Text-Induced Corpus Clean-up
Python Perl CSS Shell
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.




Text-induced Corpus Clean-up is the stand-alone command line version of the spelling correction and OCR post-correction system we have been developing since about 2008.

TICCLops is TICCL as an 'online processing system'. It is a fully fledged web application and web service due to [CLAM] (http://proycon.github.io/clam/), developed by Maarten van Gompel.

The current repository mainly houses the wrapper scripts, variably Perl and Python scripts.

The main TICCL modules were developed by Ko van der Sloot in C++ in various projects over the past years. These are available from: [TICCLtools at LanguageMachines] (https://github.com/LanguageMachines/ticcltools)

We have prepared TICCL for work in many languages, mainly on the basis of available open source lexicons due to Aspell. The language specific files are available here:

[All languages] (http://ticclops.uvt.nl/TICCL.languagefiles.ALLavailable.20160421.tar.gz) [Dutch] (http://ticclops.uvt.nl/TICCL.languagefiles.nld.20160421.tar.gz) [English] (http://ticclops.uvt.nl/TICCL.languagefiles.eng.20160421.tar.gz) [Finnish] (http://ticclops.uvt.nl/TICCL.languagefiles.fin.20160421.tar.gz) [French] (http://ticclops.uvt.nl/TICCL.languagefiles.fra.20160421.tar.gz) [Friesian] (http://ticclops.uvt.nl/TICCL.languagefiles.fry.20160421.tar.gz) [German] (http://ticclops.uvt.nl/TICCL.languagefiles.deu.20160421.tar.gz) [German Fraktur] (http://ticclops.uvt.nl/TICCL.languagefiles.deu-frak.20160421.tar.gz) [Greek (antique)] (http://ticclops.uvt.nl/TICCL.languagefiles.grc.20160421.tar.gz) [Greek (modern)] (http://ticclops.uvt.nl/TICCL.languagefiles.ell.20160421.tar.gz) [Icelandic] (http://ticclops.uvt.nl/TICCL.languagefiles.isl.20160421.tar.gz) [Italian] (http://ticclops.uvt.nl/TICCL.languagefiles.ita.20160421.tar.gz) [Latin] (http://ticclops.uvt.nl/TICCL.languagefiles.lat.20160421.tar.gz) [Polish] (http://ticclops.uvt.nl/TICCL.languagefiles.pol.20160421.tar.gz) [Portuguese] (http://ticclops.uvt.nl/TICCL.languagefiles.por.20160421.tar.gz) [Romanian] (http://ticclops.uvt.nl/TICCL.languagefiles.ron.20160421.tar.gz) [Russian] (http://ticclops.uvt.nl/TICCL.languagefiles.rus.20160421.tar.gz) [Spanish] (http://ticclops.uvt.nl/TICCL.languagefiles.spa.20160421.tar.gz) [Swedish] (http://ticclops.uvt.nl/TICCL.languagefiles.swe.20160421.tar.gz)

Unpack in your main TICCL directory. A subdirectory data/int/ will be created to house the required files for the specific language(s).

We also provide some sample corpora. These come with TICCLops for demonstration and test purposes. These are available here:

[All demo and test sets] (http://ticclops.uvt.nl/TICCL.SampleCorpora.20160504.ALL.tar.gz) [DjVu format] (http://ticclops.uvt.nl/TICCL.SampleCorpora.DJVU.20160504.tar.gz) [FoLiA XML format] (http://ticclops.uvt.nl/TICCL.SampleCorpora.FOLIA.20160504.tar.gz) [OCR] (http://ticclops.uvt.nl/TICCL.SampleCorpora.OCR.20160504.tar.gz) [PDF with embedded page images] (http://ticclops.uvt.nl/TICCL.SampleCorpora.PDF.20160504.tar.gz) [Tab-separated list] (http://ticclops.uvt.nl/TICCL.SampleCorpora.TSV.20160504.tar.gz) [Text files] (http://ticclops.uvt.nl/TICCL.SampleCorpora.TXT.20160504.tar.gz) [TIFF image files] (http://ticclops.uvt.nl/TICCL.SampleCorpora.TIFF.20160504.tar.gz)

Running TICCL from the command line

Use TICCLops.PICCL.pl to run TICCL from the command line.

Examples of input parameters with some further documentation are to be found in file: TICCL.Black.config .

Edit file TICCL.Template.config in order to specify the specific settings and file locations for your own installation.

You can then start TICCL by running the following command: $ perl TICCLops.PICCL.pl TICCL.Template.config

Running TICCL from the command line -- specific parameter settings##

The first parameter set in TICCL's config file, or the first parameter i.e. ARGV[0] when all parameters are specified on the command line, is typically a combination of alphabetical characters that when set, activate particular TICCL C++ modules.

As default we have: abcmdef. This means all of TICCL's C++ modules will be run sequentially, but two of them will be run in 'word focus mode'. This refers to our main TICCL publication (Reynaert 2010). In its essence, this is the cheapest option to run TICCL if the corpus you want to have (OCR post-)corrected is not very big.

We next describe what each letter invokes.

Parameter setting a

The letter a in fact works in tandem with line 2 in the config file, parameter -b on the full command line. The character a specifies that your input will be running text. This running text may be in three formats, either FoLiA XML, else some unspecified XML format or finally plain text format. For FoLiA XML, specify FOLIA in line 2, for unspecified XML put XML, for plain text put TXT.

For unspecified XML, the text will be extracted from the <t> nodes in the XML.

For FoLiA XML, the C++ ticcltools module FoLiA-stats is invoked. In fact, this module can also provide n-gram lists.

For plain XML or TXT the C++ ticcltools module TICCL-stats is invoked. This is to data less well developed than the FoLiA module.

In fact, the module invoked will derive a frequency list for your corpus from your input files. This entails that in case you already have a frequency file, or even in case that you only have a frequency file for your corpus (which might be the case if you acquired OCRed books from the Hathi Trust collection, for instance), you can skip this step and proceed from the next.

Parameter setting b

The letter b envokes the module TICCL-unk. This performs an elementary clean-up of the input frequency list. In fact, this removes character strings that have no hope of ever being corrected into what is commonly understood to be a meaningful word in any natural language.

Parameter setting c

The letter c invokes C++ module TICCL-anahash. This converts the cleaned frequency file into the TICCL anagram hash. Can be used in conjunction with the letter m (please, see there) to provide a speedier further processing for smaller corpora.

Parameter setting d

The letter d invokes C++ module TICCL-indexerNT if combined with the letter n, C++ module TICCL-indexer if not. TICCL-indexerNT is a multi-threaded implementation of focus word correction candidate retrieval. TICCL-indexer is an implementation of character confusion correction candidate retrieval. This may be set to work in parallel by also invoking the letter n, in which case it will create a temporary directory to split the character confusion list in as many parts as you defined the number of threads for TICCL to work with in the TICCL configuration file.

Note that currently output of TICCL-indexer on the basis of a large corpus may be a very large, uncompressed numerical file.

Parameter setting e

The letter e invokes TICCL-LDcalc, which primarily steps back to the symbolic representations of the word types, performs some basic filtering and gathering of statistics about the word types as well the actual character confusions observed between focus words and the correction candidates retrieved.

Parameter setting f

The letter f invokes TICCL-rank. This C++ module uses a large range of features in order to try and rank the correction candidates as best as possible.

We have a thorough evaluation of TICCL's ranking on the basis of ablation experiments high on our list of todos.

Parameter setting g

The letter g invokes FoLiA-correct. This naturally presupposes your texts are in FoLiA XML format. TICCL currently provides no actual text correction/editing facilities for formats other than FoLiA XML.

Parameter setting m

If specified, the letter m makes the module TICCL-anahash also output the list of word strings that TICCL needs to search correction candidates for. This is effected on the basis of the artificial frequency we let the system assign to word types that are present in the validated lexicon. This list serves as the focus words list for the subsequent module TICCL-indexerNT. For smaller corpora this results in less work and therefore shorter running times than when one runs TICCL-indexer, which exhaustively gathers all correction candidates for all the words in the cleaned up corpus frequency list, i.e. performs character confusion based corpus clean-up.

TICCLops available online as demonstrator system

We have a demonstrator system online in an official Dutch CLARIN Center at [TICCLopsCLAM] (http://ticclops.clarin.inl.nl/ticclops/). An interface developed specifically for philosopher-users is available here: [TICCLopsPhilosTEI] (http://ticclops.clarin.inl.nl/ticclops/).

In fact, both systems are geared at letting you upload scanned images of book pages, to have them automatically converted into electronic text by way of the Tesseract OCR engine and then to have them OCR post-corrected by TICCL. For e.g. PDF files for book chapters, you will receive back a full, reconstituted book in both FoLiA and TEI formats. Please take into account that this process requires, well, the time required.