Kaili Müürisep

Institute of Computer Science

University of Tartu

kaili@cs.ut.ee

The Constraint Grammar formalism was originally designed by Fred Karlsson [Karlsson 1990] for grammar-based parsing of running text. The most exhaustive constraint grammar is written for English. 95% of all words become after morphological disambiguation unambiguous and correctly analysed. 90% of words get correct and unambiguous syntactic labels [Karlsson et al. 1995].

& 9;There may be distinguished four main modules in the process of parsing: morphological analysis, disambiguation of morphosyntactic ambiguities, determination of sentence-internal clause boundaries and determination of syntactic functions.

& 9;For morphological analysis of Estonian, we use the morphological analyser written by H.-J. Kaalep. The morphological analyser provides each input word with all possible morphosyntactic interpretations (readings). Contextually illegitimate interpretations are discarded by morphological disambiguator [Puolakainen 1996], that also assigns clause-boundary tags to first word of each clause. After the stage of morphological disambiguation the analysed sentence looks like this:

Translation: It happened in the summer before last course (=last year in university).

Word-form is in the first line, readings are in following lines. (Here each word-form has only one reading.) Reading begins with base form that is followed by morphological tags. Root and endings are separated by ‘+’.

The problem of determination of syntactic functions may be broken up into two subtasks: assignment of possible functional tags to each morphological reading and applying of syntactic constraints that discard contextually illegitimate function tags.

& 9;

Syntactic tags that are assigned to each word in the sentence indicate the syntactic functions of words. Constraint Grammar employs surface-near dependency-oriented syntax - no phrase structure will be built. In Estonian Constraint Grammar (ESTCG) [Müürisep 1996] following syntactic tags are used:

Predicator: @FV - predicator; @FV> - negation that belongs to predicate.

Attribute: @A - attribute

Adverbial: @Q - adverbial

All syntactic tags begin with @-symbol.

The modifier tags have a pointer, ‘>‘ or ‘<‘, indicating in which direction the head is to be found. For example, the syntactic tag @<P means that this word is head of nominal phrase that belongs to preposition phrase. In ESTCG I do not use direction indicators in attribute tags, but it seems to be useful in future.

The output of morphological disambiguation is input to morphosyntactic mapping rules that add to each morphological reading appropriate syntactic tags. For instance, a noun in nominative case can be subject, object, predicative, attribute or adverbial. The next example shows, how a sentence looks like after applying morphosyntactic mapping rules.

Each mapping rule states for each needed combination of morphological features which its range of syntactic tags is. The mapping rule may be also constrained by context conditions that set up complementary conditions to the words in neighbourhood.

& 9;The mapping rule

states: if the reading is pronoun in nominative case and somewhere in the right context inside the clause can be found a verb which permits the use of predicative, then add these tags. This rule is applied to the first word of sample sentence.

& 9;There are 213 mapping rules in ESTCG, most of them are context-sensitive. 2.51 tags per word will remain after applying mapping rules.

& 9;After the mapping operation syntactic constraints are applied. If analysis succeeds, all morphological readings have exactly one syntactic tag. It may occur that some readings have more than one syntactic tag. In this case it was impossible to determine proper syntactic function using only context and lexicon information. Constraint Grammar avoids risks more than remaining ambiguities. The following sample sentence is correctly and unambiguosly analysed.

Morphological disambiguation constraints and syntactic constraints are formulated in the same way. Both kinds of constraints consist of operator, target and context-conditions. The target defines, which syntactic function the constraint is about. The operator defines, which operation to perform on the syntactic tags. There are two kinds of operations in syntactic constraints. Syntactic selection constraint has operator ‘=s!’ that selects target tag and removes other(s). Syntactic deletion constraint has operator ‘=s0’ that deletes target tag. The context-conditions are formulated in same way as in morphosyntactic mappings.

& 9;Context-condition consists in the simplest case of two parts: a position and a set. The position refers to the word that will be tested. The set defines which kind of tags the word should have. The position is calculated relatively to the target. Position of target is 0, position of next word to right is 1, position of previous word is -1 e.t.c. If position is preceded by asterisk (‘*’), the position number n refers to some position rightwards of n (if n is positive) or some position leftwards of n (if n is negative). It is also possible to negate the condition. In that case, condition begins with key-word ‘NOT’ and states that the word in given position should not have the tags belonging to given set.

& 9;The syntactic selection constraint

(@w=s! (@<P) (NOT *-1 PP)(NOT *1 PP) **CLB-C)

states: for any syntactically ambiguous word, select syntactic tag @<P (complement of preposition) if there are not other possible complements of adpositions (words with tags @P> or @<P) in the left and right context inside the clause.

& 9;This rule was applied to word kursust (English course) in sample sentence.

The syntactic deletion constraint

((@w = s0 (@CN)& 9;(0 SubstPron) (NOT *-1 SubstPron|Nom)

& 9;(NOT *1 SubstPron|Nom) (NOT *-1 Verb1-2))

states: for any syntactically ambiguous word, delete syntactic tag @CN (predicative in nominative case) if current word is noun or pronoun, if there are not noun or pronouns in nominative case in the left context, if there are not nouns and pronouns in nominative case in the right context, if there are not verbs in the first or second person in the left context. This rule was applied to word see (English it) in sample sentence.

& 9;ESTCG contains 275 syntactic constraints. 1.47 tags per word will remain after applying syntactic rules.

& 9;Table 1 shows how many words become syntactically unambiguous, how many words have 2 syntactic tags, how many 3 tags.

- 68 % of all words become syntactically unambiguous.

- 99% of all words retain the correct syntactic tag.

During composing and testing ESTCG syntactic constraints, I have used test corpus that contains ca 500 sentences, which are previously manually analysed. 22 words (0.32%) lost their correct function during analysis. The errors are mostly caused by ellipsis, some errors occurred during determination of apposition.

The most frequent combinations of tags are shown in table 2.

It is very difficult to distinguish adverbialattributes and adverbials. Approximately every 8th word has both labels. If two nouns are located side by side then it is difficult to determine, which word is attribute, which word is adverbial or they are both independent adverbials.

& 9;Sometimes it is difficult to distinguish object and attribute in genitive case.

Example: See toimub ta (@OG @A) koduses kabinetis.

& 9;Ma leidsin ta (@OG @A) kabinetist.

There is no information about transitivity/intransitivity of verb in the lexicon. The Estonian grammarians say, that it is impossible to write exhaustive list of transitive verbs, because one verb can occur in one sentence as transitive verb but in another sentence as intransitive verb.

& 9;It is complicated to disambiguate subjects and objects for same reason.

Use of attribute and subject tags in parallel is caused by lack of special tag for apposition.

& 9;Disambiguation of subjects and predicatives in nominative case is possible only using heuristic rules - subject precedes predicative.

Finally, I would like to say, that my research had experimental nature. It is necessary to do a lot of work to achieve that Estonian constraint syntax becomes applicable to real text analysis.