https://github.com/tomtung/happy-tree-align

Burkett, D., & Klein, D. (2012). Transforming trees to improve syntactic convergence
https://github.com/tomtung/happy-tree-align

Last synced: 20 days ago
JSON representation

Burkett, D., & Klein, D. (2012). Transforming trees to improve syntactic convergence

• Host: GitHub
• URL: https://github.com/tomtung/happy-tree-align
• Owner: tomtung
• Created: 2013-01-15T00:48:37.000Z (almost 12 years ago)
• Default Branch: master
• Last Pushed: 2013-05-03T23:08:19.000Z (over 11 years ago)
• Last Synced: 2024-10-05T21:07:40.307Z (about 1 month ago)
• Language: Scala
• Homepage:
• Size: 266 KB
• Stars: 1
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# Happy Tree Align [![Build Status](https://api.travis-ci.org/tomtung/happy-tree-align.png)](https://travis-ci.org/tomtung/happy-tree-align)

This is a implementation of the method described in [Burkett, D., & Klein, D. (2012) Transforming trees to improve syntactic convergence](http://www.cs.berkeley.edu/~dburkett/papers/burkett12-tree_transform.pdf). For binlingual corpora, it learns syntax tree transformation rules to improve their agreement with corresponding word alignments. This improves the performance of syntactic machine translation systems.

The name comes from the author's keynote [Happy Trees are Better than Correct Trees](http://www.cs.berkeley.edu/~dburkett/slides/burkett12-tree_transform-slides.pdf).

## Usage

The program can be built with [sbt](http://www.scala-sbt.org/).

`sbt one-jar` will produce a stand-alone jar package callable form command line.

```

Usage: happy-tree-align learn [options]    

  -n  | --n-trans 

        max number of transformations to learn (default 200)

  -o  | --out 

        output path (print to stdout by default)

  

        path to training tree file

  

        path to training alignment file

  

        optional, path to dev tree file

  

        optional, path to dev alignment file

Usage: happy-tree-align apply [options]   

  -o  | --out 

        output path. print to stdout by default

  -n  | --n-trans 

        use first n transformations (use all by default)

  

        transformation sequence file

  

        input tree file

  

        optional, path to alignment file. if provided, report agreement scores after each transformation

Usage: happy-tree-align score [options]  

  -a | --all

        output scores for all trees

  -o  | --out 

        output path. print to stdout by default

  

        tree file

  

        alignment file

```

## Peformance

In our experiment, using a 3000-sentence training data set, and a 1000-sentence development data set, on a duo core 2.53GHz machine, it takes 16 minutes to learn the first 100 rules, and 70 minutes to learn all 3979 rules.

The implementation is parallelized, so it should take less time with more cores available.