Skip to content

Latest commit

 

History

History
106 lines (82 loc) · 5.31 KB

tokenization-benchmark.md

File metadata and controls

106 lines (82 loc) · 5.31 KB
Raw

Word Tokenization Benchmark for Thai (obsolete)

A framework for benchmarking tokenization algorithms for Thai. It has a command-line interface that allows users to conveniently execute the benchmarks as well as a module interface for later use in their development pipelines.

Metrics

Character-Level (CL)

  • True Positive (TP): no. of starting characters that are correctly predicted.
  • True Negative (TN): no. of non-starting characters that are correctly predicted.
  • False Positive (FP): no. of non-starting characters that are wrongly predicted as starting characters.
  • False Negative (FN): no. of starting characters that are wrongly predicted as non-starting characters.
  • Precision: TP / (TP + FP)
  • Recall: TP / (TP+FN)
  • f1: ...

Word-Level (WL)

  • Correctly Tokenized Words (CTW): no. of words in reference that are correctly tokenized.
  • Precision: CTW / no. words in reference solution
  • Recall: CTW / no. words in sample -**** f1: ...

Benchmark Results

Vendor Approach Datasets
DeepCut CNN
PyThaiNLP-newmm dictionary-based
Sertis-BiGRU Bi-directional RNN

Installation (WIP)

pip ...

Usages (to be updated)

  1. Command-line Interface

    PYTHONPATH=`pwd` python scripts/thai-tokenisation-benchmark.py \
--test-file ./data/best-2010/TEST_100K_ANS.txt \
--input ./data/best-2010-syllable.txt

# Sample output
Benchmarking ./data/best-2010-deepcut.txt against ./data/best-2010/TEST_100K_ANS.txt with 2252 samples in total
============== Benchmark Result ==============
                metric       mean±std       min    max
         char_level:tp    47.82±47.22  1.000000  354.0
         char_level:tn  144.19±145.97  1.000000  887.0
         char_level:fp      1.34±2.02  0.000000   23.0
         char_level:fn      0.70±1.19  0.000000   14.0
  char_level:precision      0.96±0.08  0.250000    1.0
     char_level:recall      0.98±0.04  0.500000    1.0
         char_level:f1      0.97±0.06  0.333333    1.0
  word_level:precision      0.92±0.14  0.000000    1.0
     word_level:recall      0.93±0.12  0.000000    1.0
         word_level:f1      0.93±0.13  0.000000    1.0

  2. Module Interface

    from pythainlp.benchmarks import word_tokenisation as bwt

ref_samples = array of reference tokenised samples
tokenised_samples = array of tokenised samples, aka. from your algorithm

# dataframe contains metrics for each sample
df = bwt.benchmark(ref_samples, tokenised_samples)

Related Work

Development

# unitests
$ TEST_VERBOSE=1 PYTHONPATH=. python tests/__init__.py

Acknowledgement

This project was initially started by Pattarawat Chormai, while he was interning at Dr. Attapol Thamrongrattanarit's lab.