RST Discourse Parsing on GUM Corpus (test)
76.4Span F1Llama 2 (70B)
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Llama 2 (70B)Parsing Strategy=Bottom-up, Backbone=Llama 2, Parameters=70B, Quantization=QLoRA 4-bit2024.03 | 76.4 | 64.7 | 56.4 | 55.2 | |
| Llama 2 (70B)Parsing Strategy=Top-down, Backbone=Llama 2, Parameters=70B, Quantization=QLoRA 4-bit2024.03 | 75.8 | 64 | 55.8 | 54.8 | |
| Llama 2 (13B)Parsing Strategy=Top-down, Backbone=Llama 2, Parameters=13B, Quantization=QLoRA 4-bit2024.03 | 74.9 | 62.5 | 53.8 | 52.5 | |
| Llama 2 (13B)Parsing Strategy=Bottom-up, Backbone=Llama 2, Parameters=13B, Quantization=QLoRA 4-bit2024.03 | 74.8 | 63.4 | 54 | 52.8 | |
| Kobayashi et al.Parsing Strategy=Top-down, Backbone=DeBERTa-base, Parameters=140M2024.03 | 74.4 | 62.2 | 50.9 | 48.7 | |
| Llama 2 (7B)Parsing Strategy=Bottom-up, Backbone=Llama 2, Parameters=7B, Quantization=QLoRA 4-bit2024.03 | 74.4 | 63 | 53.4 | 52.1 | |
| Kobayashi et al.Parsing Strategy=Bottom-up, Backbone=DeBERTa-base, Parameters=140M2024.03 | 73.4 | 60.9 | 50.3 | 48.5 | |
| Llama 2 (7B)Parsing Strategy=Top-down, Backbone=Llama 2, Parameters=7B, Quantization=QLoRA 4-bit2024.03 | 72.8 | 60.9 | 52.1 | 50.9 | |
| Guz et al.Parsing Strategy=Bottom-up, Backbone=SpanBERT-base, Parameters=110M2024.03 | 69.9 | 57 | — | — | |
| Liu et al.Parsing Strategy=Top-down, Backbone=XLM-ROBERTa-base, Parameters=125M2024.03 | 68.6 | 54.9 | — | — |