Abstract:
Semantic parsing of geometry problems is the first step towards automated
geometry problem solvers. Existing systems for this task heavily depend on
language-specific NLP tools and use hard-coded parsing rules. Moreover, these
systems produce a static set of facts and record low precision scores. In this study,
we present the two-step memory network, a novel neural network architecture for
deep semantic parsing of GWPs. Our model is language independent and optimized
for low-resource domains. Without using any language-specific NLP tool,
our system performs as good as existing systems. We also introduce on-demand
fact extraction, where a solver can query the model about entities during the
solving stage. This is impossible for existing systems; the set of extracted facts
with these systems are static after the parsing stage. This feature alleviates the
problem of having an imperfect recall.
We also investigate data augmentation techniques for low resource domains
to alleviate the difficulties in applying deep learning techniques in the domain
above. We also introduce an enhanced metric for evaluating language generative
models alleviating the the limitations of exiting metrics. Analysing the results,
we come up with a ranking of models on their suitability to be used o low resource
domains
