Compositional Language Understanding with Text-based Relational Reasoning

Study of Reasoning

Inductive Logic Programming

Relational Reasoning

Propositional Satisfiability (SAT Solver)

Proposal: CLUTRR

In the dataset, the task is to learn the compositional relations directly from text. If the model learns compositional elements, it should be able to re-use it to solve larger problems.

Dataset construction:

  1. Create a family of relation graphs where nodes are entities and edges are relations.
  2. Choose any two nodes, sample a path between them.
  3. Replace the edges with relations which are chosen randomly from a dictionary of templates.
  4. Predict the relation between the start and the sink/end of the path.

2018-12-07-compositionalQA-dataset.jpg

Add noise as distractors

Generalization in CLUTRR

Setup