MIT Press

Table 8

Neural (S_token and S_type) logography measures computed on the Bible corpora. The first measure (denoted B) corresponds to the RNN configuration from Table 5 and is copied here to ease the comparison. The rest of the measures were computed for a single transformer model V, where V_A and V_M measures were computed by combining over multiple heads in the last decoder block by averaging and choosing the maximum element, respectively.

Language	B			V_A			V_M
Language	$S_{tok}^{B}$	$S_{typ}^{B}$	Acc.	$S_{tok}^{V_{A}}$	$S_{typ}^{V_{A}}$	Acc.	$S_{tok}^{V_{M}}$	$S_{typ}^{V_{M}}$
Chinese	1.00	1.00	0.85	0.87	0.84	0.95	0.80	0.74
Chinese (Cangjie)	0.74	0.71	0.87	0.79	0.76	0.94	0.65	0.61
Chinese (tok.)	0.55	0.37	0.89	0.90	0.81	0.93	0.91	0.82
Chinese (tok., Cangjie)	0.51	0.32	0.78	0.84	0.75	0.90	0.84	0.73
English	0.40	0.32	0.95	0.81	0.73	0.98	0.75	0.68
Finnish	0.19	0.12	0.96	0.34	0.24	0.98	0.37	0.26
French	0.57	0.36	0.89	0.52	0.35	0.94	0.51	0.33
Hebrew (Biblical)	0.65	0.50	0.94	0.44	0.30	0.96	0.39	0.28
Hebrew (Modern)	0.72	0.56	0.87	0.53	0.35	0.92	0.49	0.31
Japanese	0.97	0.88	0.94	0.94	0.85	0.96	0.92	0.83
Japanese (Cangjie)	0.88	0.65	0.92	0.90	0.83	0.96	0.90	0.84
Korean (jamo)	0.26	0.21	0.96	0.33	0.26	0.99	0.33	0.27
Russian	0.46	0.29	0.89	0.38	0.25	0.93	0.40	0.27
Swedish	0.35	0.20	0.90	0.43	0.27	0.91	0.47	0.33

Language	B			V_A			V_M
Language	$S_{tok}^{B}$	$S_{typ}^{B}$	Acc.	$S_{tok}^{V_{A}}$	$S_{typ}^{V_{A}}$	Acc.	$S_{tok}^{V_{M}}$	$S_{typ}^{V_{M}}$
Chinese	1.00	1.00	0.85	0.87	0.84	0.95	0.80	0.74
Chinese (Cangjie)	0.74	0.71	0.87	0.79	0.76	0.94	0.65	0.61
Chinese (tok.)	0.55	0.37	0.89	0.90	0.81	0.93	0.91	0.82
Chinese (tok., Cangjie)	0.51	0.32	0.78	0.84	0.75	0.90	0.84	0.73
English	0.40	0.32	0.95	0.81	0.73	0.98	0.75	0.68
Finnish	0.19	0.12	0.96	0.34	0.24	0.98	0.37	0.26
French	0.57	0.36	0.89	0.52	0.35	0.94	0.51	0.33
Hebrew (Biblical)	0.65	0.50	0.94	0.44	0.30	0.96	0.39	0.28
Hebrew (Modern)	0.72	0.56	0.87	0.53	0.35	0.92	0.49	0.31
Japanese	0.97	0.88	0.94	0.94	0.85	0.96	0.92	0.83
Japanese (Cangjie)	0.88	0.65	0.92	0.90	0.83	0.96	0.90	0.84
Korean (jamo)	0.26	0.21	0.96	0.33	0.26	0.99	0.33	0.27
Russian	0.46	0.29	0.89	0.38	0.25	0.93	0.40	0.27
Swedish	0.35	0.20	0.90	0.43	0.27	0.91	0.47	0.33

View Large

This Feature Is Available To Subscribers Only