Certain words and structures in language signal to the reader or listener that the statement following them is true. Consider the difference between these two headlines: “Scientists realized that smoking harms your body” versus “Scientists suggested that smoking harms your body.” The first headline presents “smoking harms your body” as an established fact, whereas the second frames it as a conjecture yet to be proven. While distinguishing between such nuances is effortless for humans, computers can struggle with these linguistic subtleties. In a series of four studies, Salimifar and colleagues conducted linguistic and computational experiments to examine these lexical and structural elements, known as presupposition triggers, in greater detail.
Recognizing presupposition triggers is crucial for making inferences, and improving computational models’ ability to identify them can enhance their performance in inferential tasks. However, the relationship between these triggers and their inferences can be more complex than the previous example suggests. Consider what happens when the earlier statement is negated: “Scientists did not realize that smoking harms your body.” Here, the assertion about scientists' realization no longer holds, but what about the claim that “smoking harms your body”? Does it remain a fact? This challenge, known as the projection problem, highlights the intricate nature of presuppositions in language.
This dissertation explores how certain presupposition triggers behave in different linguistic environments and examines how humans and computers differ in their ability to recognize presuppositional inferences. The findings, derived from linguistic and computational experiments as well as a dataset of human annotations, reveal that, just as humans often disagree on these inferences, computers also struggle to determine the correct interpretation.
Saeedeh Salimifar succesfully defended her thesis on 28 February 2025.
Trial lecture
Designated topic: "Natural language contents: Taxonomies and diagnostics (where contents include entailments, conversational implicatures, different types of projective content, etc.)"
Evaluation committee
- Professor, dr. Judith Tonhauser, Universität Stuttgart (first opponent)
- Senior Lecturer, dr. Todor Koev, Ben Gurion University (second opponent)
- Professor, dr. J Pritty Patel-Grosz, University of Oslo (committee administrator)
Chair of the defence
- Professor Patrick Georg Grosz
Supervisors
- Professor Dag T.T. Haug, Universitetet i Oslo
- Jacopo Romoli, Heinrich-Heine-Universität Düsseldorf