Response Generation in Dialogue using a Tailored PCFG Parser
Caixia Yuan Xiaojie Wang Qianhui He
School of Computer Science
Beijing University of Posts and Telecommunications
{yuancx, xjwang}@bupt.edu.cn
alisonchinabupt@gmail.com
Abstract
This paper presents a parsing paradigm for
natural language generation task, which
learns a tailored probabilistic context-free
grammar for encoding meaning represen-
tation (MR) and its corresponding natural
language (NL) expression, then decodes
and yields natural language sentences at
the leaves of the optimal parsing tree for
a target meaning representation. The ma-
jor advantage of our method is that it does
not require any prior knowledge of the M-
R syntax for training. We deployed our
method in response generation for a Chi-
nese spoken dialogue system, obtaining
results comparable to a strong baseline
both in terms of BLEU scores and human
evaluation.
1 Introduction
Grammar based natural language generation (NL-
G) have received considerable attention over the
past decade. Prior work has mainly focused on
hand-crafted generation grammar (Reiter et al.,
2005; Belz, 2008), which is extensive, but also ex-
pensive. Recent work automatically learns a prob-
abilistic regular grammar describing Markov de-
pendency among fields and word strings (Konstas
and Lapata, 2012a, Konstas and Lapata, 2013),
or extracts a tree adjoining grammar provided an
alignment lexicon is available which projects the
input semantic variables up the syntactic tree of
their natural language expression (Gyawali and
Gardent, 2014). Although it is a consensus that at
a rather abstract level natural language generation
can benefit a lot from its counterpart natural lan-
guage understanding (NLU), the problem of lever-
aging NLU resources for NLG still leaves much
room for investigation.
In this paper, we propose a purely data-driven
natural language generation model which exploits
a probabilistic context-free grammar (PCFG) pars-
er to assist natural language generation. The ba-
sic idea underlying our method is that the generat-
ed sentence is licensed by a context-free-grammar,
and thus can be deduced from a parsing tree which
encodes hidden structural associations between
meaning representation and its sentence expres-
sion. A tailored PCFG, i.e., a PCFG easily tailored
to application-specific concepts, is learned from
pairs of structured meaning representation and its
natural language sentence and then used to guide
generation processes for other previously unseen
meaning representations. Table 1 exemplifies a
record from the application under consideration.
Our model is closest to (Konstas and Lapata,
2012a) and (Konstas and Lapata, 2013) who re-
formulate the Markov structure between a mean-
ing representation and a string of text depicted
in (Liang, et al., 2009) into a set of CFG rewrite
rules, and then deduce the best derivation tree for
a database record. Although this Markov structure
can capture a few elements of rudimentary syntax,
it is essentially not linguistic grammars. Thus the
sentences produced by this model are usually un-
grammatically informed (for instance, its 1-BEST
model produces grammatically illegal sentences
like “Milwaukee Phoenix on Saturday on Satur-
day on Saturday on Saturday”). (Konstas and La-
pata, 2013) claims that long range dependency is
an efficient complementary to CFG grammar, and
incorporates syntactic dependency between words
into the reranking procedure to enhance the perfor-
mance. Although conceptually similar, our mod-
el directly learns more grammatical rewrite rules
from hybrid syntactic trees whose nonterminal n-
odes are comprised of phrasal nodes inheriting
from a common syntactic parser, and conceptual
nodes designed for encoding target meaning rep-
resentation. Therefore, the learning aspect of two
models is fundamentally different. We have a sin-
gle CFG grammar that applies throughout, where-