Speed Comparison

This wiki shows the tokenization speed of Vibrato and other tokenizers and morphological analyzers.

Experimental setup

We compare Vibrato 0.5.0 with MeCab and its reimplementations:

For Vibrato and MeCab, we evaluate two system dictionaries: IPADIC 2.7.0 and UniDic 3.1.1. For Lindera, we evaluate two versions: IPADIC and UniDic. sudachi.rs is evaluated for SudachiDict-core.

Further, we evaluate two compact versions of Vibrato UniDic models (distributed in our release page):

We also compare pointwise prediction-based tokenizers:

For Vaporetto and KyTea, we used the compact SVM model based on BCCWJ and UniDic downloaded from KyTea Models page.

We tokenize all sentences in I Am a Cat (by Soseki Natsume), which is available at Aozora Bunko, and report the elapsed time averaged on 100 runs.

The benchmark code can be found here.

The following is the specification of the used machine:

Library (dict)	Elapsed time [ms]	STD
Vibrato 0.5.0 (ipadic-mecab 2.7.0)	42	1.24
Vibrato 0.5.0 (unidic-cwj 3.1.1)	75	1.71
Vibrato 0.5.0 (unidic-cwj 3.1.1, raw-connector)	1364	5.14
Vibrato 0.5.0 (unidic-cwj 3.1.1, dual-connector)	170	2.50
MeCab 2020-09-14 (ipadic-mecab 2.7.0)	87	1.24
MeCab 2020-09-14 (unidic-cwj 3.1.1)	179	2.88
Lindera 0.23.0 (ipadic)	97	1.13
Lindera 0.23.0 (unidic)	156	2.11
sudachi.rs 0.6.4-a1 (core, 20210802)	220	4.74

KyTea 2020-04-03 (jp-0.4.7-5)	169	2.83
Vaporetto 0.6.1 (jp-0.4.7-5)	21	0.51
rust-tinysegmenter 0.1.1	166	1.69

Note that Vibrato UniDic models differ in size as follows. Thus, you can use the model with the time-space tradeoff of your choice.

Library (dict)	Model size [MB]
Vibrato 0.5.0 (unidic-cwj 3.1.1)	717
Vibrato 0.5.0 (unidic-cwj 3.1.1, raw-connector)	252
Vibrato 0.5.0 (unidic-cwj 3.1.1, dual-connector)	300