Dilbilimi

Permanent URI for this community

https://digitalarchive.library.bogazici.edu.tr/handle/123456789/11623

Browse

Browsing Dilbilimi by Subject "Computational linguistics."

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    De Re De dicto ambiguities in Turkish : disambiguation by LLMs and processing by humans
    (Thesis (M.A.) - Bogazici University. Institute for Graduate Studies in Social Sciences, 2023., 2023) Marşan, Büşra.; Atlamaz, Ümit.; Demirok, Ömer Faruk.
    This study aims to investigate De Re De Dicto ambiguities in Turkish from different points of views: Formal semantics, computational linguistics and psycholinguistics. From the formal semantics perspective, a survey of different accounts that aim to explain compositional semantics of De Re De Dicto ambiguities is provided. With free word order and optionally non-overt accusative marker interacting with information structure, Turkish poses an interesting case for De Re De Dicto ambiguities as their resolution is closely tied to many factors: Discourse and non-discourse elements in the conversational context, information structure, world knowledge, and belief states of the utterer. From the computational linguistics perspective, a challenge data set in Turkish for natural language inference (NLI) is offered to test swiftly growing large language models’ (LLMs) abilities in recognizing syntactic and morphological cues along with the world knowledge. For this purpose, a seed data of 743 sentence triplets were manually created and annotated by native speakers: One potentially De Re De Dicto ambiguous sentence, one De Re paraphrase and one De Dicto paraphrase. Then this seed data was augmented using a natural language augmentation algorithm specifically created for this task. An experiment was conducted to assess probing abilities of CNLI-TR and compare it with NLI-TR, the only other available NLI data set for Turkish. From the psycholinguistics perspective, a self-paced reading experiment was conducted to investigate whether theoretical claims that De Re is "harder" to compositionally derive hold from a sentence processing perspective. Following disambiguating contexts, participants were shown De Re De Dicto ambiguous sentences and were asked comprehension questions targeting specific readings parallel to disambiguating contexts. Reading time of De Re De Dicto ambiguous sentences and response time of comprehension questions were compared across disambiguating contexts to explore whether there was a significant and credible processing difference.