Overview

DELPH-IN has a collection of sofware systems that iterpret its descriptive formalisms. There is not one single engine for all needs, however, but rather a toolbox of component systems with different characteristics, for various use cases. Interactive grammar development, for example, requires flexible debugging and visualization tools, whereas batch parsing, say, will capitalize more on premium efficiency, as well as on robustness measures and interfaces that enable application building.

For some grammars, there are Web-accessible on-line demonstrators, e.g. at http://erg.delph-in.net for the ERG. These demonstrators usually utilize a combination of DELPH-IN processing engines, but they are limited in how much computational resources they will have available and how much interactive debugging and visualization they support.

Following is an overview of processing levels in parsing (as implemented in the ERG and the PET parser):

http://www.delph-in.net/2013/tutorial/pipeline.png

TODO: Add schematic of processing pipeline in the realization direction.

Not all DELPH-IN engines implement all sub-formalisms and processing directions. A mildly out-of-date summary by Woodley Packard provides a useful overview.

LKB: Linguistic Knowledge Builder

The LKB (Copestake 2002) has served as the main platform for interactive grammar development in the past 15 or so years. It supports parsing and generation and makes available a range of visualization and debugging tools. However, the LKB lacks some robustness measures (notably chart mapping and unknown word handling in parsing) as well as some more recent algorithmic improvements (e.g. selective unpacking with non-local features). Also, it is about one order of magnitute less efficient than engines optimized primarily for premium run-time performance.

The LKB is written in ANSI Common Lisp, building the Common Lisp Interface Manager (CLIM) for its GUI. To compile the LKB from source including the GUI, Franz Allegro Common Lisp is required. Pre-compiled binaries are available for Linux. The LKB was originally developed by Ann Copestake and since the mid-1990s has been substantially extended and revised for use in DELPH-IN with the help of, among others, John Carroll, Rob Malouf, and Stephan Oepen.

PET: Platform for Efficient Experimentation with HPSG Processing Techniques

PET (Callmeier 2002) has served as the main batch parsing engine and application delivery vehicle for the past 12 or so years. PET provides (almost) no visualization facilities and is typically invoked from the command line, reading inputs from stdin and reporting results to stdout. Furthermore, PET provides interfaces to [incr tsdb()] and a socket-based as well as an XMLRPC server interface.

PET only supports parsing, but (unlike the LKB) it makes available a range of pre-processing and robustness measures that facilitate broad-coverage analysis of running text. These include

  • more accurate characterization in REPP
  • the ability to call out to a PoS tagger to annotate a token sequence
  • chart mapping for lightweight NE recognition and token normalization
  • generic lexical entries
  • lexical filtering (to remove redundant lexical items)
  • selective unpacking with non-local features

PET lacks support for post-parsing idiom filtering (based on MRS templates), however.

PET is written in ANSI C and C++, with a number of (relatively common) external dependencies. Pre-compiled binaries are available through the LOGON tree. PET was originally developed by Ulrich Callmeier; since around 2001, Stephan Oepen, Bernd Kiefer, Yi Zhang, Peter Adolphs, and Rebecca Dridan, among others, have made substantive code contributions.

Other Engines

In 2013, two additional engines exist that implement the DELPH-IN formalism.

The Answer Constraint Engine (ACE; see the AceTop page) has been under development for about ten years, by Woodley Packard, and has recently been released into the DELPH-IN open source repository under the MIT License. ACE is an efficiency-oriented run-time engine, somewhat similar in philosophy to PET, but implements both parsing and generation (and pretty much all sub-formalisms and processing layers sketched above, including idiom filtering).

ACE provides an interface to [incr tsdb()]. It is written in ANSI C with a few external dependencies. Pre-compiled binaries are available from the ACE download page as well as through the LOGON Tree.

Another Grammar Engineering Environment (agree; see the AgreeTop page), implementing the DELPH-IN formalism, was recently released under the MIT open-source license. agree is built on top of the C# and .NET platform and aims to combine advanced GUI capabilities with good run-time effiency, capitalizing heavily on multi-threading. agree supports both parsing and generation and most, though not all of the relevant sub-formalisms and processing layers. The main agree developer is Glenn Slayden, with assistance from Spencer Rarrick.

[incr tsdb()]: Profiling and Treebanking Environment

The tee ess dee bee plus plus ([incr tsdb()]) software is a cross-processor environment to support

  • the maintenance of annotated test suites and test corpora
  • batch processing (using various engines) with detailed records of system behavior
  • interactive, graphical analysis of individual test runs and constrastive analysis of multiple runs
  • discriminant-based treebanking and (semi- or fully automatic) updating of treebanks
  • training and evaluation of discriminative parse and realization ranking models

[incr tsdb()] has been developed since the mid-1990s by Stephan Oepen; it combines some ANSI Common Lisp with ANSI C and Tcl/Tk. Pre-compiled binaries are available for Linux.

LUI: Linguistic User Interface

Parse and Realization Ranking

Last update: 2013-07-30 by EmilyBender [edit]