A Comparison of Four Character-Level String-to-String Translation Models for (OCR) Spelling Error Correction

Abstract

We consider the isolated spelling error correction problem as a specific subproblem of the more general string-to-string translation problem. In this context, we investigate four general string-to-string transformation models that have been suggested in recent years and apply them within the spelling error correction paradigm. In particular, we investigate how a simple ‘k-best decoding plus dictionary lookup’ strategy performs in this context and find that such an approach can significantly outdo baselines such as edit distance, weighted edit distance, and the noisy channel Brill and Moore model to spelling error correction. We also consider elementary combination techniques for our models such as language model weighted majority voting and center string combination. Finally, we consider real-world OCR post-correction for a dataset sampled from medieval Latin texts.

Proceedings of the SIGFSM Workshop on Statistical NLP and Weighted Automata

“…The average performance of the perceptron tagger in this experiment is superior to the performance of the AliSeTra system as reported by Eger et al (2016). The difference in performance is, however, not statistically significant.…”

“…UC refers to the unstructured classifier presented in Section 3.1, PT to the perceptron tagger presented in Section 3.2 and AliSeTra to the system presented by Eger et al (2016).…”

“…Like DirecTL+, it also views string-tostring translation as a pipeline of segmentation and sequence labeling. (4) The final system surveyed by Eger et al (2016) represents the stringto-string translation task as a series of contextual edit operations on the input string (Cotterell et al, 2015). The operations are compiled into a weighted finite-state machine.…”

“…Systems 1, 2 and 3 surveyed by Eger et al (2016) form an interesting contrast to our systems because we do not use segmentation of the input string. In this sense, our system is simpler.…”

Data-Driven Spelling Correction using Weighted Finite-State Methods

“…Note that when all input forms are incorrect (as in the case of the Twitter data), CR corresponds exactly to the evaluation metric word accuracy (WACC) used by Eger et al (2016) because the count f p is 0. WACC = tp tp + f n Tables 2 and 3 show the results of the experiments on the Finnish OCR data and Twitter data.…”