KalParso: informal Hinglish planner entries with exact dates and times

Planner entries in Roman-script Hindi-English (and some English and Devanagari), each with the moment it was written and the exact date and time it refers to. "parso shaam saade 5 baje dentist" written on Saturday 31 October 2026 at 14:00 means Monday 2 November, 17:30. Kal means tomorrow or yesterday depending only on the verb's tense, which is what most systems get wrong.

Code, models and the paper's numbers: https://github.com/imsg8/kalparso

Configurations

Config	Split	Rows	What
synthetic	train	29,911	Template-generated entries with character-level time spans
synthetic	validation	1,988	Same generator, training wording
synthetic	test	1,999	Held-out wording: frames and actions never seen in training
human	calibration	262	Real entries; used only to fix grounding conventions
human	test	748	Real entries, held out (706 scored; 42 are out of scope, such as ranges or festivals)

from datasets import load_dataset
train = load_dataset("imsg8/kalparso", "synthetic", split="train")
test = load_dataset("imsg8/kalparso", "human", split="test")

Fields

Both configs: id, text, anchor (when the entry was written, YYYY-MM-DDTHH:MM), weekday (of the anchor), type (reminder, note or goal), direction (future or past, the tense), date (YYYY-MM-DD or null), time (HH:MM or null), flags.

Synthetic only: spans, a list of {start, end, label} character spans with labels DAY (kal, parso), WD (weekdays), DATE (15 tareekh), TIME (saade 5 baje), PART (shaam), DUR (2 ghante baad).

Human only:

• fine_type: reminder, deadline, edit, past event, note or goal.
• kind: point, by (a deadline, scored on the date) or out_of_scope.
• part_only: the clock time is a default for a part of the day, so date-only scoring is possible.
• late_night, date_strict, time_strict: for entries written after midnight, the gold answer follows the speaker's day, and the strict calendar answer is also given. Both are accepted.
• notes: the annotator's explanation.

How it was made

• Synthetic: src/generate.py in the code repository builds entries from templates and the lexicon in src/lexicon.py, then computes each date with the deterministic grounder from the gold spans, so labels follow the written conventions (spec.md) by construction. Rows whose text matches a human entry (ignoring case and punctuation) are removed (scripts/decontaminate.py).
• Human: 1,010 entries typed and labelled by the first author in the style of everyday planner entries, part of them paraphrased from real planning messages with names, places and other identifying details removed. They keep their informal spelling, including misspelled time words (cal for kal, sat for saat). They are grouped under 20 anchor moments between 31 October 2026 and 5 September 2027, and 113 sentences recur under different anchors on purpose. The entries contain no personal information. The split is by sentence, so a repeated sentence never crosses it (human/split.json).

Use

• Never train on human/test. Report date-and-time accuracy on its 706 scored rows with src/eval_human.py.
• The grounding conventions (what agle Friday, a bare 8 baje or kal raat 2 baje mean) are written down in spec.md. They were fixed on the calibration split only.
• Known limits: one speaker's conventions; Roman-script Hinglish mostly; no formal Devanagari Hindi.

Licence and citation

CC BY 4.0. Shivang Gulati and Archana Praveen Kumar, Manipal Institute of Technology, MAHE. Paper under review; see CITATION.cff in the code repository.