search

shashiongithub / sidenet

SideNet: Neural Extractive Summarization with Side Information
BSD 3-Clause "New" or "Revised" License
57 stars 10 forks source link

Word level extractive summarization? #6

Open Hellisotherpeople opened 5 years ago

Hellisotherpeople commented 5 years ago

How difficult would it be to turn your code into a word-level model? That is, to select the best words to include in the summary, instead of at the sentence level. This task would be extremely useful to people in the Competitive Debate community, who need to underline evidence based on a tag-line, and cannot include words not found in the text (making abstractive summarization impossible)

shashiongithub commented 5 years ago

The Idea presented here could be easily applied to word level extraction. I am not sure if you want to keep the convolutional sentence decoder (it will depend on the encoder if you want to model the input document as a seq of sentences or words). The decoder will simply tag one word at a time instead of a sentence.

Hellisotherpeople commented 5 years ago

I'm not especially picky on how it's done. Do you have any real interest in going further with this? I have a dataset of debate evidence which is highlighted and has abstractive side information to include. I also have a burning passion to get this exact model created

Maybe a paper could come out of it?

shashiongithub commented 5 years ago

Hi Allen, Sounds very interesting. However, I have moved from Edinburgh recently. I would be very happy to discuss with you in depth about any implementation issues.

Hellisotherpeople commented 5 years ago

Can I contact you via the email you provided here? I'd like to get a word level version of sidenet working.

shashiongithub commented 5 years ago

Sure no problem!