The GELEA2LT Research Seminar aims to provide a forum for presenting and discussing ongoing research by both group members and invited researchers. It seeks to foster deeper engagement with topics related to the group’s research areas while broadening and diversifying the perspectives and approaches adopted.

First Seminar of the Research Group

Date: 2 December 2026

Schedule: 13.30 to 15.00

Location: Sala Jordi Arbonès, Facultat de Traducció i Interpretació

Programme

13:30–14:15 – Li Weijia: LLMs for Teaching and Learning Chinese for Tourism Purposes: The CARES Evaluation Framework and Pilot Study

14:15 – Christian Olalla: The Use of Artificial Intelligence for Annotating Ancient Egyptian Verb Forms

30-minute presentations followed by 15 minutes of discussion.

Moderators: Sergi Álvarez and Mireia Vargas-Urpí

Open to all. Certificates of attendance will be issued upon request.

Presentation Abstracts

LLMs for Teaching and Learning Chinese for Tourism Purposes: The CARES Evaluation Framework and Pilot Study, presented by Li Weijia, PhD candidate in the Translation and Intercultural Studies doctoral programme at the UAB.

This presentation reports on the progress of an ongoing doctoral research project investigating the evaluation of large language models (LLMs) for generating materials and activities for teaching and learning Chinese for tourism purposes. The research aims to determine the extent to which model outputs meet the communicative demands of the tourism sector, learners’ pedagogical needs, and the linguistic and didactic requirements of the tasks assigned to the models. The analysis focuses on the observable characteristics of the generated materials, without yet assessing their effects on actual student learning.
The presentation will outline the objectives of the doctoral research and its methodological approach, with particular emphasis on CARES (linguistic accuracy and appropriateness; pedagogical appropriateness; relevance to specific purposes; feedback effectiveness; and safety, reliability and robustness), an evaluation framework developed specifically for this study. This framework provides a structured approach to analysing model outputs in terms of their suitability for the specialised context, their linguistic and pedagogical quality, and the potential usefulness of the feedback provided. The presentation will explain how CARES is integrated into the evaluation procedure and how it relates to other analytical tools, such as checklists and automated estimates of lexical difficulty.

The main methodological lessons learned from a three-phase pilot study involving 89 runs will also be discussed. The study aimed to explore technical feasibility and examine how different models and configurations respond to increasingly demanding tasks. A distinction will be made between formal compliance with instructions and the appropriateness of the generated activities. The criteria for selecting a model that can be executed and for controlling its configuration will also be addressed. Finally, the findings will be used to inform decisions regarding the main study.

The Use of Artificial Intelligence for Annotating Ancient Egyptian Verb Forms, presented by Christian Olalla, senior lecturer in the Department of Translation, Interpreting and East Asian Studies at the UAB and member of GELEA2LT.

In the grammar of Middle Egyptian, a grammatical form known as sDm=f can perform several functions without any morphological variation in its written form. Thus, sDm=f could be translated as “he listens”, “he listened”, “he was listening”, “he would listen”, “he will listen” or “he was listened to”, with no change in its written form. Determining the function of an sDm=f form depends on both syntactic and grammatical factors (syntactic position, presence or absence of an enunciative auxiliary) and contextual factors.

The complexity of this verb form and the difficulty of determining its functions have led different grammars to analyse it in different ways. For example, Gardiner (1927) distinguishes two types of sDm=f (perfective and imperfective), Hutter (2017) identifies five (perfective, circumstantial imperfective, nominal imperfective, prospective and subjunctive), while Allen (2014) groups all its functions under a single sDm=f form.

This diversity of approaches has direct implications for corpus annotation, that is, assigning a grammatical function to each occurrence of the sDm=f form in a text. Unlike text tagging for lexical disambiguation, where the possible labels, such as lemmas or grammatical categories, are drawn from a shared inventory, this task has two implications: (1) a grammatical theory must be selected in advance to serve as the basis for annotating the forms, and (2) when syntactic and grammatical factors are insufficient, context becomes essential for determining the type of sDm=f form. This complexity raises questions about whether language models can perform this annotation task, how accurately, under what conditions and at what cost.

This presentation outlines the design of a study that aims to address these questions. The starting point is a corpus of sDm=f forms extracted from a literary text, The Eloquent Peasant (pBerlin 3023, pBerlin 3025; pAmherst I, pAmherst II, pButler, rº [pBM 10274], pRamesseum A [pBerlin 10499]; c. 1850 BCE). Two researchers classified all occurrences of the form using an annotation manual that defines the categories and assignment criteria based on the grammars of Gardiner, Allen and Hutter. The same manual is provided via an API to two AI models from the same family but differing in size and cost (Claude Sonnet and Claude Opus). In each round, the models receive the manual accompanied by a different number of examples (0, 10, 30 or 60), selected from occurrences that they will not subsequently be asked to classify. The rounds are independent, allowing the comparison to assess how much guidance each model requires and whether providing more examples can compensate for a model’s lower capacity. The models’ responses are compared with the human annotations and two benchmarks: the result obtained by always selecting the most frequent category and the level of agreement reached between the two researchers. The financial cost and estimated energy consumption of each round are also recorded.