Skip to content

Error detection

Logion's error detection feature predicts the likelihood that each word in a given sequence is an error. Here, error means a word that has been mistranscribed at some point in the textual history. Logion further offers suggestions as to what a possible "correct" word may be.

How to use Logion error detection

From the main menu, go to to to the error detection window by clicking Error detection on the right-hand side of the main menu. Once you are in the Error Detection page, follow these steps to generate an error report for your text.

  1. Select a model from the drop-down menu in the upper-left of the window. If this is your first time, we recommend beginning with Base BERT. This model is trained on a wide selection of premodern Greek and is suitable for general error detection.

  2. Select a Levenshtein distance from the drop-down menu to the right of the model selection menu. We recommend starting with a Levenshtein distance of 1. To learn more about Levenshtein distance, see Logion's explainer.

  3. Type/paste your text into the text area. Unlike, Logion's word prediction feature, don't use ? to represent any missing words. Enter text comprised of only Greek characters and other punctuation marks.

  4. Click the blue Detect Errors button below the text area. Note the error detection process can take several minutes depending on one's local hardware. To read more on how hardware affects processing speed, see Logion's hardware guide.

Logion displays results below the blue Detect Errors button.

How to read error reports

Text is color-coded to signify each given word's likelihood of it having been mistranscribed at some point in the textual history. Text is colored on a gradient of green-yellow-orange-red. Green means the word is unlikely to be mistranscribed; red means the word is very likely to be mistranscribed. To see what the model suggests as a potential replacement word, click a given word. Replacement word suggestions are displayed on the right-hand side of the window beside the model's projected chance-confidence ratio for that word pair. To see a different word's results, simply click that different word.