<-- Back to Current Category
English Language Processing
By: Sasson Margaliot, Tue May 23rd, 2006
The research in the field of Natural Language Processing usually
assumes the existence of a syntactic "generative engine" that
combines words and word-like elements into syntactic structures,
and then sometimes displaces them by syntactic movement. A
Linguistic Parser must "undo" all the effects of syntactic
movement, which results in a structure where the relations
between the words are represented more directly.
In recent years, a new theoretical framework was introduced, in
which the syntactic combinatorial system does not stop at the
level of the words. Instead, the same "generative engine"
continues all-the-way-down into morphology. The various parts of
the same word correspond to the different areas of the syntactic
tree, and then are brought together by multiple applications of
movement.
Within this "constructionalist" framework, the syntactic
is not a tree of words - it is a tree made of sub-lexical
elements like roots, prefixes, suffixes, etc. The components of
a single verb are spread all over the parse tree.
Correspondingly, the function of a Linguistic Parser is
different. The goal of the parsing is to "reconstruct" every
such sub-lexical element into its original place in the
syntactic tree.
By undoing the effects of syntactic movement, "constructionalist
parsing" produces the syntactic trees where atomic constituents
of every word are distributed through such a "reconstructed"
syntactic tree. Every constituent is restored into its
appropriate location in the tree, where the context for its
semantic contribution is found in immediately adjacent
locations.
Deep Parsing makes it possible for all the relations between the
elements in the tree to be strictly local. The primary
distinctive property of "constructionalist parsing" is the fact
that the meaning of all the elements of a syntactic tree is
determined locally.
The argument structures of all the verbs are represent by a
small number of "functional elements". These "functional
elements" introduce arguments and determine the semantic roles
of the immediately adjacent Noun Phrases.
The state-of-art algorithms of pattern matching generally fail
to produce good results on natural language texts, before or
after parsing. But when a text undergoes "constructionalist
parsing", the output is more accessible to standard "pattern
matching" methods and analysis, because an appropriate
representation of the text is generated.
About the author:
Sasson Margaliot is the founder of Linguistic Agents Ltd
(www.linguisticagents.com)
