Abstractive Text Summarization Using Sequence-to-Sequence RNNs and Beyond | TDLS

Toronto Deep Learning Series, 7 August 2018
For slides and more information, visit tdls.a-i.scien...
Paper Review: arxiv.org/abs/...
Speaker: / ehsanamjadian
Organizer: / amirfz
Host: PWC
Paper abstract:
In this work, we model abstractive text summarization using Attentional Encoder-Decoder Recurrent Neural Networks, and show that they achieve state-of-the-art performance on two different corpora. We propose several novel models that address critical problems in summarization that are not adequately modeled by the basic architecture, such as modeling key-words, capturing the hierarchy of sentence-to-word structure, and emitting words that are rare or unseen at training time. Our work shows that many of our proposed models contribute to further improvement in performance. We also propose a new dataset consisting of multi-sentence summaries, and establish performance benchmarks for further research.

[@EhsanAmjadian]

Alright, this offers clarification on the "pointer network optimization" slide which displays the cost function, specifically concerning the gi indicator value in the optimization/cost function equation. First let's take a step back and talk about the switch itself. When the switch is on (i.e., si=1) the network generates and otherwise it copies. Now when it comes to the cost function, if indicator gi=1 then you'll penalize by the first term only meaning the generator; and when gi=0 you penalize by the second term in the sum only which is responsible for the copy mechanism.

[@EhsanAmjadian]

Also, this is to confirm that the 1sent and 2sent chunks on the results slides do indeed refer to the source document cut-off length for the specific method.

[@UmairKhan-kr9xe]

Sir how could we automatically / directly know from an output summary, that this is a good quality summary generated???

[@TafiVids]

ROUGE is the most used metric for summary evaluation. It uses the n-gram overlap of a reference and candidate summary as scoring. PYRAMID is another for human evaluation, mostly for multi-document summarization

[@EhsanAmjadian]

Good answer by @TafiVids . Here is the Original ROUGE paper. There are a couple of open source implementations in Perl (the original) and Java: github.com/kavgan/ROUGE-2.0 as well as Python: pypi.org/project/rouge-score/