Game Development Reference
In-Depth Information
Fig. 4.4
Co-occurrence
distribution of the LSG
network term relationships
Fig. 4.5
Co-occurrence
distributions of nonsense
(semantically not sound) term
pairs, done for three word sets
50
45
40
35
30
25
20
15
10
5
0
Co-occurrence rate of within term pairs (noisy pairs)
Corpus 50.000
Corpus 5.000
Corpus 500
terms may be raised to concepts, the associations may be labeled with relationship
types (i.e. predicates). We were interested, what types of relationships are present in
term network created by LSG (e.g. are there some dominant types?). Such findings
may trigger future game modifications or recommend the LSG term network as seed
for other semantics acquisition approaches (out of scope of this work).
We have conducted experiments, in which we acquired types for LSG network
term relationships. We did so in two ways:
1. We confronted the LSG term network with existing common knowledge base
(we chose the ConceptNet ontology
1
): we took each LSG relationship's terms
and queried the knowledge base, whether it contains the relationship and of what
type it is. The ConceptNet defines 23 possible predicates, one of them being
a general “related to” predicate. The number of predicates was suitable for our
experiments, because it gave more detailed insight on the relationships, but at the
same time, did not cause too much sparsity within the data.
2. Using the same set of predicates, we assigned the relationship types manually,
using two judges for evaluation. As a side effect, we also re-examined the term
network validity (over larger dataset, than in previous experiments).
Search WWH ::
Custom Search