LR-RNN is a Python implementation of a morphological analysis algorithm introduced in the paper "Sequential Morphological Analysis of Hiragana Strings using Recurrent Neural Network and Logistic Regression".

This implementation enables you to reproduce the experimental results presented in the paper.

The proposed method specializes in analyzing highly ambiguous sequences, particularly Japanese text written in hiragana. Compared to conventional analyzers, LR-RNN achieves superior performance in both word segmentation and part-of-speech tagging when processing hiragana sentences. Below are the key evaluation results from the paper:

Word Segmentation:

Word Segmentation & Part-of-Speech Tagging:

Sequential Morphological Analysis of Hiragana Strings using Recurrent Neural Network and Logistic Regression
Authors: Shuhei Moriyama, Tomohiro Ohno
Published in: Journal of Natural Language Processing, 2022, Volume 29, Issue 2, Pages 367-394.
Language: Japanese

To set up LR-RNN, you need:

• Docker
• NVIDIA Container Toolkit

1. Navigate to the docker/ directory in the repository:

cd docker/

2. Build a Docker image:

docker build -t lrrnn .

This command creates a Docker image named lrrnn.

1. Verify your current working directory is LR-RNN/:

pwd

2. Run the following command to start a Docker container:

docker run --gpus=all --rm -it -v "$(pwd)":/workspace -p 6006:6006 lrrnn

3. Check if the files and directories in the current working directory are mounted to the /workspace directory within the container:

ls /workspace

All commands in the following sections should be executed within this container.

The steps below outline the process for replicating the results presented in Section 3 of the paper.

Kyoto University Text Corpus Version 4.0 is required to construct the dataset used in the experiments in the paper. For more information on the Kyoto University Text Corpus, please refer to the link above.

Save the Kyoto University Text Corpus under the LR-RNN/ directory:

LR-RNN
├── KyotoCorpus4.0
│   └── dat
│       └── syn
│           ├── 950101.KNP
│           ├── 950102.KNP
│           └── ...
├── README.md (this document)
└── ...

Long story short, run the following script to create the dataset:

./reproduction/scripts/build_datasets.sh

The experiment employs 5-fold cross-validation. Therefore, the script creates five datasets from the Kyoto University Text Corpus. Each dataset is created by extracting a different set of documents from the corpus.

After running the script, the following directory structure is created:

LR-RNN
├── data
│   ├── test_950112
│   ├── test_950113
│   ├── test_950114
│   ├── test_950115
│   └── test_950116
├── README.md (this document)
└── ...

Each dataset contains the following files:

test_950112
├── kana
│   ├── gt
│   │   ├── test.txt
│   │   ├── train.txt
│   │   └── valid.txt
│   └── raw
│       ├── test.txt
│       ├── train.txt
│       └── valid.txt
└── kanji
    ├── gt
    │   ├── test.txt
    │   ├── train.txt
    │   └── valid.txt
    └── raw
        ├── test.txt
        ├── train.txt
        └── valid.txt

• The kana and kanji directories contain the data in hiragana and kanji, respectively.
  • We use the kana data in the following sections.
• The raw directory contains the unsegmented text data, and the gt directory contains the segmented text data.
  • The gt directory is used for training and evaluation as the ground truth.
  • The raw directory is used for prediction as the input data.
• The train.txt, valid.txt, and test.txt files contain the training, validation, and test data, respectively.

The ground truth files (gt directory) use the following format:

• One sentence per line (ends with \n)
• Morphemes separated by spaces
• Each morpheme contains seven fields separated by slashes:
  1. Surface form in hiragana
  2. Part-of-speech (Major category)
  3. Part-of-speech (Subcategory)
  4. Group
  5. Conjugation type
  6. Kanji representation
  7. Reading in hiragana

The files in the raw directory are formatted as follows:

たろうはとうきょうだいがくにいった
じろうはきょうとだいがくにいった

• Each line ends with the newline character \n.
• Each line contains an unsegmented sequence of characters.

To train the proposed method using the datasets created in the previous section, run reproduction/scripts/train.sh.

./reproduction/scripts/train.sh

The script runs 5 sets of training on each dataset. When the script finishes, trained model files are output to the data/<each-dataset-name>/models/ directory.

The script also outputs training logs for neural networks to the models/ directory. The log files are named events.out.tfevents.*. You can visualize the training progress using TensorBoard by running:

tensorboard --host=0.0.0.0 --logdir=data/test_950112/models/<directory-starting-with-rnn>

Then access TensorBoard through your browser at http://localhost:6006.

To analyze each test data using the trained models, run reproduction/scripts/predict.sh.

./reproduction/scripts/predict.sh

After the script finishes, prediction results are output to the predictions/ directory.

LR-RNN
├── predictions
│   ├── test_950112_kana.txt
│   ├── test_950113_kana.txt
│   ├── test_950114_kana.txt
│   ├── test_950115_kana.txt
│   └── test_950116_kana.txt
└── ...

reproduction/scripts/evaluate.sh evaluates the prediction results.

./reproduction/scripts/evaluate.sh

The script outputs Precision, Recall, and F-measure on each dataset to the standard output. The average of the evaluation results will be similar to the results in the paper.

That's it! You have successfully reproduced the results in the paper.

• The contents of this repository, excluding the pre-trained model files, are licensed under the MIT License.
• The pre-trained model files in Releases are available for research purposes only.
• See LICENSE for the complete MIT License terms.