# 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

• For a path with k relations and k+1 nodes, add m distractor sentences per node.
• Distractor sentence is an explanation of attributes which are not relevant to reasoning such as “Sam likes to play soccer.”

Generalization in CLUTRR

• Explicitly control the number of relations (k) in a path.
• Train on smaller number (k=3) and test on larger numbers (k=4,5).
• Generalizations are expressed as picking up compositional elements of relations from k=3, and then generalize them on k=4,5..
• Training and testing have different distributions.

Setup

• Story/Input sentences: $S=(s_1,s_2,…,s_n)$ where $s_i=(w_1,..,w_m)$
• Subset of words $w_i$ are entities = ${e_1,..,e_n}$ which are anonymized in cloze-style.
• Each sentence $s_i$ describes a relation R.
• Predict relation between a pair of query entities $(e_i,e_j)$