This is the GKR4NLI symbolic NLI engine. It computes the inference relation between two sentences. It is based on a version of Natural Logic (Valencia, 1991; MacCartney & Manning, 2007) and on the Graphical Knowledge Representation (GKR; Kalouli & Crouch, 2018). Althought GKR4NLI is a stand-alone system, it was introduced as the symbolic component of Hy-NLI, a hybrid NLI engine. Find more details in our paper:
Kalouli, A.-L., R. Crouch and V. de Paiva. 2020. Hy-NLI: a Hybrid system for Natural Language Inference. In Proceedings of COLING 2020 (link coming soon).
Author/developer: Aikaterini-Lida Kalouli ([email protected]) and Richard Crouch ([email protected])
If you use this software in writing scientific papers, or you use this software in any other medium serving scientists or students (e.g. web-sites, CD-ROMs) please include the above citation.
If you would like to have a quick taste of how GKR4NLI looks like, check out our online demo (note that the online demo presents GKR4NLI as a component of the hybrid system, and not as a stand-alone engine): http://hynli.nlitoolkit.de/
If you are interested in the explainability of GKR4NLI and of our hybrid system, check out our online demo: http://xplainli.nlitoolkit.de/
Copyright 2020 Aikaterini-Lida Kalouli and Richard Crouch. GKR is a free-software discributed under the conditions of the Apache License 2.0, without warranty. See LICENSE file for more details. You should have received a copy of the LICENSE file with the source code. If not, please visit http://www.apache.org/licenses/ for more information.
(this installation assumes you have a Java runtime environment (version 8 or higher, https://www.oracle.com/technetwork/java/javase/downloads/jre8-downloads-2133155.html ) and the Gradle Build tool (https://gradle.org/) in place. If not, please make sure to install those first)
-
Clone this repository into your desired directory:
git clone https://github.com/kkalouli/GKR4NLI.git. -
Download WordNet from http://wordnet.princeton.edu/wordnet/download/. In order to use WordNet, you must configure the files GKR4NLI/sem.mapper/src/main/resources/gkr.properties and GKR4NLI/gnli/src/main/resources/gnli.properties. In particular you must change the following line:
wn_location=/usr/local/Cellar/wordnet/3.1/dict
Change the location and set the directory in which the WordNet dictionary is installed.
- Download and install JIGSAW as described in https://github.com/pippokill/JIGSAW Although you will have downloaded WordNet by now, make sure you follow the JIGSAW instructions for modifying the JIGSAW config file in JIGSAW/resources/wn_file_properties.xml . Also, in the jigsaw.properties file replace the relative paths of wn_file_properties.xml, nlp.tokenModel, nlp.posTagModel and nlp.stopWordFile with the absolute paths of their location in your system. In the same file, make sure that the property wsd.shortOutput is set to false (wsd.shortOutput=false).
In order to use JIGSAW, you must also configure the files GKR4NLI/sem.mapper/src/main/resources/gkr.properties and GKR4NLI/gnli/src/main/resources/gnli.properties. In particular you must change the following line:
jigsaw_props=/Users/kkalouli/Documents/libraries/JIGSAW-master/resources/jigsaw.properties
Change the properties file name and set the absolute path in which your jigsaw.properties file is installed.
- Download and install SUMO as described in https://github.com/ontologyportal/sigmakee (you don't need to install the apache-tomcat related stuff if we you don't want to run the SUMO interface through your localhost). Download the file WordNetMappings30-all.txt provided here and put it in SUMO's folder WordNetMappings (this is a combined version of the four separate files WordNetMappings30-noun/verb/adj/adv that SUMO itself provides and is used as a consolidated verison when a term cannot be found in the file of its corresponding POS.)
In order to use SUMO, you must also configure the files GKR4NLI/sem.mapper/src/main/resources/gkr.properties and GKR4NLI/gnli/src/main/resources/gnli.properties. In particular you must change the following line:
sumo_location=/Users/kkalouli/Documents/workspace/sumo
Change the location and set the location path in which you installed the parent SUMO folder.
-
Download the jar file edu.mit.jwi_2.4.0.jar from https://projects.csail.mit.edu/jwi/
-
Create a new folder GKR_libs somewhere in your system. Copy in this folder the following files:
- the edu.mit.jwi_2.4.0.jar jar file you just downloaded
- the JIGSAW.jar file (found in the folder dist of your JIGSAW installation)
- the maxent-3.0.0.jar file (found in the folder dist/lib/ of your JIGSAW installation)
- opennlp-tools-1.5.0.jar (found in the folder dist/lib/ of your JIGSAW installation)
- The GKR parser uses the easy-bert software by Rob Rua (https://zenodo.org/record/2652964#.X6pxb5NKjR0). The software is imported through gradle and the bert vocabulary file for the BERT uncases model is already provided in /GKR4NLI/sem.mapper/src/main/resources/vocab.txt . If you wish to use a different model, you will need to download the desired model from https://github.com/robrua/easy-bert and unzip it. Once you unzip it, you will see a file vocab.txt within the folder. You will need to configure the files GKR4NLI/sem.mapper/src/main/resources/gkr.properties and GKR4NLI/gnli/src/main/resources/gnli.properties. In particular, change the following line:
bert_vocab=/Users/kkalouli/Documents/project/sem.mapper/src/main/resources/vocab.txt
to point to the location of the vocab.txt file. You will also need to modify the build.gradle files of sem.mapper and gnli to import the model you want (implementation 'com.robrua.nlp.models:easy-bert-uncased-L-12-H-768-A-12:1.0.0' )
-
Go back into the cloned directory and find the build.gradle file of the gnli folder. Change the following line
compile fileTree(dir: '/Users/kkalouli/Documents/libraries/GKR_libs/', include: ['*.jar'])to point at the location in which you created the GKR_libs folder. -
Go back into the cloned directory and find the build.gradle file of the sem.mapper folder. Change the following line
compile fileTree(dir: '/Users/kkalouli/Documents/libraries/GKR_libs/', include: ['*.jar'])to point at the location in which you created the GKR_libs folder. -
Go back into the cloned directory and find the GKR4NLI/gnli/src/main/resources/gnli.properties file. Change the following line
sumo_kb=/Users/kkalouli/Documents/.sigmakee/KBsto point at the location in which you installed the .sigmakee folder of SUMO.
You are all set! You can now build the project:
Navigate to the cloned directory (GKR4NLI) and do: gradle build
If the build is successful, you can simply import the project in Eclipse or any other platform you wish. For Eclipse: import it as a gradle project (for changes made within eclipse, the 'gradle cleanEclipse' command should refresh the eclipse view)
In the subproject gnli there is a class named InferenceComputer. In the main method at the end of this class you can choose to run only one pair or a testsuite with pairs: just comment in or out the corresponding method. The testsuite should contain one pair per line and the two sentences of the pair should be separated by a tab. The pair should also be annotated with one of E, C or N. So, the file should look like:
sentenceA \t sentenceB \t N \n
sentenceC \t sentenceD \t E \n
etc (the inference label is not used anywhere during the computation -- the association rule mining learning is done offline --, but this input is more convenient for the later evaluation of the results)
For troubleshooting, comments, ideas and discussions, please contact kalouli(at)cis.lmu.de