Evolving Explanatory Novel Patterns for Semantically-Based Text Mining - Natural Language Processing and Text Mining - page 154

AB

IF goal(establish(a))

method(use(b))
method(study(c))

AC

IF goal(establish(a))

method(use(b))
method(study(c))

THEN produce(x')

BC

IF goal(produce(x))

THEN determine(z)

object(perform(y))

THEN determine(z)

Fig. 9.2. Semantically guided Swanson Crossover

b) Default Semantic Crossover: if the previous transitivity does not apply then
the recombination is performed as long as both hypotheses as a whole have
high semantic similarity which is defined in advance by providing minimum
thresholds.

• Mutation: aims to make small random changes on hypotheses to explore new

possibilities in the search space. As in recombination, we have dealt with this
operation in a constrained way, so we propose three kinds of mutations to deal
with the hypotheses' different objects:
a) Role Mutation: one rhetorical role (including its contents: relations and ar-
guments) is selected and randomly replaced by a random one from the initial
role database.
b) Predicate Mutation: one inner predicate and its argument is selected and
randomly replaced with another predicate-argument pair from the initial
predicate databases.
c) Argument Mutation: since we have no information about arguments' seman-
tic types, we choose a new argument by following a guided procedure in
which the former argument is randomly replaced with that having a high
semantic similarity via LSA. [33].

• Population Update: we use a non-generational GA in which some individuals

are replaced by the new offspring in order to preserve the hypotheses' good
material from one generation to other, and so to encourage the improvement of
the population's quality.

Evaluation

Since each hypothesis in our model has to be assessed by different criteria, usual
methods for evaluating fitness are not appropriate. Hence Evolutionary Multi-
Objective Optimisation (EMOO) techniques which use the multiple criteria defined
for the hypotheses are needed. Accordingly, we propose EMOO-based evaluation
metrics to assess the hypotheses' fitness in a domain-independent way and, un-
like other approaches, without using any external source of domain knowledge. The
different metrics are represented by multiple criteria by which the hypotheses are
assessed.
In order to establish evaluation criteria, we have taken into account different
issues concerning plausibility (Is the hypothesis semantically sound?, Are the GA
operations producing something coherent in the current hypothesis?), and quality