Information Technology Reference
In-Depth Information
statements from complex sentences using a set of hand-crafted Tregex patterns. The
simplified statements satisfy the testing purpose of identifying explicitly written fact.
The statement generation matches the source, clarified and simplified sentences
with manually-defined rules. The discourse-based rules either recompose pairs of
arguments with wrong discourse relations to form false statements or reorder the
argument pairs and reunite with other discourse connectives in the same relations to
create true statements. If the logical relation stays true, the generated statement is true
and vice versa. In the experiment, the rules are applied only if the discourse relations
that are involved are explicit because there is still room for improvement in the rec-
ognition of implicit discourse relations and because QG is more precision-favored.
The relations we include in the transformation rules are: conjunction, cause, contrast,
concession, condition and comparison. These allow us to provide more variety to the
second testing purposes and to generate choices that test more than one piece of in-
formation. The SST (SuperSense Tag) -based rules transform a sentence by replacing
a noun/verb phrase by another noun/verb phrase with the same SST for their head
words. The generated statements should be plausible but false and should act as
choices that assess learners' ability to identify explicitly written fact. A few sample
rules are listed in Table 1. Rule #1 and #2 are discourse-based while #3 is SST-based.
Fig. 1 (B) is the result of instantiating Rule #1 and Fig. 1 (C) is the statement generat-
ed by applying Rule #3. The head of the noun phrases
Chinese medicine
and
Western
medicine
are categorized to the same SST (
B-noun.cognition
). The two choices have
both been paraphrased whereas Fig.1 (A) has not, which has only undergone simplifi-
cation and leaves the wording largely the same. The full set of rules can be found
in [21].
Table 1.
Sample rules
#
Rule
T/F
1
[Arg1] CONTRAST [Arg2] → [Arg2] CAUSE [Arg1]
False
[Arg1] CONJUNCTION [Arg2] CONTRAST [Arg3]
→ [Arg1] CONCESSION [Arg3]
2
True
3 Sentence={...NP
1
…} & SST
NP1
= SST
NP2
→ Sentence={...NP
2
…}
False
3.2
Paraphrase Generation System
This system generates a ranked list of sentential paraphrases given an input sentence
and a source article. It enables the overall system to produce lexically different state-
ments and to avoid direct usage of text from the input article that would be easily
answerable by word match. The architecture is given on the right of Fig. 3.
Paraphrases are 'sentences or phrases that convey approximately the same meaning
using different words' [4]. Abiding by the definition, the correctness should remain
unchanged for any true or false statement after paraphrasing. Research on paraphras-
ing is mainly divided into two lines, paraphrase extraction and paraphrase generation.
Search WWH ::
Custom Search