This repository contains code used for a project creating typed semantic compositions for Universal Dependencies syntactic trees.
Any dependencies are included in requirements.txt.
The main program is pmbparsing.py. Results in the paper were obtained by running it from the command line on 6 cpus and on PMB 4.0.0, with options -p pmb-4.0.0 -o results -f filepairsfull.txt -n 2 -b data/mappings -m 20000, where data/mappings is the same folder of data patterns used by UD-Boxer to label synsets, pmb-4.0.0 contains the Parallel Meaning Bank data, and filepairsfull is a csv where the program will print pairs of one PMB reference DRS and one computed DRS, for easy cmparisn later.
Evaluation is done through Counter, using some helper functions in counterutils.py. Evaluation results in the paper were obtained by running this from the command line on 8 cpus and with options -i filepairsfull.txt -n 8.
The file clfutils.py provides certain functions that help one work with DRS structures in CLF format. In particular, it provides functions for converting from SBN (the output of the baseline UD-Boxer) to CLf. UD-Boxer's output was evaluated by converting to CLF using these functions and then evaluating using counterutils.py.
The files in /conllus and the file demos.py serve to give demonstrations of the program. See the help option for demos.py for more information. The files in /conllus demonstrate some of the semantic phenomena that the system was designed to capture.
The files reldenotations.csv and postemplates.csv contain, respectively, handwritten denotations for Universal Dependencies relations and handwritten templates for what Universal Dependencies denotations of different parts-of-speech that were use to create the results in the paper. The file tokensubs.csv contains some further information which is mainly useful certain stages of meaning computation.
All other files provide specific tools that help with parts of the meaning computation done by pmbparsing.py. That is,
conlluutils.py- provides some tools for working with Universal Dependencies trees in conllu format (format from the conllu package)postprocessing.py- provides some tools that clean up final computed meanings after most of the meaning composition is done.preprocessing.py- provides some tools that clean up Universal Dependencies trees before most of the meaning composition can be done.sdrt.py- provides an extension to the NLTK DRT package to allow for Segmented Discourse Representation Theory, which we need to represent the relations in the PMB.semtypes.py- provides a class of semantic types and defines their behavior. Labeling with semantic type will help other algorithms decide in what order to compose meanings.treetodrs.py- provides the main brunt of the computation work, taking words and combining them by applying functions defined by relations.
The files in /conllus all come from the GUM corpus, some of them subsequently edited.
The version of Python's NLTK DRT module used here is slightly modified, which was done to prevent certain variable clashes during beta-reduction. The modified file is recorded here as drt-MODIFIED.py in this repository. The only change is an insertion at what once was line 976 of the original drt.py: lines 976-994 of drt-MODIFIED.py are my own doing. Note that NLTK is distributed under the Apache-2.0 license.
The version of Counter (Rik van Noord, Lasha Abzianidze, Hessel Haagsma, and Johan Bos. Evaluating scoped meaning representations. LREC 2018 [PDF]) used here is also slightly modified. This is to facilitate integrating it into a larger Python program. The modified file is included as counder-MODIFIED.py in this repository. Note that Counter is distributed under the MIT license.