Skip to content

VistaExtractionTop

Jump to bottom
anonymous edited this page Oct 9, 2011 · 7 revisions

Propbank/Treebank extraction

The lkb2standard tool extracts the constituency representation of the parse tree.

Getting and running the tool

You may download the tool here.

The following libraries are required:

common
Helper functions for handling the tsdb exported files. This is included in the download.

tregex
For searching/manipulating trees. You may find it here.

jargs
Command line handling. You may find it here.

To run the tool, use the following command line:

java \
 -cp /path/to/src:/path/to/libraries/* \
 pt.ul.fc.di.nlx.lkb2standard.Simplifier \
 -c <CONFIGFILE> -v <VERSION>

Output to stdout, unless the OUTPUT field in the configuration file is defined.

Additional comments

The tool needs to find a variety of data. The relevant paths are defined in the configuration file. These paths are closely tied to the way our repository is organized, but they can be redefined/overridden in the configuration file. The VERSION number (provided through the command line above) is used when piecing together the path in order to select specific versions of the dataset.

The most important fields in the configuration file are:

ADJUDICATION
The directory that contain the test suite subdirectories which, in turn, store the .gz files (one per sentence) with the exported data.

CINTIL
In the repository there is a directory that contains a subdirectory for each test suite. Each of these subdirectories contains, among many other files, the items.txt file, which stores an annotated version of all the sentences in that test suite.

LEXICON
The file containing the lexicon of the grammar (.tdl format).

The SSTREEB and SUITEGZ fields are Java regular expressions that can be used to run the tool over just those test suites and .gz files matching the given pattern.

When processing a .gz file from a subdirectory in ADJUDICATION, the tool must find the corresponding subdirectory in CINTIL. The mapping of subdirectory names is done in the Java code (file Traverser.java, method sstreebDir2cintilDir). This code consists of a set of string replacement rules that does something like transforming "SSTreeB-bCINT-123-V03-EXPORT" into "cintil/subCintil/cintilSS-123-V03". At a given point, these rules should be given in the configuration file, but currently it requires changing the source code.

The tool looks for an items.txt file in CINTIL. This file is not produced by the grammar (it's specific to our group). The file contains a set of @-separated fields, with one sentence per line (similar to the item file). A single line sample is shown below:

3@CorpusSSNumTB@newspaper@none@1@S@Burnet graduou-se em medicina na Universidade de Melbourne em 1924.@1@22@Burnet/PNM[B-PER] graduou/GRADUAR/V#ppi-3s[O] -se/CL#gn3[O] em/PREP[O] medicina/MEDICINA/CN#fs[O] em_/PREP[O] a/DA#fs[O] Universidade/PNM[B-ORG] de/PREP[I-ORG] Melbourne/PNM[I-ORG] .*//PNT[O]@NLX@2011-07-23

The first field is the sentence identifier, which is the same number that gives the name to the .gz file exported by tsdb. Most of the other fields are ignored by the tool, except for the 7th and the 10th. These are the sentence in plain format and the annotated sentence (in a format used in our group).

Further additional comments

Though I'd love it if this tool was found to be immediately useful to everyone, I'm aware that, as development progressed and the tool became more complex, it also became more dependent on specific details of our grammar (e.g. particular derivation rules and category names). I think that this is inevitable, but perhaps the tool can still be of use for those wanting to tackle a similar task.

More specifically, things may go wrong when applying directly these tools to data from other grammars given this tool has some hard-coded references to:

  • In exported files: names of grammar rules and tags of the tag set specific of our grammar
  • In the lexicon: specific lexical type of our grammar
  • In annotated sentences: reference to specific Portuguese lexical items and specificities of our grammar input internal format

Extra further additional comments

If you need to reference this work, please use the following citation:

To check the design options of such constituency view, you may find useful the following handbook:

To see examples of constituency representation extracted with this tool and how a treebank extracted with its help looks like, you may want to check the online search service over our CINTIL-Treebank, which is found here.

Home | Forum | Discussions | Events

Clone this wiki locally