commit	b71b51fe1e63950dc84d18a3face5c5eefc5915f	[log] [tgz]
author	Akron <nils@diewald-online.de>	Mon Mar 04 15:39:02 2024 +0100
committer	Akron <nils@diewald-online.de>	Mon Mar 04 15:39:02 2024 +0100
tree	0791bf0c2fc829dfa85817d861a908a41bfec273
parent	0b9aa666d6e824452935c84e8f88637f5017499c [diff]

tree: 0791bf0c2fc829dfa85817d861a908a41bfec273

Readme.md

Datok - Finite State Tokenizer

Introduction to Datok

Implementation of a finite state automaton for high-performance large-scale natural language tokenization, based on a finite state transducer generated with Foma.

The repository currently contains precompiled tokenizer models for

The focus of development is on the tokenization of DeReKo, the german reference corpus.

Datok can be used as a standalone tool or as a library in Go.

Performance

Speed comparison of german tokenizers

Chart showing speed comparison of different tokenizers and sentence splitters for German. Effi refers to tokenizing and/or sentence splitting of one issue of Effi Briest. Datok is optimized for large batch sizes, while other tools may perform better in other scenarios. For further benchmarks, especially regarding the quality of tokenization, see Diewald/Kupietz/Lüngen (2022).

Tokenization

Usage: datok tokenize --tokenizer=STRING <input>

Arguments:
  <input>    Input file to tokenize (use - for STDIN)

Flags:
  -h, --help                  Show context-sensitive help.

  -t, --tokenizer=STRING      The Matrix or Double Array Tokenizer file
      --[no-]tokens           Print token surfaces (defaults to true)
      --[no-]sentences        Print sentence boundaries (defaults to true)
  -p, --token-positions       Print token offsets (defaults to false)
      --sentence-positions    Print sentence offsets (defaults to false)
      --newline-after-eot     Ignore newline after EOT (defaults to false)

The special END OF TRANSMISSION character (\x04) can be used to mark the end of a text.

Caution: When experimenting with STDIN and echo, you may need to disable history expansion.

Conversion

Usage: datok convert --foma=STRING --tokenizer=STRING

Flags:
  -h, --help                Show context-sensitive help.

  -i, --foma=STRING         The Foma FST file
  -o, --tokenizer=STRING    The Tokenizer file
  -d, --double-array        Convert to Double Array instead of Matrix
                            representation

Library

package main

import (
	"github.com/KorAP/datok"
	"os"
	"strings"
)

func main () {

	// Load transducer binary
	dat := datok.LoadTokenizerFile("tokenizer_de.matok")
	if dat == nil {
		panic("Can't load tokenizer")
	}

	// Create a new TokenWriter object
	tw := datok.NewTokenWriter(os.Stdout, datok.TOKENS|datok.SENTENCES)
	defer tw.Flush()

	// Create an io.Reader object refering to the data to tokenize
	r := strings.NewReader("Das ist <em>interessant</em>!")

	// The transduceTokenWriter accepts an io.Reader
	// object and a TokenWriter object to transduce the input
	dat.TransduceTokenWriter(r, tw)
}

Conventions

The FST generated by Foma must adhere to the following rules, to be convertible by Datok:

Character accepting arcs need to be translated only to themselves or to ε (the empty symbol). I.e. they will either be unchanged part of the output or ignored (e.g. whitespace characters).
Multi-character symbols are not allowed, except for the @_TOKEN_BOUND_@, that denotes the end of a token.
ε accepting arcs (transitions not consuming any character) need to be translated to the @_TOKEN_BOUND_@.
Two consecutive @_TOKEN_BOUND_@s mark a sentence end.
Flag diacritics are not supported.
Final states are ignored. The @_TOKEN_BOUND_@ marks the end of a token instead.

A minimal usable tokenizer written in XFST and following the guidelines to tokenizers in Beesley and Karttunen (2003) and Beesley (2004) would look like this:

define TB "@_TOKEN_BOUND_@";

define WS [" "|"\u000a"|"\u0009"];

define PUNCT ["."|"?"|"!"];

define Char \[WS|PUNCT];

define Word Char+;

! Compose token bounds
define Tokenizer [[Word|PUNCT] @-> ... TB] .o.
! Compose Whitespace ignorance
       [WS+ @-> 0] .o.
! Compose sentence ends
       [[PUNCT+] @-> ... TB \/ TB _ ];

read regex Tokenizer;

Hint: For development in Foma it's easier to replace @_TOKEN_BOUND_@ with a newline symbol.

Building

To build the tokenizer tool, run

$ make build

To create a foma file from the example sources, first install Foma, then run in the root directory of this repository

$ cd src && \
  foma -e "source de/tokenizer.xfst" \
  -e "save stack ../mytokenizer.fst" -q -s && \
  cd ..

This will load and compile the german tokenizer.xfst and will save the compiled FST as mytokenizer.fst in the root directory.

To generate a Datok FSA (matrix representation) based on this FST, run

$ datok convert -i mytokenizer.fst -o mytokenizer.datok

To generate a Datok FSA (double array representation*) based on this FST, run

$ datok convert -i mytokenizer.fst -o mytokenizer.datok -d

The final datok file can then be used as a model for the tokenizer.

This may take quite some time depending on the number of arcs in the FST and is therefore not recommended in most cases.

Technology

Internally the FSA is represented either as a matrix or as a double array.

Both representations mark all non-word-character targets with a leading bit. All ε (aka tokenend) transitions mark the end of a token or the end of a sentence (2 subsequential ε). The transduction is greedy with a single backtracking option to the last ε transition.

The double array representation (Aoe 1989) of all transitions in the FST is implemented as an extended DFA following Mizobuchi et al. (2000) and implementation details following Kanda et al. (2018).

References

Please cite this work as:

Diewald, Nils (2022): Matrix and Double-Array Representations for Efficient Finite State Tokenization. In: Proceedings of the 10th Workshop on Challenges in the Management of Large Corpora (CMLC-10) at LREC 2022. Marseille, France, pp. 20-26.

The library contains sources for a german tokenizer based on KorAP-Tokenizer.

For speed and quality analysis in comparison to other tokenizers for German, please refer to this article:

Diewald, Nils, Marc Kupietz, Harald Lüngen (2022): Tokenizing on scale - Preprocessing large text corpora on the lexical and sentence level. In: Proceedings of EURALEX 2022. Mannheim, Germany, pp. 208-221.

The benchmarks can be reproduced using this test suite.

License

Datok is published under the Apache 2.0 License.

The german and english tokenizer shipped is based on work done by the Lucene project (published under the Apache License), David Hall (published under the Apache License), Çağrı Çöltekin (published under the MIT License), and Marc Kupietz (published under the Apache License).

The english clitics list is based on Zwicky & Pullum (1983).

The foma parser is based on foma2js, written by Mans Hulden (published under the Apache License).

Bibliography

Aoe, Jun-ichi (1989): An Efficient Digital Search Algorithm by Using a Double-Array Structure. IEEE Transactions on Software Engineering, 15 (9), pp. 1066-1077.

Beesley, Kenneth R. & Lauri Karttunen (2003): Finite State Morphology. Stanford, CA: CSLI Publications.

Beesley, Kenneth R. (2004): Tokenizing Transducers. https://web.stanford.edu/~laurik/fsmbook/clarifications/tokfst.html

Hulden, Mans (2009): Foma: a finite-state compiler and library. In: Proceedings of the 12th Conference of the European Chapter of the Association for Computational Linguistics, Association for Computational Linguistics, pp. 29-32.

Mizobuchi, Shoji, Toru Sumitomo, Masao Fuketa & Jun-ichi Aoe (2000): An efficient representation for implementing finite state machines based on the double-array. Information Sciences 129, pp. 119-139.

Kanda, Shunsuke, Yuma Fujita, Kazuhiro Morita & Masao Fuketa (2018): Practical rearrangement methods for dynamic double-array dictionaries. Software: Practice and Experience (SPE), 48(1), pp. 65–83.

Zwicky, Arnold M., Geoffrey K. Pullum (1983): Cliticization vs. Inflection: English N’T. Language, 59, pp. 502-513.