Responses from user studies
Anonymized answers of respondents from the user studies run within the project. The project's analyses were computed on this table, including the calibration of the authorship verdict this application reports.
Each row is one blank in a cloze test: the expression the respondent typed, the original expression from the text, and how the two were scored. The table contains no data identifying the respondents.
What the columns mean
test_id
|
Identifier of the test; all rows of one filled-in test share it. |
|---|---|
test_type
|
Variant of the test the respondent took (authorML, previewML, randomML, reviewML). |
word_types
|
Part-of-speech pattern of the blanked expression, e.g. ADJ+NOUN. |
similarity
|
Score of the answer against the original: 1 for an exact match, 0 for no match, and the cosine similarity of the word embeddings in between. |
answered
|
Expression the respondent typed into the blank. |
content
|
Original expression from the text. |
sentence
|
Sentence the blank comes from, with the blank marked as <<BLANK>>. |
selector
|
How the blank was chosen: unigram, bigram, trigram, or ml for a language-model pick. |
ml_pos
|
Rank of the original word among the language model's predictions for the blank; -1 when it was not among them. |
prob, log_prob
|
Probability the language model assigned to the original word, and its natural logarithm. |
Language, lang_code, lang_code_2
|
Language of the text in three encodings: c/e/s, CZ/EN/SK, and CS/EN where Slovak texts are counted with Czech. |