In July 2012, on the occasion of the 50th anniversary of the Association for Computational Linguistics (ACL), a special workshop was devoted to “issues related to preserving, analysing and exploiting the scientific heritage of the ACL”. Spurred by this event, a community effort emerged, aiming to provide the full ACL Anthology as a high-quality corpus with rich markup, following the TEI P5 guidelines. The goals of this initiative are threefold: (a) to provide a shared resource for experimentation on scientific text; (b)to serve as a basis for advanced search over the ACL Anthology, based on textual content and citations; and, by combining the aforementioned goals, (c)~to present a showcase of the benefits of natural language processing to a broader audience.

This community effort, dubbed the ACL Anthology Corpus (AAC), continues the tradition of projects like the ACL Anthology Reference Corpus (ACL ARC), ACL Anthology Network (ANN), or the ACL Anthology Searchbench (ACL ASB).

As of mid-2012 at least, AAC is in its early stages in many respects. This page aims to provide a stable on-line access point to AAC-related information, launched at the time of the 2012 Annual Meeting of the ACL.

For the time being, this page has been redirected into the wiki of the DELPH-IN community (whose members stand behind the ACL Anthology Searchbench). However, for long-term validity, please always refer to the canonical URL for the AAC initiative as: http://www.delph-in.net/aac/.

The main idea behind what was called a Contributed Task at ACL 2012 was to combine techniques from Optical Character Recognition (OCR) and ‘native’ text stream extraction from born-digital PDF documents, to let alternate approaches complement each other and aim for the creation of a rich XML format. This method would rely on OCR exclusively only in cases where no born-digital PDFs are available—in case of the ACL Anthology mostly papers published before the year 2000.

Details on specific sub-tasks and examples of challenges to high-quality text and format extraction from the ACL Anthology, for the time being, remain available from the ACL 2012 pages. However, we expect to migrate all relevant information and files to a more permanent communication infrastructure (see below), hence please monitor this page for updates.

Although many of the subtasks sketched above did not find volunteers in this round, the 2012 Contributed Task, in our view, is an on-going, long-term community endeavor. Results to date, if nothing else, confirm the general suitability of (a) using TEI P5 markup as a shared target representation and (b) exploiting the complementarity of OCR-based techniques (Schäfer & Weitz, 2012), on the one hand, and direct interpretation of born-digital PDF files (Berg, et al., 2012), on the other hand. Combining these approaches has the potential to solve the venerable challenges that stem from inhomogeneous sources in the ACL Anthology—e.g. scanned, older papers and digital newer papers, generated from a broad variety of typesetting tools.

However, as of mid-2012 there still is no ready-to-use, high-quality corpus that could serve as a shared starting point for a broad range of Anthology-based NLP activities. In fact, we remain slighly ambivalent about our recommendations for utilizing the current state of affairs and expected next steps—as we would like to avoid much work getting underway with a version of the corpus that we know is unsatisfactory. Further, obviously, versioning and well-defined release cycles will be a prerequisite to making the corpus useful for comparable research, as discussed by Bird, et al. (2008).

In a nutshell, we see several possible avenues forward. For the ACL 2012 Contributed Task, we collected various views on the corpus data (as well as some of the source code used in its production) in a unified SVN repository. Following the open-source, crowd-sourcing philosophy, we make this repository openly available to all interested parties for future development (seel below), further augmenting it with support infrastructure like, for example, a mailing list and shared wiki.

At the same time, our experience from the past months suggests that it is hard to reach sufficient momentum and critical mass to make substantial progress towards our long-term goals, while contributions are limited to loosely organized volunteer work. A possibility we believe might overcome these limitations would be an attempt at formalizing work in this spirit further, for example through a funded project (with endorsement and maybe financial support from organisations like the ACL, ICCL, AFNLP, ELRA, or LDC). To decide on the way forward, we will solicit comments and expressions of interest during ACL 2012, including of course from the R50 workshop audience and participants in the Contributed Task.

All supporting files (documents and software) for the AAC initiate are distributed through a shared Subversion (SVN) repository. For read-only access, please use a command like the following:

  svn co http://svn.delph-in.net/aac/trunk aac

Note that the repository, in mid-2012, contains some 55,000 files, so please allow adequate time (between ten minutes and a couple of hours, depending on network throughput) for the initial checkout.

Use of SVN is intended to ease the creation of stable release snapshots or development branches for individuals or groups of collaborating developers. To gain write access to the repository, please contact Stephan Oepen.

In case you would like to participate in on-going and future AAC-related efforts, please contact Ulrich Schäfer.

To receive updates on the ACL Anthology Corpus initiative, please subscribe to our (very low-traffic) mailing list.