Head-Driven Phrase Structure Grammar (HPSG) is a formally well-defined
approach to grammatical theory that seeks to provide a cognitively
plausible model of human languages as systems of declarative
Typed feature structures play a central role in this modeling. The
HPSG community views explicit formalization as a pre-requisite for
theory evaluation, and values efficient large-scale grammar implementations as well as psycholinguistic models of language processing.
Computational HPSG implementations are useful not only in research
(grammar comparison and hypothesis testing), but also in applied domains like machine translation, question answering, language tutoring, etc.
A comprehensive and detailed exposition of the state-of-the-art in HPSG theorizing is provided by the Head-Driven Phrase Structure Grammar: The Handbook volume, published (open-access) by Language Science Press. The proceedings of the annual International conference on HPSG are available here and for a bibliography of HPSG research go to this repository.
Some of the leading ideas of current work in this theory are the following.
- 'Abstract' structures
(e.g. empty categories and abstract functional projections like IP, NegP, SigmaP, etc.) are avoided wherever
possible, in favor of minimal grammatical structures. Similarly, non-monotomic operations like movement are rejected, and instead
modeled like all other phenomena, via features. Different languages are analyzed first and foremost on their own terms, and if any cross-linguist commonalities do exist, they emerge as empirically motivated linguistic generalizations.
HPSG is not predicated on (neurogenetically implausible) linguistic
innateness assumptions, and
the construction-based nature of the theory is consistent with
a wealth of empirical results in language acquisition,
grammar induction, and
probabilistic models of language.
For an overview of the history of phrase structure grammar leading
up to HPSG see Blevins and Sag (2013).
- Linguistic objects are modeled as feature structures organized via a system of types and constraint inheritance, drawing key insights from
research in object-oriented paradigms. The organization of typed feature structures (i.e. the geometry of linguistic descriptions) plays a significant role
in predicting the impossibility of certain kinds of
linguistic phenomena. Although the formalism is very expressive (basically, a kind of first-order logic with sorted types),
the feature geometry of the theory severely restricts - in a linguistically motivated fashion -
what the grammar can express. This principled bottleneck not only makes linguistic predictions, but it also enables efficient
computational implementations of linguistic grammars.
For example, as Kay and Sag (2014)
show, their account
of idioms correctly predicts that some idiomatic dependencies can cross clausal boundaries, but only in a very limited way.
A few other examples of geometric prediction are discussed below.
- In HPSG, the grammar is nothing but a static set of declarative constraints
(stated as a certain sub-set of first-order logic statements), not an algorithm. The notion of derivation does not exist as part of the knowledge of language. Rather, derivations arise from the interaction between the language processing modules, the input, and the grammar constraints. In other words, it is up to the performance modules (i.e. comprehension and production) to consult the competence grammar module in order to determine which structures are licit. Consequently, the grammar is compatible with comprehension and production (see Sag and Wasow (2011, 2015)), and is by definition ‘crash-proof'. Consequently, there is no need for post-derivational filters or computationally
demanding evaluation metrics that compare many alternative derivations in order to guarantee that
only grammatical (convergent) outputs can be derived and interpreted at interface levels.
- All linguistic constraints are stated in terms of
local mother-daughter configurations. For example, category selection, role
assignment, case assignment, head agreement and semantic selection all obey a
particular kind of locality determined by valence selection features. This
is a kind of geometric prediction. For example, unbounded/long-distance dependencies
are modeled not via movement, but rather in terms of certain feature specifications that
are present throughout the 'path' from filler to gap. This feature-based
theory correctly predicts the existence of grammatical phenomena sensitive to
such specifications, as amply documented in languages such as Irish, Chamorro, Icelandic, Kikuyu and Thompson Salish (among many others). This is another instance of
geometric prediction. For more on long-distance dependencies in HPSG
and Levine (2006), and for more on locality in HPSG see Sag
- Agreement phenomena have
been classified by Pollard and Sag (1994) as syntactic concord, anaphoric
agreement, or pragmatic agreement. Their theory of indices predicts,
inter alia, the absence of case agreement in anaphoric agreement. Work
and Zlatic (2003) and various others has
refined these proposals,
extending the range of languages considerably and correcting mispredictions.
- Generalizations about
constraints on the binding of referentially dependent elements are stated in
terms of relative obliqueness (o-command), at the lexical level, rather than configurational
superiority (c-command). See for example
and Sag (1998).
- Within HPSG, words are rich in
information. Lexical information is not simply listed, however; rather it is
organized in terms of multiple inheritance hierarchies and lexical rules that
allow complex properties of words to be derived from the logic of the
lexicon. Current research is developing extensions of hierarchical lexicons
that allow most kinds of lexical rule to be eliminated and linking patterns to be derived
in a general fashion from semantic properties. See for example
and Davis (2003). Word structure and phrase structure are
governed by partly independent principles. Words and phrases are two
kinds (subtypes) of sign. For more recent work on morphology in HPSG
see Crysmann and Bonami (2015) and
and Copestake (2015).
HPSG can model constructions, as well as signs,
in terms of feature structures. This allows constructions to be analyzed
via multiple inheritance hierarchies, which in turn provides a way of
modeling the fact that constructions cluster into groups with
a 'family resemblance' that corresponds to a constraint on a common supertype.
This view is similar to the Berkeley-style conception of
'Construction Grammar' developed by Charles Fillmore, Paul Kay, and colleagues.
More recently, a variant of HPSG called
Construction Grammar (SBCG)
has emerged from collaborative work between Paul Kay, Ivan Sag, and Laura Michaelis, which gives more prominence to constructions and is less reliant
on the notion of syntactic 'head'. The latter allows for an elegant account of non-coordinate
exocentric constructions such as free relatives, certain partitives, and comparative correlatives, among others
(cf. with Sag (1999), Wright
and Kathol (2003), Borsley (2011)).
- HPSG has explored modes of
serialization which are not based on the model of traditional phrase structure
grammar (where sentences are word strings defined derivatively in terms of
phrase structure). HPSG's linearization theory has sought to capture word order phenomena
in various languages which, albeit correlated with notions of hierarchically defined
constituency, are ultimately not derivable from the latter.
This has implications for the treatment of discontinuous
constituency, as the linear syntactic
organization is to some extent dissociated from the combinatorial
relationships among the items serialized. See
for example Kathol
(2004) and Wetta
(2014), for construction-based accounts of word order phenomena in Germanic languages.