DeepBank

This page describes the DeepBank project, which has produced linguistically rich syntactic and semantic annotations of Wall Street Journal newspaper text, now available in several output formats, as version 1.0. For downloads and details of the output formats, please see this MetaShare site.

Background

The DeepBank project has the goal of annotating the one million words of 1989 Wall Street Journal text (the same set of sentences annotated in the original Penn Treebank project) with the English Resource Grammar, augmented with a robust approximating PCFG for complete coverage. DeepBank contains rich linguistic annotation on both syntactic and semantic structures of the sentences and is available in a variety of representation formats (see the description on formats below).

The project is hosted at the Department of Computational Linguistics of Saarland University and the Language Technology Lab of the German Research Center for Artificial Intelligence in Saarbrücken, Germany, and in close collaboration with CSLI Stanford. Other institutes, including (but not limited to) Humboldt University of Berlin and University of Oslo, have also contributed to the development and release of the resource. In the long term, the DeepBank will be further supported by the DELPH-IN community with updates and maintenance.

Development Methodology

The project is technically built on top of resources developed in the long-term grammar and software engineering effort maintained under the collaborative umbrella of DELPH-IN. Following earlier practice in the development of Redwoods treebanks, manual annotations are done using the discriminant-based treebanking environment provided by [incr tsdb()] to identify the correct full analysis among the candidate analyses proposed by the English Resource Grammar.

For the first public release of DeepBank, most of the data has gone through at least two rounds of human annotation with independent annotators. Also, the linguistic analyses in DeepBank were made independently from the previous treebank annotations of the same data (i.e. PTB), distinguishing it from PTB-derived treebanks including the Enju HPSG treebank, CCGBank, and the CoNLL syntactic dependency bank, to name a few.

For completeness of the annotations over the full corpus, the public release of DeepBank also includes analyses (trees) licensed by an approximating PCFG for the sentences of the WSJ corpus not correctly analysed by the current version of the ERG. Semantic structures are also composed robustly for these sentences, which comprise some 15% of the 50,000-sentence total.

Stages of Development

The development of DeepBank started in the fall of 2008 as an internally funded project at the Department of Computational Linguistics, Saarland University and the LT-Lab of DFKI, under the supervision of Valia Kordoni and Yi Zhang. Thanks to the partial financial support of the Erasmus Mundus European Masters Program in Language and Communication Technologies (LCT), part-time student annotators were employed and trained for the first round of annotation. Dan Flickinger, the main ERG developer, has provided grammar updates throughout the project. He also went through a thorough (second) round of annotation updates to arrive at the first public release of DeepBank. Both the ERG and DeepBank have significantly evolved over the years of the project, but the dynamic nature of the annotation method has kept them synchronized through the update cycles.

By the summer of 2012, the development of DeepBank reached a mature stage where a significant amount of the data had gone through two rounds of careful annotation. The resource was made available for internal DELPH-IN review (alpha release) by several sites, including the University of Oslo, the University of Washington, Melbourne University, the University of Barcelona, then Bulgarian Academy of Science, and the University of Lisbon. Many suggestions and detailed feedback helped in preparation for the first full public release of DeepBank.

At the end of November 2012, a substantial portion of DeepBank (WSJ sections 00-15) was made open for public preview through a beta release announced at TLT in Lisbon. The beta version (v0.9) is still available for download upon request. This beta-release only includes annotation for WSJ sections 00-15 in the original [incr tsdb()] format. Further sections and other formats are included in the public release, v1.0, described below.

Version 1.0: Available Formats

The public release (v1.0) of DeepBank includes annotation in multiple formats. The combination of the raw [incr tsdb()] profiles with a corresponding version of the ERG enables automatic reconstruction of all detailed analyses. The HPSG derivations and the MRSes are recorded in these profiles and can be extracted directly.

For convenience of usage, DeepBank is also available in other representation formats (though not all details are preserved in the converted representations), including the (modified) Penn-style constituent tree representation with labeled brackets, and the CoNLL-style syntactic and semantic dependency representation with tabbed format. The conversion software is available to the public and maintained collaboratively between Oslo and Saarbrücken.

Downloads are available through MetaShare at this site.

For further information or feedback, please feel free to subscribe to the mailing list, or contact the developers.

Acknowledgements

The primary work on the development of this resource was initiated and carried out through CoLi and the DFKI in Saarbrücken, including project organization and management, the first rounds of treebank construction, the PCFG approximations for non-ERG-parsed items, and some exports to alternative formats. Additional contributions were made at the Center for the Study of Language and Information (CSLI) at Stanford University, for improvements to the ERG and a second round of treebank annotations; and at the University of Oslo in several enabling roles, including preparation of the raw text, tokenization technology and rules, parsing, support for the annotation tools, partial funding of the annotation effort at Stanford, and packaging of the resulting annotations. Ongoing support and expansion of this resource will be provided by these three institutions, together with Humboldt University as a participating institution in the ongoing QTLeap project.

We are grateful to the Erasmus Mundus European Masters Program in Language and Communication Technologies (LCT, EM Grant Number: 2007-0060) for financial support of the project in Saarbrücken, and to the WeSearch: Language Technology for the Web project for support of the contributions made at the University of Oslo and, in part, funding of work at Stanford.

We are equally grateful to the following student annotators for their diligent and patient work. All remaining errors in the treebank are of course ours.

The involvement of Yi Zhang in the project is also partially sponsored by the German Cluster of Excellence on "Multimodal Computing and Interaction" (MMCI) funded by the DFG, and the Deependance project funded by BMBF (01IW11003).

Publications

  1. Dan Flickinger, Valia Kordoni and Yi Zhang. DeepBank: A Dynamically Annotated Treebank of the Wall Street Journal. In Proceedings of TLT-11, Lisbon, Portugal, 2012.

  2. Stephan Oepen, Dan Flickinger, Kristina Toutanova, and Christopher D. Manning. LinGO Redwoods: A Rich and Dynamic Treebank for HPSG. In Journal of Research on Language and Computation 2.4, pages 575-596. 2004.
  3. Angelina Ivanova, Stephan Oepen, Lilja Øvrelid, and Dan Flickinger. Who did what to whom? a contrastive study of syntacto-semantic dependencies. In Proceedings of the Sixth Linguistic Annotation Workshop, pages 2–11, Jeju, Republic of Korea, 2012.
  4. Yi Zhang and Hans-Ulrich Krieger. Large-scale corpus-driven PCFG approximation of an HPSG. In Proceedings of the 12th International Conference on Parsing Technologies, pages 198–208, Dublin, Ireland, 2011.
  5. Yi Zhang, Valia Kordoni. Discriminant Ranking for Efficient Treebanking. In Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), Beijing, China, 2010.
  6. Valia Kordoni, Yi Zhang. Disambiguating Compound Nouns for a Dynamic HPSG Treebank of Wall Street Journal Texts. In Proceedings of the Seventh conference on International Language Resources and Evaluation (LREC'10), Malta, 2010.
  7. Valia Kordoni, Yi Zhang. Annotating Wall Street Journal Texts Using a Hand-Crafted Deep Linguistic Grammar. In Proceedings of the Third Linguistic Annotation Workshop, Singapore, 2009.

DeepBank (last edited 2014-01-16 19:16:51 by DanFlickinger)

(The DELPH-IN infrastructure is hosted at the University of Oslo)