wenmin-wu/nlp-keystroke-essay-reconstruction

Overview

When a dataset only gives you keystroke events — activity, cursor_position, text_change — and not the final text, you can deterministically rebuild the essay (or at least its anonymized structure) by replaying each event against a running string buffer. This unlocks a whole second feature family: sentence / paragraph / word-length statistics that are impossible to compute from event logs alone. Used in the Kaggle "Linking Writing Processes to Writing Quality" competition to add ~20-40 text-shape features on top of event-based ones.

Quick Start

def reconstruct_essay(df):
    df = df[df['activity'] != 'Nonproduction']
    out = {}
    for uid, g in df.groupby('id'):
        essay = ""
        for act, cur, txt in g[['activity','cursor_position','text_change']].values:
            if act == 'Replace':
                old, new = txt.split(' => ')
                essay = essay[:cur-len(new)] + new + essay[cur-len(new)+len(old):]
            elif act == 'Paste':
                essay = essay[:cur-len(txt)] + txt + essay[cur-len(txt):]
            elif act == 'Remove/Cut':
                essay = essay[:cur] + essay[cur+len(txt):]
            elif 'M' in act:  # Move From [a,b] To [c,d]
                pass  # splice slices per parsed offsets
            else:  # Input
                essay = essay[:cur-len(txt)] + txt + essay[cur-len(txt):]
        out[uid] = essay
    return out

Workflow

1. Drop Nonproduction rows (mouse clicks, focus changes) — they don't modify text
2. Group events by session id and iterate in order
3. For each activity type, splice the buffer at cursor_position using text_change length
4. Keep the buffer as anonymized characters — the output is a string of q's and whitespace, same shape as the real essay
5. Feed the reconstructed essay to downstream sentence/paragraph/word statistics

Key Decisions

• Anonymized is fine: you only need lengths, counts, and boundaries — not content. The q-replaced text preserves everything statistical you need.
• Replace vs. Input: Replace carries both old and new string separated by =>. Parse them to compute the delta correctly, otherwise offsets drift.
• Move is the hardest: parse From [a,b] To [c,d] and splice three slices. Skip it if rare — most sessions have <1% Move events.
• vs. text_change aggregation: counting text_change values misses deletions and the final layout; only replay recovers structure.

References

• LGBM (X2) + NN + Fusion
• Silver Bullet | Single Model | 165 Features