Dice: Joint Reasoning for Multi-Faceted Commonsense Knowledge

Dice is a reasoning framework for deriving refined and expressive commonsense knowledge from existing CSK collections.

The interface below allows to explore Dice knowledge computed from two popular CSK collections, ConceptNet and Quasimodo.

Popular subjects

What the interface shows

The interface shows for each subject, the parents and siblings selected for the joint reasoning, along with their weight.

It then lists the statements from the existing CSK collections, along with 5 kinds of scores:

Dice scores can be shown both as absolute scores, or as percentiles. One can also see statements inferred from related concepts (ConceptNet-enriched, Quasimodo-enriched).

On clicking a specific statement, details on how its scores were computed, are shown. In particular, one can see related properties used to reinforce basic scores, prior cues for each statement, and evidence scores, i.e., score before the joint reasoning step. At the bottom, individual clauses from the reasoning framework are shown.

You may also try the dialogue generator, that produces mini dialogues using the commonsense knowledge statements.

The full datasets are available here.


Try to look for bears. The "Related concepts" snippet shows that a bear is a carnivore, and a wild animal. It has black bears and polar bears as siblings. Below we see all its properties, e.g. it attacks persons.

Clicking on Salient sorts the properties according to their salience score. The top-salient property about bears is that they are big, it has a score of 1.00. It is also very plausible and typical, both dimensions being scored 0.99. However, this fact is a little less remarkable, with a score of 0.78. In order to understand better why this fact is less remarkable, let's dig deeper.

Click on the property is big to access the detailed page about the statement. The radar chart presenting the cues shows that the entropic cue of the property is very low (close to 0.1), which already gives lower evidence for remarkability.

Below the charts, are reported the clauses used in the reasoning. It appears that bear siblings, polars bears and giant pandas, are also quite big, which makes it less remarkable. Yet, those clauses have low weights: 0.06 and 0.05, meaning they only have a small impact on the reasoning. Thus, the remarkability score does not suffer that much.


Dice has been developed at the Max Planck Institute for Informatics by Yohan Chalier, Simon Razniewski and Gerhard Weikum.

Further details can be found in this research paper.