Dependency-basded Hybrid Tree for Semantic Parsing

This repository is the implementation of the paper "Dependency-based Hybrid Tree for Semantic parsing" appeared in the Empirical Methods in Natural Language Processing (EMNLP), 2018.

Requirements

• Java 1.8
• Perl and swipl (running the script to query GeoQeuery database)
• Torch (optional, required if running the neural version)

Building

Under the maven environment, just simply type:

	mvn clean package

You can obtain the jar file dht-1.0.jar under target directory to run the experiment. Alternatively, you can use IDE such as Eclipse or IntelliJ to build this package. In the worst case, if you do not know how to build, we provide the jar file for download.

Usage

Feature-based model

We compiled a jar file (including all the required external library in Maven) dht-1.0.jar.

java -cp target/dht-1.0.jar org.statnlp.example.depsemtree.DepHybridTree --thread 40 --language en

You should be able to obtain exactly same performance in the paper with the above command. To change another language (e.g., th, de, el, zh, id, sv, fa), simply replace en with other languages indicated in the paper. For further customized configuration settings (e.g., L2, features, etc), please check the main class and we will list the details with another README document soon.

Using embeddings as feature values

First, download the embedding file from this link and put them under nn-crf-interface/neural_server/polyglot/ directory. Then you can run the jar file using the following command:

java -cp target/dht-1.0.jar org.statnlp.example.depsemtree.DepHybridTree --thread 40 --language en --useEmbFeats true

You should be able to obtain exactly same performance in the paper with the above command.

Neural network-augmented model

Make sure you have installed the latest Torch package.

java -cp target/dht-1.0.jar org.statnlp.example.depsemtree.DepHybridTree --thread 40 --language en --type bilinear