摘要
为了促进和进一步发展RST（修辞结构理论）式的篇章分析模型，我们需要一个强大的基线模型，该模型可以作为报告可靠实验结果的参考。本文通过整合现有的简单解析策略（自上而下和自下而上）与各种基于变压器的预训练语言模型，探索了一个强大的基线模型。从两个基准数据集获得的实验结果表明，解析性能在很大程度上依赖于预训练语言模型而非解析策略。特别是，当使用DeBERTa时，自下而上的解析器相比当前最佳解析器实现了显著的性能提升。我们进一步分析了句内和跨句解析以及核心性预测，发现采用跨度掩码方案的语言模型特别提升了解析性能。
代码仓库
nttcslab-nlp/rstparser_emnlp22
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| discourse-parsing-on-instructional-dt-instr | Top-down (XLNet) | Standard Parseval (Full): 40.2 Standard Parseval (Nuclearity): 55.2 Standard Parseval (Relation): 47.0 Standard Parseval (Span): 74.3 |
| discourse-parsing-on-instructional-dt-instr | Bottom-up (DeBERTa) | Standard Parseval (Full): 44.4 Standard Parseval (Nuclearity): 60.0 Standard Parseval (Relation): 51.4 Standard Parseval (Span): 77.8 |
| discourse-parsing-on-instructional-dt-instr | Top-down (BERT) | Standard Parseval (Full): 30.9 Standard Parseval (Nuclearity): 44.6 Standard Parseval (Relation): 37.6 Standard Parseval (Span): 65.3 |
| discourse-parsing-on-instructional-dt-instr | Top-down (SpanBERT) | Standard Parseval (Full): 36.7 Standard Parseval (Nuclearity): 54.5 Standard Parseval (Relation): 42.7 Standard Parseval (Span): 73.7 |
| discourse-parsing-on-instructional-dt-instr | Top-down (DeBERTa) | Standard Parseval (Full): 43.4 Standard Parseval (Nuclearity): 57.9 Standard Parseval (Relation): 50.0 Standard Parseval (Span): 77.3 |
| discourse-parsing-on-instructional-dt-instr | Bottom-up (SpanBERT) | Standard Parseval (Full): 40.5 Standard Parseval (Nuclearity): 53.8 Standard Parseval (Relation): 46.0 Standard Parseval (Span): 72.9 |
| discourse-parsing-on-instructional-dt-instr | Top-down (RoBERTa) | Standard Parseval (Full): 41.5 Standard Parseval (Nuclearity): 56.1 Standard Parseval (Relation): 48.7 Standard Parseval (Span): 75.7 |
| discourse-parsing-on-instructional-dt-instr | Bottom-up (BERT) | Standard Parseval (Full): 32.9 Standard Parseval (Nuclearity): 46.3 Standard Parseval (Relation): 39.5 Standard Parseval (Span): 66.6 |
| discourse-parsing-on-instructional-dt-instr | Bottom-up (XLNet) | Standard Parseval (Full): 40.7 Standard Parseval (Nuclearity): 56.4 Standard Parseval (Relation): 47.4 Standard Parseval (Span): 73.6 |
| discourse-parsing-on-instructional-dt-instr | Bottom-up (RoBERTa) | Standard Parseval (Full): 41.4 Standard Parseval (Nuclearity): 55.5 Standard Parseval (Relation): 47.9 Standard Parseval (Span): 73.2 |
| discourse-parsing-on-rst-dt | Top-down (XLNet) | Standard Parseval (Full): 54.8 Standard Parseval (Nuclearity): 67.4 Standard Parseval (Relation): 57.0 Standard Parseval (Span): 77.8 |
| discourse-parsing-on-rst-dt | Top-down (SpanBERT) | Standard Parseval (Full): 52.2 Standard Parseval (Nuclearity): 65.4 Standard Parseval (Relation): 54.5 Standard Parseval (Span): 76.5 |
| discourse-parsing-on-rst-dt | Bottom-up (DeBERTa) | Standard Parseval (Full): 55.4 ± 0.4 Standard Parseval (Nuclearity): 68.0 ± 0.5 Standard Parseval (Relation): 57.3 ± 0.2 Standard Parseval (Span): 77.8 ± 0.3 |
| discourse-parsing-on-rst-dt | Bottom-up (RoBERTa) | Standard Parseval (Full): 53.7 Standard Parseval (Nuclearity): 66.5 Standard Parseval (Relation): 55.4 Standard Parseval (Span): 76.1 |
| discourse-parsing-on-rst-dt | Bottom-up (SpanBERT) | Standard Parseval (Full): 52.7 Standard Parseval (Nuclearity): 65.3 Standard Parseval (Relation): 54.9 |
| discourse-parsing-on-rst-dt | Top-down (DeBERTa) | Standard Parseval (Full): 54.4 Standard Parseval (Nuclearity): 67.9 Standard Parseval (Relation): 56.6 Standard Parseval (Span): 78.5 |
| discourse-parsing-on-rst-dt | Bottom-up (XLNet) | Standard Parseval (Full): 54.2 Standard Parseval (Nuclearity): 65.9 Standard Parseval (Relation): 56.3 |
| discourse-parsing-on-rst-dt | Top-down (RoBERTa) | Standard Parseval (Full): 53.8 Standard Parseval (Nuclearity): 66.6 Standard Parseval (Relation): 55.8 Standard Parseval (Span): 77.3 |
| discourse-parsing-on-rst-dt | Top-down (BERT) | Standard Parseval (Full): 46.6 Standard Parseval (Nuclearity): 59.1 Standard Parseval (Relation): 48.3 Standard Parseval (Span): 69.8 |
| discourse-parsing-on-rst-dt | Bottom-up (BERT) | Standard Parseval (Full): 46.0 Standard Parseval (Nuclearity): 57.8 Standard Parseval (Relation): 47.8 Standard Parseval (Span): 68.3 |