Referential Choice: Predictability and Its Limits

Area of Science:

• Computational Linguistics
• Cognitive Science
• Natural Language Processing

Background:

• Referential choice, the selection of linguistic devices like pronouns or noun phrases, is crucial for coherent discourse.
• Understanding the factors influencing referential choice aids in modeling human language production and comprehension.

Purpose of the Study:

• To predict referential choice in discourse production using a machine learning approach.
• To investigate the limits of predictability in referential choice and identify factors contributing to ambiguity.

Main Methods:

• Developed a machine learning model incorporating 25 features, including referent properties (animacy, protagonism) and antecedent characteristics (syntactic role, distance).
• Conducted corpus analysis and designed an experimental study with human participants to validate prediction accuracy and explore non-categorical choices.

Main Results:

• The machine learning algorithm achieved nearly 90% prediction accuracy when compared to original referential choices in the corpus.
• Experimental results indicated that prediction accuracy decreases in situations offering multiple plausible referential options, suggesting non-categorical choice points.

Conclusions:

• While machine learning can accurately predict most referential choices, complete predictability is hindered by the inherent flexibility and ambiguity in language.
• The study highlights that divergences from predicted referential devices often occur in contexts where multiple linguistic options are equally valid, reflecting cognitive processing of discourse.