sahilfarib/bangla-agreement-probes

Bangla agreement and honorific-register minimal pairs

4,790 minimal pairs probing Bangla subject–verb agreement in person and honorific register. Released with the MRL 2026 paper State Tracking vs. Attention in a Morphologically Rich Low-Resource Language: A Controlled Bangla Case Study with Mamba-3 (Farib, Meem, Islam, Arefin, and Anik).

Each pair has a grammatical sentence (sen) and a minimally different ungrammatical one (wrong_sen). The two sentences differ only in the inflection of the final verb.

File	Condition	Pairs
sva.tsv	Subject–verb agreement, with an intervener of distance none/short/medium/long	3,300
attraction.tsv	Agreement attraction: a pronoun of a different person sits inside the intervener, and the ungrammatical verb agrees with it	1,190
honorific.tsv	Honorific vs. ordinary third-person register (তিনি/উনি vs. সে)	210
discourse.tsv	Register set in sentence 1, agreement tested on a subjectless (pro-drop) sentence 2	90

Example (SVA)

	Sentence	Translation
✓	আমি আগামী সপ্তাহে একা একা বই পড়ব।	"I will read a book alone next week."
✗	আমি আগামী সপ্তাহে একা একা বই পড়বেন।	The verb is in the honorific form, but the subject is "I".

Columns

id, phenomenon, subj_person, wrong_person, lure_person, distance, tense, sen, wrong_sen

Code	Meaning
p1	1st person (আমি)
p2_int	2nd person intimate (তুই)
p2_fam	2nd person familiar (তুমি)
p2_hon	2nd person honorific (আপনি)
p3_ord	3rd person ordinary (সে, or a noun subject)
p3_hon	3rd person honorific (তিনি/উনি)

tense is pres, past, or fut. distance takes these values:

• SVA: none, short, medium, long
• Attraction: attractor
• Honorific: none
• Discourse: adjacent or one_filler

Construction

Pairs are generated deterministically from hand-written conjugation tables (10 verbs × 3 tenses × 6 person/register cells) and reusable sentence frames. A native speaker reviewed the lexicon rather than each individual pair: verb morphology, compatibility of time expressions with tense, naturalness of the intervening phrases, and discourse coherence. Every flagged issue was fixed before scoring. There was one reviewer and no second annotation pass. The generator and lexicon are in the code repository (bangla_ssm/probes/), and running it reproduces these files byte for byte.

Scoring

A model is correct on a pair if it assigns a higher total log-probability to sen than to wrong_sen. Total log-probability is not length-normalized. In 20–37% of pairs, depending on the condition, the two sentences split into a different number of subword tokens under the paper's tokenizer, and raw accuracy is sensitive to that difference. The paper therefore also reports results on length-matched pairs.

Citation

@inproceedings{farib2026statetracking,
  title     = {State Tracking vs. Attention in a Morphologically Rich Low-Resource Language:
               A Controlled {B}angla Case Study with {M}amba-3},
  author    = {Sahil Al Farib and Momota Ahsana Meem and Sheikh Redwanul Islam and
               Khan Md Shams Arefin and Azizur Rahman Anik},
  booktitle = {Proceedings of the 6th Workshop on Multilingual Representation Learning (MRL 2026)},
  year      = {2026},
  note      = {To appear}
}