-
## 一言でいうと
文章要約で教師有と強化学習を併用したという話。モデルはSeq2Seq+Attention(リピート防止にDecoder側も参照)。単純な教師有だと「正解」に固執する傾向があるため、自由度を持たせた強化学習も導入(ROUGEにより良さを評価)。補完的に働かせ良好な結果が得られた
### 論文リンク
https://arxiv.org/abs/1705.04304…
-
|논문|Concept Pointer Network for Abstractive Summarization|
|--|--|
|저자|Wenbo Wang, Yang Gao, Heyan Huang, Yuxiang Zhou|
|링크|https://www.aclweb.org/anthology/D19-1304/|
|학회|EMNLP|
# Contents
##…
-
Basically combine all the different approaches into a single benchmark, similar to GLUE.
Also see if there is some way to extend evaluation beyond ROUGE.
-
### Description
I run in Google Colab the following code
```
!pip install --upgrade
!pip install -q git+https://github.com/microsoft/nlp-recipes.git
!pip install jsonlines
!pip install pyro…
-
For Abstractive Summarization Task, I run bash preprocess_cnndaily_bin.sh path/to/cnndm_raw_data after downloading [CNN dataset] and [Daily Mail dataset]. However, I met the following problem:
![file…
-
# Currently working on...
## UI implementation for viewing sources in chat
- [x] Extracting sources from RAG chain
- [x] Adding sources to the UI (references drop down menu inspiration: https://…
-
- Abstract (2-3 lines)
Text Summarization which seems pretty simple but here's the fun fact, No certain breakthroughs has been made in this field which is open-source and gives a significant ROUGE sc…
-
## 0. 論文
[Fast Abstractive Summarization with Reinforce-Selected Sentence Rewriting](https://arxiv.org/abs/1805.11080v1)
Yen-Chun Chen, Mohit Bansal
## 1. どんなもの?
生成型要約のために、最初に文を抽出しそれを書き直す形で要約を生成…
-
## 一言でいうと
要約を作成する際文を抽出してから書き換える形で作成する手法。文選択は単語分散表現をCNNで畳み込みBi-directionalにかけたものをRNNに入れ、ステップごと選択を行う。これをEncoder-Decoderに入れ書き換えを行うが、文選択が微分不可能なため強化学習で最適化を行っている
![image](https://user-images.githubuse…
-
## 一言でいうと
Actor-Criticを使った強化学習で生成型要約を行う
要約文の質を上げるために、Criticの1つとして文がモデルによって生成された文か人間の文かを2値分類するモデルを使うのが特徴
これにより生成される文にOOVや###.#といったのノイズが入りにくくなる
### 論文リンク
https://arxiv.org/abs/1803.11070
###…