=Paper=
{{Paper
|id=None
|storemode=property
|title=An Interaction Grammar for English Verbs
|pdfUrl=https://ceur-ws.org/Vol-954/paper17.pdf
|volume=Vol-954
}}
==An Interaction Grammar for English Verbs==
https://ceur-ws.org/Vol-954/paper17.pdf
An Interaction Grammar for English Verbs
Shohreh Tabatabayi Seifi
INRIA / Université de Lorraine, BP 239
54506 Vandoeuvre-les-Nancy cedex
Abstract. This paper accounts for the construction of a grammar for
English verbs using Interaction Grammars. Interaction Grammar is a
grammatical formalism based on the two key notions: polarities and con-
straint system. A polarity expresses an available resource or a lack of
resource and is used to discriminate between saturated and unsaturated
syntactic structures. A grammar is viewed as a system of constraints of
different kinds: structural, feature and polarity constraints, which should
be satisfied by the parse trees of sentences. We have developed a gram-
mar for English verbs in affirmative clauses and finally we evaluated our
grammar on the portion of a test suite of sentences, the English TSNLP,
with LEOPAR parser which is a parser devoted to Interaction Gram-
mars.
Keywords: Grammatical formalism, interaction grammar, tree descrip-
tion, polarity, unification grammar
1 Introduction
The goal of this work is to construct an Interaction Grammar for English verbs. In-
teraction Grammar(IG) [1] is a very powerful formalism which has the advantages of
both Lexical Functional Grammar (LFG) [5] and Categorial Grammar (CG) [3] and
[4]. From LFG, it takes the flexibility with the mechanism of unification to perform
syntactic composition. From CG, it takes the resource sensitivity to control syntactic
composition with polarities. It is inspired by chemical reactions. In this formalism each
lexicalized elementary tree acts as a potentially active element which can participate
in a reaction with another tree. Two nodes of different trees can merge if they are able
to neutralise each other’s active polarities and make stable combination. Whenever all
the polarities are neutralised correctly to build a unique tree, the obtained tree would
be the parse tree of the sentence.
Guillaume and Perrier [9] investigated the principle underlying IG formalism from
more theoretical point of view. In general we can see that in IG polarities are attached
to the features where as in CG they are attached to the constituents themselves. The
other more essential difference lies in the frameworks of these two formalism: CG
are usually formalized in generative deductive framework while IG is formalized in a
model–theoretic framework. Pullum and Scholz highlighted the advantages of changing
the framework [8].
Currently a very detailed French Grammar with quite high coverage has been
developed in IG formalism [9]. This work is an attempt to construct a portion of
English grammar using the same formalism.
R.K. Rendsvig and S. Katrenko (Eds.): ESSLLI 2012 Student Session Proceedings, CEUR Work-
shop Proceedings vol.: see http://ceur-ws.org/, 2012, pp. 160–159.
An Interaction Grammar for English Verbs 161
2 Interaction Grammars
In order to get a view of how Interaction Grammars work, the following notions are
required.
2.1 Tree Description
A tree description is a tree like structure which uses the notion of underspecification
relations to describe a family of trees instead of only one tree [7]. This allows to make
an unlimited number of trees from a unique underspecified tree representation. A tree
which does not have any underspecified relation is a model. For instance, one tree
description with an underspecified dominance link, illustrated in the left side of Fig. 1,
can produce three (and more) different models.
Fig. 1. A description tree with three of its models.
2.2 Feature Structure
IG associates features with the nodes of tree descriptions to put constraints and prevent
building ungrammatical parse trees. Feature structure models the internal structure of
the language where grammatical properties are represented. Not only do feature struc-
tures prevent building ungrammatical sentences, but also they carry valuable linguistics
information. Using feature structure to construct a grammar is a salient characteristics
of a formalism like LFG but unlike LFG there is no recursive feature in IG and feature
structures have only one level.
2.3 Polarities and Saturated Models
Polarities are one of the basic notions in Interaction Grammars. A polarized feature is
used to control the process of merging nodes of trees to be combined. Different kinds
of polarities are used in IG. There are features with positive or negative polarities
meaning an available resource to be consumed or an expected resource to be provided
respectively. A neutral feature indicates a feature which is not to be consumed but
162 An Interaction Grammar for English Verbs
just to participate in a simple unification process while combining with other nodes. A
saturated feature is the result of combination of a positive and a negative feature and
can be unified with no more positive or negative feature.
Finally there are virtual features which should be merged with real features during
the parsing process including positive, negative, neutral and saturated features. Virtual
polarities are used to express required contexts. Positive, negative and virtual features
constitute the active polarities because they need to combine with other polarities.
In Fig. 2 features with positive (–>), negative (<–) and virtual (∼) polarities can be
seen in the tree descriptions associated with the words been and arranged. Saturated
features indicated with symbol <==> and neutral features indicated with symbol = or
== can be seen in the non-completed parse tree in the left side of Fig. 2. The horizontal
arrows between nodes indicate linear precedence constraints and comes in two forms:
large precedence (- - >)which means there can be several nodes between these two
node and immediate precedence (–>) which obstacles locating any other node between
these two nodes.
Tree descriptions along with polarized feature structure are called polarized tree
descriptions or PTDs. The composition process is defined as a sequence of PTD super-
positions controlled by polarities. When two nodes merge, their features will be unified
according to the standard process of unification with the extra limitations coming from
the rules of combination for polarities. Saturated trees are those completely specified
trees which have no active polarities anymore. The goal of parsing is to generate all
saturated trees from the set of input PTDs which come from a particular IG grammar.
2.4 Grammar
An IG grammar is defined by a finite set of PTDs which are called the elementary
PTDs or EPTDs of the grammar. Each EPTD has one (or more) leaf which is linked
with a word of a lexicon and is called an anchored node. This means that the grammar
is strictly lexicalized and there is no EPTD without any anchor. In practice there is
more than one EPTD for each lexical entry in the grammar with respect to its different
syntactic properties.
The parser of an IG grammar has two main roles. First to select the EPTDs of the
words in the input sentence from the set of EPTDs inside the grammar then to build
all valid minimal models with the use of these EPTDs.
Fig. 2 and Fig. 3 show an example of building a model for the sentence The interview
has been arranged. The first figure is a snapshot of the intermediate results during
parsing process; a non-completed parse tree with unsaturated nodes along with the
EPTDs of the rest of the words of the sentence. The second figure shows the result
yielded by combining the EPTDs of the remaining words with the non-completed tree:
it is one valid model, the parse tree of the input sentence.
3 Building grammars with wide coverage
Tackling with real language problems needs a large scale lexicalized grammar which
is hard to build and without using automatic tools it is an overwhelming task. The
main reason for that is the huge degree of grammatical and lexical redundancy. For
instance in IG, several PTDs may share same subtrees which is due to the grammatical
redundancy of syntactic structures. Moreover different lexical entries share the same
elementary PTDs owing to the fact that they are in the same syntactic category. These
An Interaction Grammar for English Verbs 163
Fig. 2. A non-complete parse tree on the top with two of its EPTDs on the bottom
waiting to be used to make a saturated model for the sentence The interview has been
arranged.
164 An Interaction Grammar for English Verbs
Fig. 3. A parse tree (model) for the sentence The interview has been arranged.
An Interaction Grammar for English Verbs 165
redundancies turn even a small modification of the grammar into a big change in large
amount of grammar trees. In order to conquer these obstacles eXtensible MetaGrammar
(XMG) [6] is used as a facilitating tool to write the IG grammar. XMG is a tool for
constructing large scale grammars. The main feature of XMG is to distinguish between
source grammar and object grammar. A source grammar is a collection of statements
written in a human readable language which produces object grammars which are
usable by NLP systems.
3.1 Source grammar and object grammar
The terms source grammar and object grammar here are analogous with the source
and object codes in programming languages. In the current task first we wrote the
source grammar and then we compiled it with XMG into the object grammar which is
encoded in XML. XMG is widely used in construction of grammars in two formalisms:
IG and TAG [2].
In source grammar, each small fragment of tree structure is written in an individual
class.
By use of class disjunction, conjunction and inheritance, more complex classes can
be built. The compilation of terminal classes of the source grammar produces the
EPTDs of the object grammar. Each EPTD does not contain any lexical entry yet, but
has a detailed description of the word which is going to be nested in its anchor node.
3.2 Linking the grammar with the lexicon
Fig. 4. The process tool chain
The characteristics of the appropriate word to be settled in an anchor node of an
EPTD is described in the EPTD interface. An interface is a feature structure describing
the words which are able to anchor the EPTD. Every entry in the lexicon is a pair of a
word and a feature structure hence a word can anchor an EPTD if its feature structure
is unifiable with the EPTD’s interface in the exact same means of unification we have
in node merging.
The whole process chain is illustrated in Fig. 4 and the highlighted box is the part
that has been developed in our work.
166 An Interaction Grammar for English Verbs
4 The Grammar of English Verbs
To build a complete Interaction Grammar for a specific language which can be able to
parse any possible sentence, we need a set of EPTDs for wide range of words from dif-
ferent categories like verb, auxiliary, noun, adjective, adverb, prepositions, determiner,
pronoun etc. Our main effort was on the construction of a grammar for verbs in affir-
mative clauses (e.g. different tenses, moods and voices). We have used a small grammar
for noun phrase, pronouns, prepositions and adjectives to provide appropriate EPTDs
to parse a sentence [11].
The central focus in writing a grammar with IG formalism is to find a way to write
classes in a manner that with the use of heredity, conjunction and disjunction all dif-
ferent grammar trees needed in a complete grammar can be obtainable.
4.1 An Sketch of The Major Modules
Five major modules contribute building the EPTDs for verbs. Module VerbalKernel
is the kernel of all lexical and auxiliary verbs. It implements different possible voices and
moods of the verbs and also it provides syntactic functions of a verb phrase (e.g. head,
subject, modifier and etc). A subclass in a module is an intermediate tree description
that we use along with operators such as heredity, conjunction and disjunction to build
the final tree descriptions.
There are seventeen subclasses in module VerbalKernel. For instance we model
the long distance dependency between a verb and its subject with the use of under-
specification relation. All tree descriptions of the verbs which are coming out of this
module have a subject node which is an empty node for imperative verbs and a non
empty node for all other verbs.
Fig. 5. A schema of relations between some of classes in the module VerbalKernel.
An Interaction Grammar for English Verbs 167
Fig. 5 aims to illustrate different relations between some of classes in the module
VerbalKernel. Some classes are inherited from another one and some are built out of
the disjunction of two or more other classes.
Whenever there is no complement nodes in the verb tree, auxiliaries are following
the same pattern as non auxiliary verbs therefore class Auxiliary is inherited from
VerbalKernel. The EPTD associated with the auxiliary been in Fig. 2 is an example
of an auxiliary EPTD. Eight subclasses are in class auxiliary implementing all different
kinds of auxiliaries. We treated to ,the infinite marker, like auxiliaries in the current
grammar. The main reason for this approach is the complementary distribution between
to and modal auxiliaries e.g. they never appear together but they appear in the exact
same position in the sentence and both force the verb after them to be in the bare
infinitive form.
Module Complements builds all different kinds of verb complements. It contains
20 different classes and all number, category and function of verb complements are
defined in this module.
VerbalKernel and Complements will merge together in order to make a new
module Diathesis with 20 classes which describes different diathesis along with dif-
ferent verb complements e.g. 13 classes for active verbs and 7 classes for passive verbs.
For instance to build an EPTD for an active verb with a direct object a tree describing
a predicate node with one complement of type noun phrase will merge to a tree associ-
ated with active verb and give back the appropriate PTD with active verbs with direct
nominal object. Finally module Verb with 15 classes will implement trees for different
verb families according to their subcategorization frame. For example the family of
transitive verbs can anchor the tree associated with active verbs with direct object
and the tree associated with passive verbs. So the class for transitive verbs will be the
disjunction of two classes of module Diathesis. The PTDs which are coming out of
module Verb and module Auxiliary are elementary PTDs and used to attach to the
lexicon.
5 Evaluation Results
The English TSNLP (Test Suite for Natural Language Processing) [12] project has
aimed to compose a tool for Natural Language Evaluation purposes. It is a database of
phrases which carry different linguistic phenomena to be used as a reference for eval-
uation tasks. These phrases are divided into several categories according to different
linguistic phenomena. Each category has both grammatical and ungrammatical exam-
ples to provide an infrastructure not only for determining the true positives (success
in parsing the grammatical sentences), but also for counting the true negative (failure
to parse ungrammatical sentences).
The English TSNLP has 15 different categories and 3 of those contain phenomena
exclusively related to verbs which was the subject of this research. However those three
categories also contain some other structure which was not included in the current
grammar and sentences of those form were put out as well including phrasal verbs,
sentences having there as their subject and sentences containing relative clauses. (e.g.
That she leave is requested by him.) We have used LEOPAR parser [10] to construct
the models for the input sentences and the result of evaluation of the current grammar
can bee seen in table 1.
168 An Interaction Grammar for English Verbs
Grammatical Ungrammatical
Total Number 148 832
Parsing Success 115 50
Parsing Failure 33 782
Precision 86.7% 94%
Table 1. Evaluation results of the proposed IG grammar on portion of English TSNLP
5.1 Result Analysis
The major reason of failure in parsing grammatical sentences is that the construction
of verbs up to now requires that every verb tree has a subject node which should be
filled with a real subject while parsing the sentence. However there are situations in
which a verb acts like a noun or an adjective and there is no such a subject node
in its grammar tree. (e.g. He finds the office closed.) Owing to the fact that we were
focusing on construction of the verbs of the main VP of the sentence, other grammatical
phenomena which are related to verbs were not properly treated and this failure should
not be regarded as a weakness of the framework. In appropriate time all this structures
can be constructed in the grammar which is one of the goals in the future works.
The other failure is to mistakenly parse some ungrammatical sentences. One of
the main reasons, among other defeats, is that there are some sentences that are not
correct because of semantic issues which is not recognizable by our grammar.
6 Conclusions and Future Works
Interaction grammars is a powerful formalism that has advantages of both unification
grammars and categorial grammars at the same time. Writing a wide coverage grammar
for the English language needs a huge effort and this work can be regarded as the
starting point of such a project. Using tools like XMG to accomplish this goal is quite
helpful and makes the writing and then the tuning of the grammar a lot more easier
than before. Moreover, a high degree of factorization is possible when we separate
source grammar and object grammar which leads to more efficiency.
The potential future aims of this project are first to continue to construct a com-
plete grammar for English incorporating all different phenomena in order to parse any
grammatical English sentence and second try to cope with the still open problems like
coordination in the sentences within the same framework.
On the other hand English TSNLP is a relatively simple set of sentences and is
not quite similar to real corpora. Therefore some steps further would be to enrich the
grammar some how we would be able to cope with real corpora like newspaper or
spoken language corpora where the structure of sentences are more complicated and
our grammar should be able to handle several non quite grammatical sentences too .
References
1. Perrier, G.: Interaction grammars. In 18th International Conference on Computa-
tional Linguistics, CoLing 2000, Sarrebrucken, pages 600–606. (2000)
2. Crabbé, B.: Grammatical development with XMG. LACL 05. (2005)
An Interaction Grammar for English Verbs 169
3. Retoré, C.: The Logic of Categorial Grammars. ESSLLI 2000, Birmingham.(2000)
4. Steedman, M.: Categorial Grammars. A short encyclopedia entry for MIT Encyclo-
pedia of Cognitive Science. (1999)
5. Bresnan, J.: Lexical-Functional Syntax. ESSLLI 2000, Birmingham.(2000)
6. Duchier, D., Le Roux, J., Parmentier,Y.: The Metagrammar Compiler: an NLP Ap-
plication with a Multi-paradigm Architecture. In Second International Mozart/Oz
Conference, MOZ 2004, Charleroi, Belgium, pages 175–187 (2004)
7. Marcus, M.C., Hindle, D., Fleck. M.M. : D-Theory: Talking about Talking about
Trees. In 21st Annual Meeting of the Association for Computational Linguistics,
pages 129–136.(1983)
8. Pullum,G.K., Scholz, B.C. : On the Distinction between Model–Theoretic and
Generative–Enumerative Syntactic Frameworks. In Logical Aspects of Computa-
tional Linguistics, LACL 2001, Le Croisic, France, volume 2099 of Lecture Notes in
Computer Science, pages 17–43. Springer Verlag. (2001)
9. Guillaume,B., Perrier, G.: Interaction Grammars. INRIA Research Report 6621:
http://hal.inria.fr/inria-00288376/ (2008)
10. Guillaume,B., Le Roux,J., Marchand, J., Perrier,G., Fort,K., Planul,J.:A Tool–
chain for Grammarians. CoLING 08, Manchester. (2008)
11. Planul, J.: Construction d’une Grammaire d’Interaction pour l’anglais. Mémoire,
présenté et soutenu publiquement le 27 juin. (2008)
12. Lehmann,S., Oepen, S., Regnier–Prost,S., Netter,K., Lux,V., Klein,J., Falkedal,K.,
Fouvry,F., Estival, D., Dauphin, E., Compagnion, H., Baur, J., Balkan, L.,
Arnold,D.: TSNLP– Test Suite for Natural language Processing. In Proceedings
of COLING 1996, Kopenhagen. (1996)