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 constraint 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 grammar 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 Grammars.
In order to get a view of how Interaction Grammars work, the following notions are required. 2.1
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 [
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 superpositions 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
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
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
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) [
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 [
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
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.
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
different categories like verb, auxiliary, noun, adjective, adverb, prepositions, determiner,
pronoun etc. Our main effort was on the construction of a grammar for verbs in
affirmative 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 [
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 different grammar trees needed in a complete grammar can be obtainable. 4.1
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 underspecification 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.
The English TSNLP (Test Suite for Natural Language Processing) [
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 [
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
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 complete 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 .