This repository provides the converter from dependency structure to phrase structure. The converter works completely algorithmically and does not need any training.

pip install pyconll nltk

For test, you additionally need to install pytest.

You can run the converter in the following format.

python converter.py
  --source_path <conll-formatted texts>
  --output_source_dir <output source directory>
  --convert_method <flat/left/right>

If source_path includes the treebank named train/dev/test set, the converted treebank is named according to the original name such as train.txt. In addition to converted treebank, the program outputs raw text separated by space, which is named such as train.tokens.

Besides above, you need to set the labeling method (described later). Currently we have four types of labeling methods and you need to choose one from them:

--without_label
--use_pos_label
--use_merged_pos_label
--use_dep_label

Example: Converting a Universal Dependencies corpus via flat conversion and without_label method.

python converter.py
  --source_path path/to/ud-treebanks-v2.7/English_EWT
  --output_source_dir path/to/output_dir
  --convert_method flat
  --without_label

The converter takes two types of parameter: conversion method and labeling policy.

Three types of methods are implemented: flat/left/right. These conversion algorithms work for any dependency structures, so long as they are projective. The details of algorithm is described in Collins et al, 1999.

In short, flat method converts the dependency tree into the "flat" constituency tree. It converts dependents into children in the constituency tree at the same time.

Left/right methods does not convert at the same time. Instead, they merge head and dependent one by one, in left-first or right-first way. These methods produce binary tree.

We set following four types of nonterminals attached to resulting constituency tree:

• Without label: All the nonterminals are set 'X'
• POS label: Using part-of-speech tag of the head token.
• Merged-POS label: Basically using POS label, and convert DETP/PROPNP/PRONP into NOUNP.
• DEP label: Using dependency label of the head token.