Information Technology Reference
In-Depth Information
Choice Generation System, Paraphrase Generation System to Acceptability Ranker, is
implemented.
Fig. 2.
System architecture
3.1
Choice Generation System
The Choice Generation System takes an article as input text and output a set of state-
ments, each of which with a specific testing purpose. The testing purposes that are
considered in this work are: understanding the cohesion of anaphora in the context,
understanding the relationship (cause and effect, comparison, etc.) between details
and identifying factual information that is explicitly stated in the passages. The over-
view of this system is given on the left of Fig. 3. In preprocessing, the information
from the input article is extracted. The CoreNLP pipeline [11] splits the article into
sentences and provides information on coreference chains, part-of-speech tags and
syntactic trees. The PDTB-styled end-to-end discourse parser [10] recognizes intra-
and inter-sentential discourse relations and the corresponding argument spans. Know-
ing the three basic elements (two arguments and the relation between them) allows the
rules to rearrange them into new statements, with predetermined correctness. Since
it's important not to produce vague statements, each pronoun, if not specified in the
sentence, is replaced with the representative mention in the same coreference chain.
Fig. 3.
Overview of the Choice Generation System and the Paraphrase Generation System
These clarified sentences are either sent to choice generation as choice candidates,
which are intended to test the cohesion of anaphora, or enter the sentence simplifica-
tion process. We utilize the sentence simplification work [7] that extracts simplified
Search WWH ::
Custom Search