15th NODALIDA, Joensuu, May 20-21, 2005
Joakim Nivre, Johan Hall, Jens Nilsson: MAMBA meets TIGER: reconstructing a Swedish treebank from antiquity
Treebanks have become an essential resource for the development, optimization and evaluation of broad-coverage syntactic parsers. Given that the number of languages for which treebanks are available is growing steadily, there is a remarkable lack of treebank resources for the Nordic languages with the notable exception of Danish, which is blessed with not only one but two treebanks of substantial size (the VISL Arboretum and the Danish Dependency Treebank). For Swedish, the lack of resources is especially surprising, since some of the earliest examples of syntactically annotated corpora, Talbanken in the 70's (Einarsson 1976a, 1976b) and SynTag in the 80's (Järborg 1986), were based on Swedish data. Talbanken contains close to 300,000 words of both written and spoken Swedish, manually annotated with partial phrase structure and grammatical functions according to the MAMBA scheme (Teleman 1974), and was a very impressive achievement at the time of its creation. By modern standards, however, Talbanken is probably best characterized as a "proto-treebank", since the annotation format makes it rather difficult to use with contemporary parsers and treebank tools.In this paper, we report on a project aiming at the reconstruction of Talbanken in a modern framework. The purpose of the project is twofold. First of all, by converting Talbanken to a more usable format, we want to create a useful resource for research on syntactic parsing of Swedish. Secondly, we want to see whether Talbanken can be used to realize the notion of a theory-supporting treebank, in the sense of Nivre (2003), i.e. a richly annotated source treebank from which we can generate target treebanks in different theoretical frameworks. More precisely, the goal is to extract two treebanks from Talbanken, one with phrase structure annotation, and one with dependency annotation.
The MAMBA annotation scheme consists of two layers, the first being a lexical analysis, consisting of part-of-speech information including morphological features, and the second being a syntactic analysis, in terms of grammatical functions. Both layers are flat in the sense that they consist of tags assigned to individual word tokens, but the syntactic layer also gives information about constituent structure, as exemplified in the annotation of the sentence Genom skattereformen införs individuell beskattning av arbetsinkomster (Through the tax reform, individual taxation of work income is introduced):
*GENOM PR AAPR SKATTEREFORMEN NNDDSS AA INFÖRS VVPSSMPA FV INDIVIDUELL AJ SSAT BESKATTNING VN SS AV PR SSETPR ARBETSINKOMSTER NN SSET . IP IP
The first column of annotation is the lexical analysis, while the second column is the syntactic analysis. The grammatical subject of the sentence is the phrase individuell beskattning av arbetsinkomster (individual taxation of work income), where the head word beskattning (taxation) is assigned the simple tag SS for subject, while the pre-modifying adjective individuell (individual) is tagged SS and AT for adjectival modifier; in the post-modifying prepositional phrase, the noun arbetsinkomster (work income) is tagged SS and ET for post-modifier, while the preposition av (of) is tagged SS, ET and PR for preposition.
The syntactic analysis in MAMBA represents an eclectic approach based on dependency grammar, Diderichsen's field model and immediate constituent analysis (Teleman 1974). This makes it very suitable as a source annotation for conversion to both phrase structure and dependency annotation. The conversion proceeds in three steps:
1. The original flat but multi-layered annotation is converted to a bare phrase structure annotation, i.e. a phrase structure with unlabeled nonterminal nodes, and edges labeled with grammatical functions. This conversion is relatively straightforward given the partially hierarchic annotation exemplified above.
2. The bare phrase structure annotation is extended to a full phrase structure representation by labeling nonterminal nodes with syntactic categories. These categories are not part of the original MAMBA annotation and have to be inferred from other parts of the annotation.
3. The full phrase structure annotation is converted to a dependency annotation using the standard technique with head-finding rules (Magerman 1995, Collins 1996) and preserving grammatical functions as edge labels. Head-finding rules are not part of the original annotation scheme and have to be constructed manually.
The extracted treebanks will be encoded in TIGER-XML, and the phrase structure version will in fact be very close to the NEGRA/TIGER annotation scheme for German (Brants et al. 2002), an adaptation of which has previously been applied to Swedish by Volk and Samuelsson (2004). This scheme allows discontinuous phrases, which do occur in the MAMBA annotation, and has edges labeled with grammatical functions. The dependency version, which consists of terminal nodes connected by edges labeled with grammatical functions, will also be encoded in TIGER-XML, following the guidelines endorsed by the Nordic Treebank Network (Kromann 2004). The reconstructed treebank, in both versions, will be made freely available for research and educational purposes. We are planning a first release in connection with NODALIDA 2005.
REFERENCES
Brants, S., Dipper, S., Hansen, S., Lezius, W. and Smith, G. (2002) TIGER Treebank. In Hinrichs, E. and Simov, K. (eds.) Proceedings of the First Workshop on Treebanks and Linguistic Theories, pp. 24-42.
Collins, M. (1996) A New Statistical Parser Based on Bigram Lexical Dependencies. In Proceedings of the 34th Annual Meeting of the Association for Computational Linguistics (ACL), pp. 184--191.
Einarsson, J. (1976a) Talbankens skriftspråkskonkordans. Lund University, Department of Scandinavian Languages.
Einarsson, J. (1976b) Talbankens talspråkskonkordans. Lund University, Department of Scandinavian Languages.
Magerman, D, M. (1995) Statistical Decision-Tree Models for Parsing. In Proceedings of the 33rd Annual Meeting of the Association for Computational Linguistics (ACL), pp. 276-283.
Nivre, J. (2003) Theory-Supporting Treebanks. In Nivre, J. and Hinrichs, E. (eds.) Proceedings of the Second Workshop on Treebanks and Linguistic Theories. Växjö University Press, pp. 117-128.
Kromann, M. T. (2004) Nordic Treebank Network TIGER-XML: Proposals for Extensions and Conventions in TIGER-XML within the Nordic Treebank Network. September 1, 2004. URL: http://www.id.cbs.dk/mtk/ntn/tiger-xml.html.
Teleman, U. (1974) Manual för grammatisk beskrivning av talad och skriven svenska. Lund: Studentlitteratur.
Volk, M. and Samuelsson, Y. (2004) Bootstrapping Parallel Treebanks. In Proceedings of the 5th International Workshop on Linguistically Interpreted Corpora, pp. 63-70.
| nodalida2005@joensuu.fi Last modified: Sat Apr 16 10:43:22 EEST 2005 |