DelphinTutorial_Formalisms
Strictly speaking, not a formalism but primarily a linguistic theory. See brief linguistic characterisation on the DelphinTutorial/Grammars page.
One of the early accomplishments of DELPH-IN partners was agreement on a joint reference formalism (i.e. a ‘designer logic’ for typed feature structures), blending elements of the formalisms proposed in Carpenter (1992), Copestake (1992), and Krieger & Schäfer (1994). The joint descriptive formalism can be informally characterized as a closed-world, conjunctive-only, multiple inheritance type system that enforces strong typing and strict appropriateness, but allows types to be associated with arbitrary (complex) constraints that are inherited and applied both at compilation and at run-time (e.g. when two types unify to a more specific, constraint-introducing subtype). HPSG well-formedness principles, immediate dominance schemata, and constituent ordering constraints are all spelled out in the type hierarchy (and cross-multiplied), yielding a set of phrase structure and lexical rules that can be interpreted as rewrite rules over complex (typed feature structure) categories by a suitable parser or generator.
Compared to the range of formal devices assumed at times in the (theoretical) HPSG literature, the DELPH-IN JRF is relatively conservative, excluding, for example, logical disjunction and negation, implications with complex antecedents, relational constraints, or cyclic feature structures. The DELPH-IN selection of descriptive devices was seeking to balance theoretical and engineering considerations alike, i.e. (i) linguistic adequacy, reflecting two decades of experience in building large computational grammars, and (ii) processing requirements, informed by much earlier work on computational efficiency.
DELPH-IN grammars are encoded in the Type Definition Language (TDL), a textual specification of type inheritance and constraints. The canonical reference for the DELPH-IN Joint Reference Formalism is the appendix of the volume Collaborative Language Engineering:
@incollection{Copestake:02:CLE,
author = {Copestake, Ann},
title = {Definitions of Typed Feature Structures},
booktitle = {Collaborative Language Engineering},
editor = {Oepen, Stephan and Flickinger, Dan and Tsujii, {Jun-ichi} and Uszkoreit, Hans},
year = 2002,
pages = {227--230},
address = {Stanford, CA},
publisher = {CSLI Publications}
}
MRS (Copestake et al 2005) is a system of semantic representations which is designed to support underspecification, notably of quantifier scope. Such underspecified representations can be taken as a description language for other kinds of semantic representations, e.g., predicate logic with generalized quantifiers.
The semantic information associated with an HPSG sign is called an MRS. While typically constructed through unification of typed feature structures (in parsing or generation), MRSs formally are not feature structures (or objects in the DELPH-IN JRF; see above). An MRS consists of:
-
a distinguished label, called the top handle
-
a bag (i.e. unordered multi-set) of elementary predications (also called EPs)
-
and a bag of handle constraints
Each EP has a label (aka handle), a predicate symbol (or pred or relation) and a list of arguments. In many EPs (and in some grammars increasingly in all EPs), one of those arguments can be considered an intrinsic variable (or ‘distinguished’ or ‘characteristic’ variable). Depending on the type of the EP, this can be either an event or instance variable. The value of the other arguments may be either a handle or a variable. Handle-valued arguments are typically either the label of another EP or related via a handle constraint to the label of another EP. Variable-valued arguments are typically intrinsic variables of other EPs within the same MRS.
The naming conventions for predicate symbols allow them to encode lemma, part of speech, and additional coarse sense distinctions as well as distinguishing between surface predicates (typically associated with particular lexical entries, where the lemmatized surface form occurs in the predicate symbol) and abstract predicates (which at times have been called ‘grammar predicates’) introduced by grammatical rules or semantically decomposed lexical entries.
Note that DELPH-IN grammars typically do not make sense distinctions in MRS predicate symbols unless the sense distinction correlates with morphosyntactic differences between lexical entries.
The Semantic Interface (SEM-I) is an optional component in DELPH-IN grammars that spells out the inventory of valid predicates, roles, variable types, variable properties, and such.
While MRSs are the primary objects of meaning representation in DELPH-IN grammars, a number of variants have been defined throughout the years. Robust Minimal Recursion Semantics (RMRS) further decomposes individual EPs (into sets of predications, one per argument in the original EP, linked together through a unique anchor assumed for each EP), to also allow underspecification of argument roles. Elementary Dependency Structures (EDS) reduce the bag of EPs and argument structure to a variable-free dependency graph. EDS can be further simplified into strictly bi-lexical semantic dependencies, dubbed DELPH-IN MRS Dependencies (DM), for compatibility with standard data-driven dependency parsing. To allow interconversion between an original MRS and a variable-free dependency graph, Dependency MRS (DRMS) overlays the EDS graph with additional scope-related information about label (in)equalities among EPs.
While MRS in the first instance is a formalism, the name can also be taken to refer to the collection of analyses that have been developed using this formalism, the vast majority of which are encoded in DELPH-IN grammars, especially the ERG. Documentation of these analyses is under way (see ErgSemantics), and the analyses are often the topic of discussion at DELPH-IN summits and smaller meetings.
As grammars need to make assumptions about tokenization, the DELPH-IN formalisms include a declarative specification language for string normalization and tokenization rules, applied prior to parsing to map from an input string to a sequence of token objects. The REPP formalism essentially defines a cascade of regular expression rewrite rules, with some additional machinery to support iteration, modularity of rule sets and configurability, as well as so-called characterization, i.e. the ability to keep track, for each output token, of character start and end points in the original input string. REPP documentation is available from the ReppTop page.
Another minor element in the DELPH-IN collection of formalisms is what is called Variable Property Mapping (VPM), a declarative, bidirectional facility to map between naming conventions for MRS elements employed grammar-internally (i.e. in the JRF encoding of MRSs built compositionally during parsing or generation) and actual, grammar-independent MRS objectss, as specified in the Semantic Interface (SEM-I). VPM documentation is available from the RmrsVpm page.