---

What is WikiReading

@ WIKIREADING- A Novel Large-scale Language Understanding Task over Wikipedia, ACL 2016

new data from structured knowledge statements

Wikidata

(item, property, value) -> (document, question, answer)

Statistics

• all instances size : 18.58M
• train/val/test(85/10/5): 16.03M, 1.89M, 0.95M
• documents: 4.7M unique, 5.31 instances, 4, 879 ; 489.2, 203 words
• properties: 867 unique, 20->75%, 180->99%; Categorical and Relational

Features

• long document
• short question
• structured knowledge

A Real Case

Baselines

State-of-the-art Methods

Coarse-to-fine Model

@ Coarse-to-Fine Question Answering for Long Documents, ACL 2017
@ Hierarchical Question Answering for Long Documents, Arxiv 2016

Intuition

long documents and low speed

Model

Sentence Selection Methods

• BoW Model
  

• Chunked BoW Model

• Convolutional Neural Network Model

Document Representation

• Hard Attention
  

• Soft Attention
  

Learning Methods

• Distant Learning

  First sentence full matching the answer
  First sentence of Document if no full match exists

• Reinforcement Learning
  

• Soft Attention Learning

Result

Discussion

70%, 10%, 20%

Sliding-Window Encoder Attentive Reader (SWEAR)

@ Accurate Supervised and Semi-Supervised Machine Reading for Long Documents, EMNLP 2017

Intuition

Long document, Chunk

Model

• Attention Method
  

Result

Feedback&Advice

• weibo：@伟康青年，@github
• mail：wavejkd@pku.edu.cn