typing37k-clean / README.md
MostLime's picture
Cleaned Typing37k: data, changelog, results, card
a3ab739 verified
|
Raw History Blame Contribute Delete
9.48 kB
metadata
license: other
language:
  - en
pretty_name: Typing37k (cleaned)
size_categories:
  - 10K<n<100K
task_categories:
  - text2text-generation
tags:
  - spelling-correction
  - typo-correction
  - mobile-typing
  - benchmark
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/typing37k_clean.jsonl

Typing37k (cleaned)

A cleaned, word-aligned typo-correction benchmark built from Typing37k:

Palin, K., Feit, A. M., Kim, S., Kristensson, P. O., & Oulasvirta, A. (2019). How do people type on mobile devices? Observations from a study with 37,000 volunteers. MobileHCI 2019. https://userinterfaces.aalto.fi/typing37k/

In Typing37k, volunteers retyped a shown sentence on their phones. This set keeps the sections typed on a QWERTY phone keyboard with no autocorrect, prediction or swipe, where the typed text has the same number of words as the target. Each typed word is aligned to its target word, and a corrector is scored per word slot.

What changed

The original targets are kept as they are. On top of them, word slots that one of two correction models (Keylime m2 and m3) got wrong were reviewed, and three kinds of change were made:

  • Gold fixed (17 target slots): the target itself had a typo or a wrong word, e.g. "Dijoubti" -> "Djibouti", "second half on October" -> "of".
  • Alternatives added (57 slots from the target review + 629 from the message review): another word is equally correct in that slot. Examples: British/US spellings, "question"/"questions" after "any", name variants, typed words that read just as well in context, and contractions whose apostrophe was lost in the target ("didn t": "didn't" is also accepted).
  • Removed (616 unique typed messages = 689 messages): at least one target word cannot be inferred from what was typed. The typist misremembered or paraphrased the prompt ("Katis's store" for "Kari's house"), skipped or added a word so the slots shift, or typed something unreadable.

Fixes and alternatives belong to the target sentence, so they apply to every typed message of that target. status = fixed marks the 12,749 rows (of 34,151) whose target has at least one changed slot.

Difficulty shift. Removing and relabelling slots changes what is left, so scores are not comparable with the original set. Slot counts, weighted by messages:

change slots
correct -> typo (target fixed, typed word = old gold) 1,435
typo -> correct (typed word accepted as an alternative) 1,837
correct word removed with its message 4,400
typo removed with its message, no system fixed it 1,051
typo, alternatives added (still a typo) 2,009
typo, gold word changed (still a typo) 89
typo removed with its message, some system fixed it 183

Almost every slot that left the typo pool was one that no system fixed. Because of that, every system's fix rate goes up, by 0.8 to 4.6 points.

How

  1. Items: every unique typed message where Keylime m2 or m3 (whole-sentence) got at least one word wrong: 12,468 messages, 14,949 word slots, on 1,490 target sentences.

  2. Target-level review (Claude Haiku, 100 calls): each of those targets was checked once for a wrong gold word or equally correct alternatives. The prompt gave the real words people typed in each slot and the system outputs as anonymous hints.

  3. Message-level review: 4,011 typo slots that none of the 43 saved system runs fixed (every baseline and Keylime run, both configurations). Claude Haiku first judged each one as a real typo or uninferable (101 calls). Slots that some system did fix count as inferable.

  4. Verification (Claude Sonnet), blind and with model outputs anonymised. Agreement with Haiku:

    • target level (Haiku vs Sonnet, blind sample of 110 slots): gold fixed 10/17 = 58.8%, alternatives 25/43 = 58.1%, no issue 50/50 = 100.0%
    • message level (Haiku vs Sonnet, blind sample of 140 slots, two-way rule): uninferable 83/90 = 92.2%, real typo 31/50 = 62.0%

    Both target-level categories were under 85%, so every target with a Haiku fix or alternative was re-judged by Sonnet, whose verdict is final (the 50 "no issue" slots agreed 100%). At message level the "real typo" category agreed only 62%: Sonnet treated many context-free word swaps (the/this, here/there) as unrecoverable. Removing those would drop the hardest real-word errors from the set, so all 4,011 leftover slots were re-judged by Sonnet with a stricter three-way rule: remove only misremembered/paraphrased or unreadable text; accept the typed word when it is also a fine reading in context; otherwise keep it as a real typo, however hard. Sonnet's verdicts: typo 2,168, uninferable 941, alt_ok 902.

  5. A few verdicts were corrected by hand. Fixes that need two words in one slot were dropped, and lost apostrophes were handled by rule. These are marked manual review in the changelog.

changelog.csv has one row per changed target slot or removed message: typed text, old gold, new gold / alternatives, verdict, reason, number of messages and the judge.

Scoring rule

A slot is a typo if the typed word is not in the slot's gold set (the gold word plus gold_alts), and it is fixed if the output is in the gold set. False correction (FC) = a slot whose typed word is in the gold set but the output is not. Comparison is case-insensitive. Punctuation and digits are not scored. Each unique (typed message, target) row is weighted by count, the number of sections with that exact typed text. An empty string in gold_alts means the slot may be empty, because the word was joined to the previous slot (only for lost apostrophes).

Counts

original cleaned
messages (sections) 204,825 204,136
unique typed messages 34,151 33,535
words 1,511,164 1,505,530
typos 41,279 39,643

The file keeps every row, including removed ones (status = removed); drop those rows for scoring.

Fields

  • pair_id: row index; sentence_id: SENTENCE_ID from the original dump
  • target_text, seps (the text between words, from the target), gold_words_original, gold_words (with fixes), gold_alts (extra accepted words per slot)
  • typed_text, typed_words, count, sections ([TEST_SECTION_ID, PARTICIPANT_ID] pairs from the original dump)
  • status: unchanged, fixed or removed; note: what changed

Results

Typo-fix rate and false-correction rate on the original and the cleaned set, using the same outputs and the same aligner as the original benchmark.

Whole sentence (each system sees the whole typed message):

system fix % (orig) FC % (orig) fix % (clean) FC % (clean)
SAGE mT5-large spelling (1.2B, ru+en) 72.28 0.771 76.86 0.646
Keylime m2 (106M, int4) 71.99 0.346 75.88 0.301
Keylime m2 + cand-gen v2 71.98 0.374 75.83 0.319
T5-large-spell (738M) 69.98 0.648 74.06 0.572
Keylime m3 (61.7M, int4) 67.97 0.446 72.33 0.330
LFM2.5-Encoder-350M Spellchecker 60.17 0.917 64.17 0.672
T5-small spoken-typo (61M) 59.07 1.042 63.02 0.889
BART-base spelling (140M) 57.25 0.540 60.23 0.470
JamSpell 3-gram 50.88 0.338 53.86 0.306
pyspellchecker 45.46 1.103 47.81 1.022
SymSpell word 44.62 1.657 46.60 1.616
Apple spellchecker (NSSpellChecker) 44.47 0.194 46.55 0.139
CoEdIT-large (783M, GEC) 43.13 1.553 46.54 1.345
SymSpell compound 43.03 2.127 44.92 2.079
vennify T5-base GEC (223M) 38.62 0.904 41.42 0.779
Hunspell en_US 35.68 1.173 37.48 1.093

Left context only (each system sees only the words typed so far):

system fix % (orig) FC % (orig) fix % (clean) FC % (clean)
Keylime m2 (106M, int4) 65.24 0.485 68.68 0.441
Keylime m2 + cand-gen v2 65.19 0.506 68.63 0.462
T5-large-spell (738M) 51.99 5.503 54.82 5.422
T5-small spoken-typo (61M) 46.69 4.935 49.34 4.854
SAGE mT5-large spelling (1.2B, ru+en) 45.21 9.146 48.17 9.005
pyspellchecker 45.45 1.103 47.81 1.022
SymSpell word 44.62 1.657 46.60 1.616
Apple spellchecker (NSSpellChecker) 44.46 0.194 46.55 0.138
JamSpell 3-gram 42.75 0.201 45.16 0.174
SymSpell compound 43.23 1.981 45.14 1.933
BART-base spelling (140M) 41.53 2.072 43.52 1.989
LFM2.5-Encoder-350M Spellchecker 36.77 1.147 39.09 1.050
Hunspell en_US 35.68 1.173 37.48 1.093
CoEdIT-large (783M, GEC) 19.40 3.772 20.60 3.679
vennify T5-base GEC (223M) 12.01 19.849 12.78 19.788

License

Derived from Typing37k (Palin et al. 2019), which the authors released as openly available. Original data © the authors.

Citation

@inproceedings{palin2019typing,
  title     = {How do People Type on Mobile Devices? Observations from a Study with 37,000 Volunteers},
  author    = {Palin, Kseniia and Feit, Anna Maria and Kim, Sunjun and Kristensson, Per Ola and Oulasvirta, Antti},
  booktitle = {Proceedings of the 21st International Conference on Human-Computer Interaction with Mobile Devices and Services (MobileHCI '19)},
  year      = {2019},
  url       = {https://userinterfaces.aalto.fi/typing37k/}
}