-
-
Hi, for the 2(3)-hop-ins(sub)-triples datasets, could you give a detail explanation about the 2-hop path or 3-hop path. Based on the raw file, I'm very confused how to get a multi-hop inference datase…
-
## 一言でいうと
不完全な知識グラフ上での関係推定を、強化学習で行う手法。グラフが不完全でありグラフ上の関係=正しい保証がなため、別途事前学習した知識グラフの分散表現で到達できなかった場合でも蓋然性に応じ報酬を与える。また、冗長な関係パスを防ぐため行動を正則化(Dropout)するようにしている。
![image](https://user-images.githubusercont…
-
ScienceQA was evaluated in your experiments. As I understand, ScienceQA is a benchmark associated with multi-modal tasks, whereas your model operates purely within the realm of text. Could you please …
-
## 論文リンク
https://arxiv.org/abs/1906.07132
## 公開日(yyyy/mm/dd)
2019/07
## 概要
A -> B -> C というような multi-hop な論理的推論を必要とするデータセットで HotpotQA というものがあるが、そのデータにはそれなりの割合で、その段階的推論をスキップする shortcut があることを発見…
-
# 一言でいうと
知識グラフ上のマルチホップ推論において、エンティティとクエリを埋め込んだ空間上で推論する。埋め込みにベータ分布を仮定することで、conjunction, disjunction, negetaionを自然に導入できるようにした。
![image](https://user-images.githubusercontent.com/1615546/97443721-b3e2…
-
**Question:**
Where is the prompt generating the actual question. I am using Claude Sonnet for generating the question test data set and it seems like the prompt can use some tweaking. The issue …
-
-
The link prediction results based on embedding models in the paper are so high, why is the Hit@10 result of the TransE model only 0.47?
-
https://openreview.net/forum?id=B1lf43A5Y7
ICLR2019 submission withdrawn.
Clear Rejectionのレビュアーがいた。