Speech Pauses and Pronominal Anaphors

This paper addresses the usefulness of speech pauses for determining whether third person neuter gender singular pronouns refer to individual or abstract entities in Danish spoken language. The annotations of dyadic map task dialogues and spontaneous first encounters are analyzed and used in machine learning experiments act to automatically identify the anaphoric functions of pronouns and the type of abstract reference. The analysis of the data shows that abstract reference is more often performed by marked (stressed or demonstrative pronouns) than by unmarked personal pronouns in Danish speech as in English, and therefore previous studies of abstract reference in the former language are corrected. The data also show that silent and filled pauses precede significantly more often third person singular neuter gender pronouns when they refer to abstract entities than when they refer to individual entities. Since abstract entities are not the most salient ones and referring to them is cognitively more hard than referring to individual entities, pauses signal this complex processes. This is in line with perception studies, which connect pauses with the expression of abstract or complex concepts. We also found that unmarked pronouns referring to an entity type usually referred to by a marked pronoun are significantly more often preceded by a speech pause than marked pronouns with the same referent type. This indicates that speech pauses can also signal that the referent of a pronoun of a certain type is not the most expected one. Finally, language models were produced from the annotated map task and first encounter dialogues in order to train machine learning experiments to predict the function of third person neuter gender singular pronouns as a first step toward the identification of the anaphoric antecedents. The language models from the map task dialogues were also used for training classifiers to determine the referent type (speech act, event, fact or proposition) of abstract anaphors. In all cases, the best results were obtained by a multilayer perceptron with an F1-score between 0.52 and 0.67 for the three-class function prediction task and of 0.73 for the referential type prediction.