Semantic frame induction through the detection of communities of verbs and their arguments

Applied Network Science

Table 1 Semantic frame induction results achieved using different contextualized representations of verb instances

	d	Edges	CC	Clusters	Purity F₁	BCubed F₁
BERT	0.21	18,945	0.66	41.13 ±0.56	89.03 ±1.02	82.76 ±1.22
Context-Free	0.57	39,441	0.97	22.03 ±0.31	95.57 ±0.58	93.35 ±0.71
Syntactic Context	0.41	27,201	0.80	30.93 ±0.25	94.32 ±0.16	91.65 ±0.20
Semantic Context	0.53	20,660	0.67	24.47 ±0.52	88.64 ±0.33	81.92 ±0.54
Free + Syntactic	0.47	37,913	0.94	22.97 ±0.41	95.83 ±0.28	93.66 ±0.32
All	0.45	21,448	0.73	34.73 ±0.51	93.93 ±0.30	91.04 ±0.59

The first row refers to BERT representations, the second block to the different levels of ELMo representations, and the last block to combinations of those levels. d refers to the neighboring threshold. CC refers to the clustering coefficient