DELFT
Complex Factoid Question Answering with a Free-Text Knowledge Graph
DELFT Abstract
DELFT is a factoid question answering system which combines the nuance and depth of knowledge graph question an- swering approaches with the broader coverage of free-text.
DELFT builds a free-text knowledge graph from Wikipedia, with entities as nodes and sentences in which entities co-occur as edges. For each question, DELFT finds the subgraph linking question entity nodes to candidates using text sentences as edges, creating a dense and high coverage semantic graph. A novel graph neural network reasons over the free-text graph—combining evidence on the nodes via information along edge sentences—to select a final answer.
Experiments on three question answering datasets show delft can answer entity-rich questions better than machine reading based models, bert-based answer ranking and memory networks.
DELFT Example
Wikipedia Graph Format
We provide two versions of Wikipedia graph, using orginal links here and using tagme entities here.
Each data file contains multiple json objects, and each object contains the following items:
Page id: the Wikipedia id of the page
Title: the Wikipedia title of the page
Text: Sentence
Anchored et : the tagged entities from Tagme or linked entities
Each tagme entity has the following: (Wikipedia title, entity start index, entity end index, tagme score, Wikipedia page id, surface form)
Each linked entity has the following: (Wikipedia title, surface form)
Code
The DELFT code is coming soon, available at https://github.com/henryzhao5852/DELFT.
Bibtex for WWW'20 paper:
@inproceedings{Zhao:Xiong:Qian:Boyd-Graber-2020,
Title = {Complex Factoid Question Answering with a Free-Text Knowledge Graph},
Author = {Chen Zhao and Chenyan Xiong and Xin Qian and Jordan Boyd-Graber},
Booktitle = {The Web Conference},
Year = {2020}
}