kylelovesllms/grammar-v2-frames

Grammar V2 (determiners): Depth-1 Syntactic Frames

All syntactic frames with at most one relative clause generated by Grammar V2, the paired Head Initial / Head Final CFG with subject-verb number agreement and optional determiners of the translation-syntax project (experiments/02_grammar_v2/src/cfg_v2/grammar.py). The grammar is Grammar V1 + Number Agreement with a DP layer: every argument (matrix subject, object, embedded subject) is a bare noun or a determiner plus noun, and the determiner agrees in number with its noun (D_sg: this, a, the; D_pl: these, the). The determiner keeps its position in Head Final while the relative clause moves in front of the noun: the dog that these birds like chases cat translates to the these birds like that dog cat chases.

Frames are category-level skeletons to be populated with lexical items; terminals carry their 1-based HI linear position (e.g. [N_pl_1]), so HF frames show where each HI slot moved. The populated datasets are grammar-v2-frames-ce-rb, grammar-v2-frames-ce and grammar-v2-frames-rb.

1336 frames (360 ambiguous / in distribution, 976 unambiguous / out of distribution; the number variants of 212 number-erased frames), each expanded into a copy row (HF Frame = HI Frame, the copy target) and a translate row (HF Frame = hierarchical translation), giving 2672 rows in frames.csv.

Columns

is_ambiguous is True iff the linear rule gives the hierarchical translation. The linear rule splits the frame at the first verb and emits each half with its determiners first, in source order, followed by its other words reversed. Ambiguous frames are translated in training; unambiguous frames are copied in training and translated only at test. first_divergence is the 1-based output position of the first word on which the linear rule and the hierarchical translation differ (empty if ambiguous).

fits_mixed, fits_that_aware_pivot and fits_prefix_copy make the same comparison for three other rules. Mixed: the linear procedure split at the matrix verb. That-aware pivot: split at the first verb that does not directly follow that. Prefix copy: the words before the first verb are copied verbatim.

Tier is ID (in distribution) for ambiguous frames, T1 for Subject / Subject Gap, T2 for Object / Object Gap and T3 for Subject / Object Gap. No pivot rule translates a frame of T2 or T3.

Determiners lists the arguments that carry a determiner (subject, object, RC = the noun inside the relative clause). Determiner Condition groups them: no-determiner, matrix (only arguments of the matrix clause), RC (only the noun inside the RC) and matrix-RC (both).

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 (only center-embedded frames have any).

Counts per Category (frames, before task expansion)

Attachment Site	Gap Type	Frames	Ambiguous	Unambiguous
None	None	40	40	0
Subject	Subject Gap	400	0	400
Subject	Object Gap	320	0	320
Object	Subject Gap	320	320	0
Object	Object Gap	256	0	256
Total		1336	360	976

Frames each rule translates correctly

Attachment Site	Gap Type	Tier	Frames	Linear	Mixed	That-aware pivot	Prefix copy
None	None	ID	40	40	40	40	40
Subject	Subject Gap	T1	400	0	400	400	0
Subject	Object Gap	T3	320	0	0	0	0
Object	Subject Gap	ID	320	320	320	320	320
Object	Object Gap	T2	256	0	0	0	0
Total			1336	360	760	760	360

Frames per Determiner Condition

Attachment Site	Gap Type	Tier	no-determiner	matrix	RC	matrix-RC	Frames
None	None	ID	12	28	0	0	40
Subject	Subject Gap	T1	72	168	48	112	400
Subject	Object Gap	T3	48	112	48	112	320
Object	Subject Gap	ID	48	144	32	96	320
Object	Object Gap	T2	32	96	32	96	256
Total			212	548	160	416	1336