kylelovesllms/grammar-v3-frames-ce

Grammar V3 (determiners and PPs): CE Populated Frames

The 6744 depth-1 syntactic frames of grammar-v3-frames (Grammar V3, experiments/03_grammar_v3/src/cfg_v3/grammar.py in the translation-syntax project), each populated with 2 randomly sampled lexicalizations per task (seeded RNG, seed=0).

Grammar V3 is the number-agreement grammar with optional determiners (Grammar V2) and prepositional phrases adjoined to NP. In Head Final the PP moves in front of the noun it modifies, inside that noun's DP, and the preposition becomes a postposition: the dog near a cat chases this bird translates to the a cat near dog this bird chases. PPs attach only to the subject and the object of the matrix clause; a noun takes at most one modifier (a PP or a relative clause), and the nouns inside a PP or a relative clause take none.

The lexicon is Grammar V2's (26 nouns, 26 intransitive and 26 transitive verbs in both numbers, 26 adverbs, the determiners this, a, these, the and the complementizer that) with 26 prepositions: near, behind, beside, above, below, under, over, beyond, around, against, among, beneath, underneath, by, past, opposite, toward, with, without, alongside, before, after, between, inside, outside, atop.

Partition

A frame is ambiguous if the linear rule gives the hierarchical translation. The linear rule splits the sentence at the first verb and emits each half with its determiners first, in source order, followed by its other words reversed. A frame whose only modifiers are PPs is ambiguous.

• ambiguous frames (2,440): 2,196 are in train with copy and translate rows, 244 are held out and are in validation with copy and translate rows
• unambiguous frames (4,304): copy rows in train, translate rows in test

Validation holds out frames, not sentences: 10% of the ambiguous frames of every category, spread over the PP sites and determiner conditions, and only frames that keep another number variant in train. Copy and translate rows of a frame never share a sentence, so no test translate input appears as a train copy input.

3,514 of the 4,392 train translations (80%) have a PP on the subject, whose words the translation reorders. In Grammar V2 every train translation copied its subject verbatim.

Copy-task variant ce: copy simple (no RC) and center embedded (subject-attached RC) structures only

Train copy rows are kept only for frames whose relative clause attaches at one of these sites: None, Subject (None = no relative clause, with or without PPs). The validation and test splits are identical across the ce-rb, ce and rb datasets. Every frame has exactly 2 sentences per task in all three, so the datasets differ in size: the dropped copy rows are not replaced.

• train: 11,952 rows = 4,392 translate + 7,560 copy, 280,020 tokens (<bos> src <task> <sep> tgt <eos> = 2 * n_tokens + 4 per row)
• validation: 976 rows (244 held-out frames, copy and translate)
• test: 8,608 rows (translate, unambiguous)

Train rows per structure

Task	Attachment Site	Gap Type	Frames	Rows per frame	Rows	Tokens
translate	None	None	756	2	1,512	31,284
translate	Object	Subject Gap	1440	2	2,880	70,284
copy	None	None	756	2	1,512	31,284
copy	Subject	Subject Gap	1680	2	3,360	80,864
copy	Subject	Object Gap	1344	2	2,688	66,304
Total					11,952	280,020

Columns

HF Sentence is the task target: the HI sentence itself for copy rows, the hierarchical (Head Final) translation for translate rows. Frames use indexed bracketed notation where [N_sg_3] means the terminal filled by HI token 3. Matrix Number is the number of the matrix subject, RC Number that of the noun the RC verb agrees with (None without an RC), and n_attractors counts nouns of the other number between a verb and the noun it agrees with.

PP Site lists the arguments of the matrix clause that carry a PP: no-PP, subject, object or subject+object. Determiners lists the nouns that carry a determiner (subject, object, subject-PP and object-PP = the noun of the PP on that argument, RC = the noun inside the relative clause). Determiner Condition groups them: no-determiner, matrix (only nouns outside the relative clause), RC (only the noun inside the RC) and matrix-RC (both). Tier is ID (in distribution) for ambiguous frames, T1 for Subject / Subject Gap, T2 for Object / Object Gap and T3 for Subject / Object Gap.

Frame counts per category (before population)

Attachment Site	Gap Type	Frames	Ambiguous	Unambiguous
None	None	840	840	0
Subject	Subject Gap	1680	0	1680
Subject	Object Gap	1344	0	1344
Object	Subject Gap	1600	1600	0
Object	Object Gap	1280	0	1280
Total		6744	2440	4304

Sentence counts per category (after population)

Over every split and both tasks. Sample factor is the number of sentences per frame, copy and translate together; it is lower where the variant dropped a structure's train copy rows.

Attachment Site	Gap Type	Frames	Ambiguous	Unambiguous	Sentences	Sample factor	RC number = matrix number	RC number differs
None	None	840	840	0	3,360	4	—	—
Subject	Subject Gap	1680	0	1680	6,720	4	6,720	0
Subject	Object Gap	1344	0	1344	5,376	4	2,688	2,688
Object	Subject Gap	1600	1600	0	3,520	2–4	1,754	1,766
Object	Object Gap	1280	0	1280	2,560	2	1,280	1,280
Total		6744	2440	4304	21,536		12,442	5,734

Translate sentences per PP site

Split	Attachment Site	Gap Type	no-PP	subject	object	subject+object	Sentences
train	None	None	72	288	230	922	1,512
train	Object	Subject Gap	576	2,304	0	0	2,880
validation	None	None	8	32	26	102	168
validation	Object	Subject Gap	64	256	0	0	320
test	Subject	Subject Gap	800	0	2,560	0	3,360
test	Subject	Object Gap	640	0	2,048	0	2,688
test	Object	Object Gap	512	2,048	0	0	2,560

Translate sentences per determiner condition

Split	Attachment Site	Gap Type	no-determiner	matrix	RC	matrix-RC	Sentences
train	None	None	152	1,360	0	0	1,512
train	Object	Subject Gap	258	1,470	174	978	2,880
validation	None	None	16	152	0	0	168
validation	Object	Subject Gap	30	162	18	110	320
test	Subject	Subject Gap	336	1,680	224	1,120	3,360
test	Subject	Object Gap	224	1,120	224	1,120	2,688
test	Object	Object Gap	192	1,088	192	1,088	2,560