Update README.md
Browse files
README.md
CHANGED
|
@@ -6,4 +6,9 @@ encoder-decoder model initialized with [huBERT](huggingface.co/SZTAKI-HLT/hubert
|
|
| 6 |
Our quantitative evaluation on automatic and human transcripts of news
|
| 7 |
(with automatic and human-made punctuation, [Tündik et al. (2019)](https://www.isca-speech.org/archive/interspeech_2019/tundik19_interspeech.html), [Tündik and Szaszák (2019)](https://www.isca-speech.org/archive/interspeech_2019/szaszak19_interspeech.html)) shows that the model is
|
| 8 |
robust with respect to errors in either automatic speech recognition or
|
| 9 |
-
automatic punctuation restoration.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 6 |
Our quantitative evaluation on automatic and human transcripts of news
|
| 7 |
(with automatic and human-made punctuation, [Tündik et al. (2019)](https://www.isca-speech.org/archive/interspeech_2019/tundik19_interspeech.html), [Tündik and Szaszák (2019)](https://www.isca-speech.org/archive/interspeech_2019/szaszak19_interspeech.html)) shows that the model is
|
| 8 |
robust with respect to errors in either automatic speech recognition or
|
| 9 |
+
automatic punctuation restoration. In fine-tuning and inference, we followed [a jupyter notebook by Patrick von
|
| 10 |
+
Platen](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb). Most hyper-parameters are the same as those by von Platen, but we
|
| 11 |
+
found it advantageous to change the minimum length of the summary to 8 word-
|
| 12 |
+
pieces (instead of 56), and the number of beams in beam search to 5 (instead
|
| 13 |
+
of 4). Our model was fine-tuned on a server of the [SZTAKI-HLT](hlt.bme.hu/) group, which kindly
|
| 14 |
+
provided access to it.
|