Character-based PCFG Induction for Modeling the Syntactic Acquisition of Morphologically Rich Languages

Introduction

This is the code repository for the paper Character-based PCFG Induction for Modeling the Syntactic Acquisition of Morphologically Rich Languages, including unsupervised PCFG induction models as well as manually corrected syntactic annotations for Korean child-directed speech.

Major dependencies include:

• Python 3.7+
• PyTorch 1.7.0+
• TensorBoard 2.3.0+
• Bidict

Training

main.py is the main training script. Sample commands for training the induction model can be found under the exps directory:

python main.py train --seed -1 \
                     --train_path data/Jong.010322.linetoks \
                     --train_gold_path data/Jong.010322.linetrees \
                     --model_type char --model char_jong_c90 \
                     --num_nonterminal 45 --num_preterminal 45 \
                     --state_dim 128 --rnn_hidden_dim 512 \
                     --max_epoch 45 --batch_size 2 \
                     --optimizer adam --device cuda \
                     --eval_device cuda --eval_steps 2 \
                     --eval_start_epoch 1 --eval_parsing

seed: Seed for the run, -1 can be used for a random seed.

train_path: Path to training sentences, which should be whitespace-tokenized like the following:

이리로 와 엄마랑 보자 .
이게 누구에요 ?
아우 이쁘다 우리 종현이네 .

train_gold_path: Path to gold trees for training sentences, which are used only for evaluation:

(S (S (ADVP (npd+jca 이리로)) (VP (pvg+ef 와))) (S (VP (NP (ncn+jcj 엄마랑)) (pvg+ef 보자))) (sf .))
(S (NP (npd+jcs 이게)) (VP (npp+jp+ef 누구에요)) (sf ?))
(S (S (IP (ii 아우)) (VP (paa+ef 이쁘다))) (S (ADJP (npp 우리)) (VP (nq+jp+ef 종현이네)) (sf .)))

model_type: Use char for the NeuralChar model and word for the NeuralWord model described in the paper.

model: The name of the directory in which the output will be saved saved. For example, if char_jong_c90 is used, then the run info will be saved under outputs/char_jong_c90_0 for the first run, outputs/char_jong_c90_1 for the second run, and so on.

num_nonterminal and num_preterminal: The NeuralChar and NeuralWord models do not make a distinction between preterminal categories and other nonterminal categories. Set these two values accordingly so that they add up total number of categories to be used (e.g. 45 and 45 for a total of 90 categories).

state_dim: Dimensionality of category embeddings.

rnn_hidden_dim: Dimensionality of the LSTM hidden state for the NeuralChar model's emission probabilities.

eval_steps: Number of training epochs between each evaluation.

eval_start_epoch: Number of training epochs before first evaluation.

eval_parsing: Whether or not to parse training sentences as part of evaluation. If set to False, only the likelihood will be reported.

Descriptions of other arguments can be found in model_args.py.

Syntactic Annotations of Korean Child-Directed Speech

The data directory contains manually corrected syntactic annotations of Korean child-directed speech from the Ryu corpus of the CHILDES database. The annotation scheme follows that of Choi (2013).

Questions

For questions or concerns, please contact Lifeng Jin (lifengjin@tencent.com) or Byung-Doh Oh (oh.531@osu.edu).