levels of the two treebanks. In order to comply with UD guidelines, it has been necessary to infer grammatical classes (e.g., that of determiner) and syntactic categories (e.g., that of mood) which are not explicitly represented in the input LFG structure bank. Conversely, it has also been useful to add to the usual UD categories a few language-specific features in order to preserve detailed information available in the input (e.g., that of the three masculine ‘subgenders’ or emphatic forms of some broadly pronominal lexemes). Finally, the longest chapter of this part, Chapter 7, presents – in excruciating detail – the two stages of the conversion of syntactic LFG structures to dependency representations assumed in UD. First, the derivation of a dependency representation which closely mirrors the input LFG structures is described in Section 7.1. Second, the consecutive transformations of this intermediate representation resulting in the final fully UD-compliant structure are discussed in Section 7.2.Part III consists of the sole Chapter 8, which offers a stand-alone presentation of the resulting UD trebank of Polish. Apart from describing the kinds of morphosyntactic and syntactic information available in the treebank, it also characterises the underlying data and gives quantitative information about the size of the corpus and the kinds of texts it contains. As this is not the first UD treebank of Polish, this chapter also contains a comparison of this LFG-derived UD treebank to an earlier treebank of Polish, itself the result of (a few steps of) conversion from a constituency treebank. The most conspicuous difference – apart from the larger size of the LFG-derived treebank – is the fact that the treebank presented here makes extensive use of the enhanced representation scheme made available in the current version 2 of Universal Dependencies. As discussed in Chapter 9, concluding the monograph, this feature of UD makes it possible to preserve various kinds of syntactic information normally not expressible in simple dependency trees, including information about grammatical control and about sharing of dependents in coordinate structures.While the concluding Chapter 9 presupposes some knowledge of the material of the previous chapters, the three main parts of this monograph are meant to be self-contained. This is especially true about Parts I and III, which present the two resources in a way that does not assume the knowledge of the other resource or of the conversion procedure. An attempt was also made to present the conversion procedure in Part II independently of the presentation of the two resources, although prior exposition to LFG and UD will certainly make reading this part easier.Both the creation of the original LFG corpus and the conversion into UD have been partially supported by the Polish Ministry of Science and Higher Education within the CLARIN ERIC programme 2016–2018 (http://clarin.eu/). The original LFG structure bank has been developed under the supervision of Agnieszka Patejuk and has been converted to UD by Adam Przepiórkowski, in collaboration with Agnieszka Patejuk. We would like to cordially thank Joakim Nivre and Dan Zeman for their infinite patience in answering a myriad of diverse UDrelated questions during the development of this treebank, and the reviewers of this monograph, Paul Meurer and Stephan Oepen, for their comments, which led to some importantimprovements. The data, lemmata and original morphosyntactic tags come from the manually annotated subcorpus of the National Corpus of Polish (http://nkjp.pl/), whose development – within a project led by Adam Przepiórkowski – was financed by the Polish Ministry of Science and Higher Education in 2007–2011, and – to a lesser extent – from the Corpus of 1960s Polish (http://clip.ipipan.waw.pl/PL196x). Parts of this monograph were written and revised during our fellowship at the Oslo Center for Advanced Study (CAS) at the Norwegian Academy of Science and Leers (https://cas.oslo.no/), within the group “SynSem: From Form to Meaning – Integrating Linguistics and Computing” led by Dag Haug and Stephan Oepen. It is very possible that, if not for our involvement in CAS, neither the UD treebank of Polish presented here, nor this monograph, would ever see the light of day.