📄 An Analysis of the Standard English Keyboard

Summary

Hiraga, Ono and Yamada examine how efficient the standard QWERTY layout really is. The work grew out of designing a Japanese-input keyboard. Instead of ranking key pairs by raw speed, they recorded inter-keystroke intervals to the millisecond from a proficient typist entering long English texts on the Superwriter system at the University of Tokyo. They first describe the time distribution of each key pair. They then fit a multivariate linear model that splits interval time into usage frequency, hand alternation, row and finger transitions, and row and finger weights. Alternating hands proves to be the strongest speed-up, which matches Dvorak's earlier claims. Their central new finding is that a key pair's timing depends on its surrounding context, likely the word. Hard sequences slow down the easy ones nearby, and an expert evens out the rhythm. Because of this, measurements taken on QWERTY understate its defects and cannot reliably predict performance on other layouts such as the Dvorak Simplified Keyboard. The authors conclude that QWERTY is far from ideal.

Key points

• The corpus has 302,392 keystrokes: 142,262 from the Alice books (Carroll, verses excluded) and 160,130 from Bohl's "Information Processing" textbook, all typed by one professional typist.
• Intervals over 500 ms were discarded as non-typing pauses. Analysis focuses on a "main set" of 30 keys, and 805 of its 900 possible pairs occurred in the data.
• Most mean intervals fall between 100 and 200 ms. The physical floor is about 60 ms, except for same-finger pairs such as reaches and taps. The distributions are sharply peaked and skewed right.
• In the regression, hand alternation has a coefficient of −40.0 ms on an intercept of 185.8, making it the dominant factor. Row transitions add time, and higher frequency lowers it.
• On QWERTY, the left hand and upper row are overloaded. 9.6% of key pairs are hurdles and 8.2% are reaches.

Interesting ideas

• I1. Raw per-item speed rankings gave no insight, so the authors factored the times into a few interpretable attributes with linear regression and then tested hypotheses on those factors.
• I2. Performance on a unit depends on its context, so per-pair measurements in one setting do not combine into a correct prediction for another.
• I3. A skilled performer's "levelling" smooths out difficulties, so averages can hide real weaknesses. Lower bounds, skew and tails reveal more than means do.
• I4. A simple threshold removed intervals caused by events outside the task, such as page turning, before any modelling.
• I5. Frequency was modelled as a separate regressor so that familiarity effects were not confused with structural difficulty.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Parser or tagger errors could be regressed on explainable features such as arc length, POS pair and frequency, so that failure causes are named rather than just ranked.
• I2 (learn): Per-relation accuracy from one treebank should not be extrapolated to a new UD dialect or domain, and end-to-end evaluation on the target setting should be required.
• I3 (answer): Test scores should report the distribution of hard cases, such as long chains and rare constructions, alongside aggregate accuracy, so that a smooth average does not mask brittle reasoning.
• I4 (learn): Silver data could be filtered for examples whose disagreement comes from extraction glitches rather than linguistic difficulty before training.
• I5 (tree): Attachment difficulty analyses could control for word or construction frequency, so that structural hardness is not mistaken for rarity.