Close

Parametric model of agreement

in the light of experimental data

This research is supported by Russian Science Foundation, RSF project 22-18-00037 realized at Lomonosov Moscow State University.

Abstract

The project is aimed at constructing a formal model of agreement in Russian. The problem of agreement modeling has several aspects. The first aspect is the limits of agreement as a model of syntactic interaction. The model should fully reflect the core instances of agreement as long as the relevant phenomena that show similar restrictions. The second aspect deals with the inner structure of agreement model and its parametrization: which components may be involved in the agreement relation and what characteristics they have, which features agreement relation may be established for, etc. Agreement model should be both flexible enough to allow parametric variation of the agreement mechanism based on a set of formal features, and restrictive enough to effectively limit this variation. We believe that the study of the attested variation in agreement can be a convenient research tool for identifying the parametric properties of agreement.
Variable agreement is postulated when there is an ambiguous calculation of agreement features, it occurs in situations of agreement with a non-canonical controller or in the presence of several potential agreement controllers. In addition, the constraints on agreement in various grammatical categories are not uniform. This variability is further affected by additional factors of the external syntax and lexical characteristics of the constituents involved in the agreement relation. The issue of variation puts forward increased demands towards the empirical data on which the model is based: while the modeling of canonical agreement can be done based only on qualitative (binary) differences (“grammatical / ungrammatical”), the modeling of variation that is supposed to be influenced by different factors and their interaction requires quantitative measurements.

Results

The team was working on solving two fundamental problems: creating a parametric model of agreement and integrating agreement mechanisms into theoretic accounts of other syntactic interactions.

 

With regard to the agreement variability parametrization we achieved the following results. First, using the corpus and experimental methods we investigated agreement in context of multiple controllers based on three constructions in Russian. We show that in relative constructions with a personal pronominal head, the conflict of person features between the head and the relative pronoun leads to a decrease in acceptability compared to syncretic agreement or no feature conflict. On the contrary, in binominal clauses syncretic agreement and matching features of potential controllers have the same level of acceptability as agreement with one of the controllers. In constructions with a coordinate subject three agreement patterns are licit: resolved agreement, agreement with the first conjunct (partial agreement) and (in the case of non-past tense) default agreement in 3rd person plural. While these three strategies are all acceptable, they are not in free variation, showing different levels of acceptability depending on word order, conjunct order, and tense form of the verb. Thus, despite all three constructions we studied involve multiple potential controllers, they differ as to feature realization.

 

Secondly, we examined the resolution rules in case of agreement with non-canonical controllers. In constructions with governing quantifiers there are three possible agreement patterns: agreement with the quantifier (subset), agreement with the restrictor (superset), and default agreement. We show that these patterns are found in both Russian and Tatar, but the conditions for their realization are different. Along the same lines, we identified morphological and syntactic factors that determine variance in agreement with a non-canonical controller in noun phrases with coordinated modifiers. These factors include regularity of the number morphology of the noun and interaction of the attributive and predicative agreement patterns.

 

Thirdly, we studied the external conditions for agreement computation. We conducted computational and experimental studies of morphological homonymy of subject and direct object. We examined the ability of neural language models to learn the pattern of predicative agreement in the contexts involving case forms’ homonymy and increasing complexity of the syntactic structure. Using corpus data we investigated parameters native speakers rely on when parsing predicate agreement in transitive structures with varying degrees of homonymy. The results of the study show that for Russian native speakers the case parameter is the most significant one and the parameters of word order and semantics compensate each other, whereas neural network algorithms focus primarily on word order.

 

With regard to the integration of the agreement model into the formal model of Russian syntax we achieved the following results. The study of locality constraints on negative concord revealed that they are close to movement constraints on wh-extraction and relativization; other non-local syntactic interactions such as NPI licensing, anaphor binding and quantifier raising are significantly different. This result corroborates the analysis reducing negative concord to the (covert) movement of a negative pronoun targeting the sentential negation projection.

 

Significant results have been achieved in the domain of modelling syntactic selection as a feature-based interaction. Studying selection of non-finite clauses in Russian, we identified a previously unnoticed type of clausal argument, namely, a small clause with a participle as a predicate and a subject which raises to the structural case position in the matrix clause. These small clauses are selected by the predicates of perception, evaluation, and causation. We developed a hypothesis according to which the constraints on raising structures are feature-based: raising implies that the non-finite predicate should be characterized by the tense feature, which makes it possible in Russian to raise subjects from small clauses, but not from infinitival clauses.

 

On the basis of a distributive analysis of presuppositional predicates, we confirmed the correlation between presuppositionality (as a lexical property of the matrix predicate) and the nominal status of its argument (the presence of the pronominal light head “to”), which was earlier proposed in the literature. Presuppositionality of a predicate determines the acceptability of a nominal projection for verbs with an accusative object and makes it obligatory in nominalizations. We suggested that the violation of this correlation is licensed by a separate pragmatic factor, which implies a correlation between the new information and the absence of a nominal projection and, conversely, the given information and the presence of a nominal projection. Thus, we proposed a model for a clausal argument selection which is based on the default principle of feature-driven syntactic selection isomorphic to the semantic type of the argument, but which also integrates morphosyntactic and informational factors, as well as potentially other non-lexical factors (such as speech register).

 

Finally, we started working on the database of agreement variability in Russian. We compiled a list of criteria to classify the cases of non-standard agreement in Russian. These criteria include the context of variability, the structural type of agreement, the case of the controller, the grammatical categories that show agreement variation. Next, based on three Russian constructions (coordinated subjects, constructions with quantity nouns and constructions with governing quantifiers) we compiled a list of possible predictors of variation resolution. At last, but not at least, we analyzed major theoretical approaches to agreement and identified parameters distinguishing between various formal models. This will make it possible for us to choose the most adequate model for agreement variability parametrization.

Team

Ekaterina Lyutikova

Project Leader

Pavel Grashchenkov

Principal Investigator

Anastasia Gerasimova

Principal Investigator

Mikhail Knyazev

Tatiana Davidjuk

Fedor Baikov

Ksenia Studenikina

Daria Belova