We present a web prototype for the Dice framework for joint consolidation of noisy multi-faceted commonsense knowledge (CSK) [1]. In the demonstration session, participants will be familiarized with the multi-faceted knowledge representation formalism used in Dice, and get the opportunity to inspect grounded constraint systems and resulting inferences for two CSK collections, ConceptNet and Quasimodo. The web prototype is available at https://dice.mpi-inf.mpg.de.
Commonsense knowledge (CSK) is a potentially important asset towards
building versatile AI applications, such as visual understanding for describing images
or conversational agents like chatbots. In delineation from encyclopedic
knowledge on entities like Trump, Sydney, or FC Liverpool, CSK refers to properties,
traits and relations of everyday concepts, such as elephants, co ee mugs or school
buses. Prior works on acquiring CSK, such as ConceptNet [
Yet they compute each concept and statement in isolation from others, and the only quantitative measure (or ranking) is a con dence score capturing whether the statement is valid. There is no information about whether a property holds for all or for some of the instances of a concept, and there is no awareness of which properties are typical and which ones are salient from a human perspective.
Dice overcomes these limitations by introducing a multi-faceted model of
CSK statements and methods for joint reasoning over sets of inter-related
statements [
Knowledge representation. While binary truth assignments are sensible for
encyclopedic knowledge (e.g., Trump either was born in NY, or was not), they
fall much short of CSK, which generalizes across individuals and time. To model
re ned semantics of CSK statements, Dice captures four facets of concept
properties: (i) Plausibility indicates whether a statement makes sense at all (like the
established but overloaded notion of con dence scores). (ii) Typicality indicates
whether a property holds for most instances of a concept (e.g., not only for cubs).
(iii) Remarkability expresses that a property stands out by distinguishing the
concept from closely related concepts (like siblings in a taxonomy). (iv) Saliency
re ects that a property is characteristic for the concept, in the sense that most
humans would spontaneously list it in association with the concept. This way,
statements are better contextualized, loosely similar as by use of quali ers in
traditional entity-centric KBs [
Joint reasoning framework. Dice identi es inter-related concepts by their
neighborhoods in the WebIsALOD [
Our constraint system includes logical clauses such as
Plausible(s1; p) ^ Related(s1; s2) ^ :Plausible(s2; p) ^ : : : ) Remarkable(s1; p) where : : : refers to enumerating all siblings of s1, or highly related concepts. The constraint itself is weighted by the degree of relatedness; so it is a soft constraint that does allow exceptions.
For example, our framework can draw the following probabilistic inferences:
Macaques eating bananas makes it likely that also stump-tailed macaques
eat bananas (parent-child propagation)
Penguins not ying is remarkable when most taxonomical siblings do y
(sibling interdependencies)
Being able to swim correlates with being able to dive (statement similarity)
Lions attacking humans being salient implies the event being at least
plausible (dimension interrelation)
Implementation and web interface. The reasoning was performed using the
Gurobi math solver library. The solving of slices of the full constraint system
took several hours on a 500 GB RAM machine, we therefore computed the
results o ine on two existing unidimensional CSK collections, ConceptNet and
Quasimodo. Further details on Dice are in [
We now provide a web interface to explore the underlying constraint systems and resulting statement sets. It is available at https://dice.mpi-inf.mpg.de. (a) Demo landing page.
(b) List of statements for subject band. 1. Inspecting multifaceted KR. At the start, attendees will be familiarized with the multi-faceted KR used in Dice. From the landing page (Fig. 1 (a)), they will be shown statements for arbitrary CSK subjects of their choice (see Fig. 1 (b)), and get the chance to explore scores along the four facets, along with initial scores from ConceptNet and Quasimodo. For example, the statements for the subject band are available at https://dice.mpi-inf.mpg.de/subject/band. This way, attendees will learn about the di erence of unidimensional and multifaceted scoring. 2. Inspecting constraint systems. In the second part of the demonstration, attendees will explore constraint systems for individual statements. From the screen in (Fig. 1 (b)), attendees will already see taxonomic parents and siblings as used for the constraint system. Upon selecting a speci c CSK statement in that screen, a new interface ((Fig. 2) opens, which on the left side (Fig. 2 (a)) shows related statements, numeric statement features, and resulting bootstrapping scores, and on the right side (Fig. 2 (b)) resulting constraint systems. For example, the constraint systems for band: become famous is available at https://dice.mpi-inf.mpg.de/fact/band/become-famous. 3. Chat templates In the third part, attendees will be shown a use case of multi-faceted CSK, dialogue. Provided with some chat templates, which allow to include CSK statements with high or low scores along the four dimensions, attendees will be able to write their own dialogues, and explore the impact of inserting statements of di erent types and scores into these (https://dice. mpi-inf.mpg.de/dialogues). (a) Scores and neighbourhood for statement band: hold concert. (b) Materialized clauses for statement band: hold concert.