Automatic Generation and Reranking of SQL-Derived Answers to NL Questions - Trustworthy Eternal Systems via Evolving, Software Data and Knowledge - page 61

Fig. 2. GeoQuery database fragment
Section 2 gives a formal description of the problem while Section 3 describes
the basic steps of our algorithm used to build clause. Section 4 shows how we
prune and weigh queries in their possible combinations to generate an ordered
set of meaningful queries among which we find the answer. Section 5 describes
tree kernels our kernel-based rerankers. Section 6 discusses the results obtained
using a reranking algorithm, while Section 7 draws some conclusions.
2TheProbem
We will begin by introducing the notion of typed dependencies and how to
obtain a collapsed list of dependencies starting from an NL sentence. Then we
will introduce the subset of Structured Query Language that our system can
deal with and, in order to formalize the problem, we will recall the notation of
corresponding operations in relational algebra.
2.1 NL Questions and Dependencies List
To represent the textual relationships of the NL sentence we use typed depen-
dency relations. The Stanford Dependencies representation [8] provides a simple
and consistent description of the binary grammar relations existing between a
governor and a dependent. As shown in the example below, each dependency is
written as abbreviated relation name (governor, dependent). The governor and
the dependent are words in the sentence associated with a number indicating
the position of the word in the sentence.
In particular we refer to collapsed representation, where dependencies involv-
ing prepositions, conjuncts, as well as information about the referent of relative
clauses are collapsed to get direct dependencies between content words.
For example, the Stanford Dependencies Collapsed ( SDC ) representation for
the question, q 1 :“ What are the capitals of the states that border the most popu-
lated state? ” is the following:
SDC q 1 = attr (are-2, what-1) , root (ROOT-0, are-2) ,
det (capitals-4, the-3) ,nsubj (are-2, capitals-4) ,
nsubj (border-9, states-7) ,rcmod (states-7, border-9) ,
det (states-13, the-10) ,advmod (populated-12, most-11) ,
amod (state-13, populated-12) ,dobj (borders-9, state-13)