Abstractive Summarization on How2 (test)
48.9Content F1Multi-source Sequence-to-Sequence Model with Hierarchical Attention
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Multi-source Sequence-to-Sequence Model with Hierarchical AttentionModel No.=8, Input=Ground-truth transcript + Action, Modality=Multimodal2019.06 | 48.9 | 54.9 | — | |
| Multi-modal ModelPretraining Data=How22020.08 | 48.9 | 59.2 | 59.3 | |
| S2S Complete TranscriptModel No.=5a, Modality=Text-only, Tokens=6502019.06 | 47.4 | 53.9 | — | |
| Pointer-Generator (PG) Complete TranscriptModel No.=5b, Modality=Text-only2019.06 | 42 | 50.2 | — | |
| BertSum with ordered trainingPretraining Data=How2, WikiHow, CNN/DM2020.08 | 36.4 | 44.02 | 48.26 | |
| Using Extracted Sentence from 2a onlyModel No.=3, Modality=Text-only2019.06 | 36 | 46.4 | — | |
| Action Features + RNNModel No.=7, Modality=Video2019.06 | 34.9 | 46.3 | — | |
| ASR output Complete TranscriptModel No.=5c, Modality=Text-only2019.06 | 34.7 | 46.1 | — | |
| Multi-source Sequence-to-Sequence Model with Hierarchical AttentionModel No.=9, Input=ASR output + Action, Modality=Multimodal2019.06 | 34.7 | 46.3 | — | |
| BertSum with random training and postprocessingPretraining Data=How2, 1/50 Sampled-WikiHow, CNN/DM2020.08 | 32.9 | 22.47 | 26.32 | |
| First 200 tokensModel No.=4, Modality=Text-only2019.06 | 27.5 | 40.3 | — | |
| BertSum (with/without pre+post processing)Pretraining Data=CNN/DM2020.08 | 26 | — | — | |
| Action Features onlyModel No.=6, Modality=Video2019.06 | 24.8 | 38.5 | — | |
| Rule-based Extractive summaryModel No.=2a2019.06 | 18.8 | 16.4 | — | |
| BertSum with random trainingPretraining Data=How2, 1/50 Sampled-WikiHow, CNN/DM2020.08 | 18.7 | 21.45 | 24.4 | |
| Next-neighbor SummaryModel No.=2b, Topic space=LDA2019.06 | 17.9 | 31.8 | — | |
| Lead 3 for How2Pretraining Data=Not Applicable2020.08 | 16.2 | 20.69 | 23.66 | |
| Random Baseline using Language ModelModel No.=12019.06 | 8.3 | 27.5 | — |