Events

Abstract

In the 19th century, European linguists discovered that it was possible to reconstruct the pronunciation of earlier languages with no surviving written records in a reproducible fashion by comparing their descendants (the Comparative Method). Starting in the 1960s, multiple attempts have been made to implement the Comparative Method computationally to assist linguists in reconstructing ancestors for the hundreds of language families existing today. However, these efforts have never achieved the accuracy and insight of a first-year student of historical linguistics. In particular, these models have relied on direct or indirect supervision, in contrast to linguistics students, who are able to operate without supervision. David presents a new way to understand the Comparative Method in terms of information theory, and shows how this framework can be used to reduce or eliminate the supervision necessary to reconstruct ancestral pronunciations (protoforms) and changes in pronunciation (sound laws). Finally, he presents results from an agent-based approach through which the entire Comparative Method can be automated without any direct supervision.

Abstract

Construction Grammar (CxG) is one of the major approaches within Cognitive Linguistics. Analyzing linguistic units as so-called constructions (pairs of form and meaning with different levels of schematicity, complexity and idiomaticity) has numerous advantages: the approach allows for the analysis of a wide variety of units (individual words, multi-word units, and even entire sentence structures can all be considered constructions), is highly adaptable with other approaches, assumes a continuum between grammar and lexicon and, accordingly, accesses the area of linguistic utterances that cannot be captured by either lexicon or grammar alone. This talk introduces the key assumptions of Construction Grammar and explains how a group of different constructions form a so-called Constructicon (a lexicon of constructions) — basically a network of constructions. The advantages and challenges of working with linguistic networks are addressed, and it is shown that the representation of a linguistic network varies depending on the parameters assigned to the relations and on the research question. Case studies illustrating the network approach in the domains of Intensification, Conditionality, and Superlatives will show different usage scenarios.

Abstract

Implicit misogyny and classism are difficult for NLP systems because they are often conveyed indirectly, through presuppositions, stereotypes, and unstated assumptions rather than explicit hateful expressions. Arianna presents work on analysing these implicit forms of harm in social media — framing misogyny detection as an argumentative reasoning task, and introducing ACID, a resource for studying how lower-SES people are perceived across cultures.

Abstract

As LLM-generated output becomes increasingly fluent, evaluating what they know and how reliably they generalise has become critical. At the discourse level, meaning is conveyed beyond single sentences, shaped by linguistic conventions, communicative goals, and variation in human interpretation. Janet presents work examining LLMs' grasp of discourse coherence, how breaking down model reasoning into discourse units can account for human label variation, and how incorporating discourse relations can improve faithfulness evaluation in long-form summarisation.