Using a Probabilistic Model of Discourse Relations to Investigate Word Order Variation

Using a probabilistic model of discourse relations to investigate word order variation Cassandre Creswell Cymfony, Inc. 600 Essjay Rd Williamsville NY 14221 USA ccreswell@cymfony.com Department of Linguistics University of Pennsylvania Philadelphia PA 19104 USA creswell@ling.upenn.edu Abstract Like speakers of any natural language, speakers of English potentially have many different word orders in which to encode a single meaning. One key fac- tor in speakers’ use of certain non-canonical word orders in English is their ability to contribute infor- mation about syntactic and semantic discourse rela- tions. Explicit annotation of discourse relations is a difficult and subjective task. In order to measure the correlations between different word orders and vari- ous discourse relations, this project utilizes a model in which discourse relations are approximated us- ing a set of lower-level linguistic features, which are more easily and reliably annotated than discourse relations themselves. The featural model provides statistical evidence for the claim that speakers use non-canonicals to communicate information about discourse structure. 1 Introduction: Non-canonical main clause word order in English Users of natural languages have many ways to en- code the same propositional content within a sin- gle clause. In English, besides the “canonical” word order, (1), options for realizing a proposi- tion like GROW(MYRA,EGGPLANTS), include top- icalization, left-dislocation, it-clefts, and wh-clefts, shown in (2–5), respectively. (1) Myra grows eggplants. (2) Eggplants, Myra grows. (3) Eggplants, Myra grows them. (4) It’s eggplants that Myra grows. (5) What Myra grows are eggplants. Corpus-based research has shown that these forms are appropriate only under certain discourse conditions (Prince, 1978; Birner and Ward, 1998); among others. These include the membership of referents in a salient set of entities (left-dislocations and topicalizations) or the salience of particular propositions (topicalizations and clefts). For exam- ple, in (6), the topicalization is felicitous because there is a salient set KINDS OF VEGETABLES and a salient open proposition, that Myra stands in some relation X with an element of that set. (6) Myra likes most vegetables, but eggplants she adores. V = {KINDS OF VEGETABLES}; P = X(m1, v2), SUCH THAT v2 ∈ V The discourse conditions licensing the use of these non-canonical syntactic forms are necessary conditions. When they do not hold, native speak- ers judge the use of the form infelicitous. They are not, however, sufficient conditions for use because salient sets and open propositions are ubiquitous in any discourse context, but these non-canonical forms are rare. Each type alone makes up < 1% of utterances, across a variety of genres (Creswell, 2003). In addition to their information structure func- tions, one additional communicative goal these word orders fulfill is providing information about how an utterance is related to other discourse seg- ments (Creswell, 2003). Native speaker intuitions about the appropriateness of non-canonicals in par- ticular contexts provide anecdotal evidence (i.e. based on listing individual examples) for this dis- course function. To provide broader support for this claim, however, we need to be able to generalize across many tokens. Ideally, a corpus annotated with discourse rela- …

Content truncated to 500 words