Difficulty-controllable Multi-hop Question Generation From Knowledge Graphs

Difficulty-controllable Multi-hop Question Generation From Knowledge Graphs

0.Summary

• automatic  generation of complex multi-hop question over knowledge graph
• Taking a subgraph and an answer as input
• incorporates difficulty estimation based on name entity popularity and makes use of this estimation to generate diffculty-controllable question.

1.Reseach Objective

• automatic  generation of complex multi-hop question over knowledge graph
• KGQA have been hindered by the limitedavailability of data.
• able to generate questions of varying difficulty levels.(困难可控的问题)

2.Background and Problem

problem:

• singel-hop is easy because only find correct  the entities and predicates (候选)are identified(确定)
• complex questions are alsomore difficult to create.most existing complex question datasets are created by either manually or in a semi-automated manner
• input not sequence but most model are based on sequence-to-sequence

example:

3.Method

transfomer

model

1.Knowledge Graph Encoding

Difficulty Level Modelling & Estimation

(1) the confidence of entity linking in the question, 

• named entity recognition

(2) these lectivity of the surface forms of entities in the subgraph

Complex Question Decoder

使用encoder 的两个矩阵 its encoder-decoder attention layer, which helps decoder in attending to(focusing on) important triples in the input subgraph

postiion embedidng

advantage：

• Our model generates complex multi-hop questions
• our method is end-to-end and fully automated without the need of manually created templates or patterns only requires a subgraph and answer.
• Our end-to-end model estimates and controls difficulty levels of generated questions
• employthe Transformer [31] as our base mode treat a graph as a graph, insted of a sequence of triples

4.Evaluation

 

5.Conclusion

Our technique takes a subgraphas input, encodes the answer, estimates the difficulty level, and generates anatural-language question from the subgraph.

conditioned on the difficulty level

We have planned a number of further research directions. Firstly, we willinvestigate a more refined estimation of difficulty levels, taking into accountmore comprehensive information such as predicates and the graph itself, butnot only entities. Secondly, taking into account additional information sourcessuch as background ontologies as entity and predicate definitions is also worthinvestigating.

6.contributions

• We propose a novel model for generating complex, difficulty-controllable questions from subgraphs of multiple triples
• Our Transformer-based model naturally treats a subgraph (a set of triples)as a graph and avoids arbitrary linear is ation into a sequence of triples.
• Our evaluation over a state-of-the-art natural-language generation model ontwo multi-hop QA datasets shows our technique is able to generate questionsof much higher quality.
• Models, dataset and code is available8to facilitate reproduction and furtherresearch on KG QA research.

7. Reference

 

8.Notes