TICCLAT

[jvdzwaan]

The TICCLAT project is about extending TICCL, software that does ocr post correction and/or spelling correction and/or word normalization based on what word forms it sees in the corpus. In this project we want to run a number of experiments to evaluate the performance of different configurations of TICCL. The sprint will be about setting up the pipeline/infrastructure to run these experiments. We will focus on the task of OCR post-correction and have a data set available. Hopefully, we'll be able to run the baseline experiment at the end of the sprint.

Together with @egpbos

[egpbos]

Main goal

Setup framework/pipeline for validating TICCL (+ extra metadata) given different configurations of input data and parameters.

Data preparation

• DBNL data (from KB): XML file with corrected OCR'ed text (TEI format) + text file with uncorrected text + metadata
  • TEI format has footnotes/endnotes, references added, complex junk, needs to be cleaned / decided how to handle
  • Alignment
  • What are differences? Determine how much are the texts alike so you can make balanced train/validation sets.
  • Split into training and validation sets
  • If necessary to get started on other task, as zeroth order attempt: just randomly split.

Define research questions:

Input for performance measurements.

• Define configurations:

  • Only test corpus input
  • • Lexicon
  • • Names
  • • Time, Location, Source (attestation) metadata
  • and combinations of several and all of these.

• What is the effect on performance of configuration X in OCR post correction tasks?

  • What is the effect of adding lexicons (with historical spelling variation)? How much does it help and in which cases (if any) does it decrease performance?

• How does configuration X perform on lexical assessment tasks? "Is this a word?"

  • Either in context or in isolation

Performance assessment pipeline

• Define and program performance tests
  • check existing work: e.g. OCR evaluation tool or Docker version and Martin Reynaert's previous work
• Define input configuration options
  • Input datasets: corpus, lexicon, names
  • Input parameters: artifreq, weights of ranking features, use of years/periods/source
• Output format/visualization
  • Check @jvdzwaan's previous KB OCR post-correction work (offline)

Measure performance

When we have data and validation pipeline, we can run!

• Baseline performance: using vanilla TICCL settings
• Configurations performance

Database setup: from TICCL to TICCLAT

https://github.com/TICCLAT/explore/blob/master/notebooks/folia2lexicon.ipynb

[eriktks]

The results for the OCR evaluation part of the sprint can be found on this address: https://github.com/TICCLAT/evaluation_pipeline/tree/master/evaluation

[romulogoncalves]

@jvdzwaan were all the goals achieved? Can we consider it done and close this issue?

[jvdzwaan]

It's a work in progress and we'll continue the work in the project. The issue can be closed.