Large language models (LLMs) are rapidly being adopted by users across the globe, who interact with them in a diverse range of languages. At the same time, there are well-documented imbalances in the training data and optimization objectives of this technology, raising doubts as to whether LLMs can accurately represent the cultural diversity of their broad user base. In this study, we look at LLMs and cultural values in particular, and examine how prompt language and cultural framing influence model responses and their alignment with human values in different countries. We do so by probing 10 LLMs with 63 items from the Hofstede Values Survey Module and World Values Survey, translated into 11 languages, and formulated as prompts with and without different explicit cultural perspectives. Our study confirms that both prompt language and cultural perspective produce variation in LLM outputs, but with an important caveat: While targeted prompting can, to a certain extent, steer LLM responses in the direction of the predominant values of the corresponding countries, it does not overcome the models’ systematic bias toward the values associated with a restricted set of countries in our dataset: the Netherlands, Germany, the United States, and Japan. All tested models, regardless of their origin, exhibit remarkably similar patterns: They produce fairly neutral responses on most topics, with selective progressive stances on issues such as social tolerance. Alignment with cultural values of human respondents is improved more with an explicit cultural perspective than with a targeted prompt language. Unexpectedly, combining both approaches is no more effective than cultural framing with an English prompt. These findings reveal that LLMs occupy an uncomfortable middle ground: They are responsive enough to changes in prompts to produce variation, but they are also too firmly anchored to specific cultural defaults to adequately represent cultural diversity.

The most widely used LLMs like GPT4 and Llama 2 are trained on large amounts of data, mostly in English but are still able to deal with non-English languages. This English bias leads to lower performance in other languages, especially low-resource ones. This paper studies the linguistic quality of LLMs in two non-English high-resource languages: Dutch and French, with a focus on the influence of English. We first construct a comparable corpus of text generated by humans versus LLMs (GPT-4, Zephyr, and GEITje) in the news domain. We proceed to annotate linguistic issues in the LLM-generated texts, obtaining high inter-annotator agreement, and analyse these annotated issues. We find a substantial influence of English for all models under all conditions: on average, 16% of all annotations of linguistic errors or peculiarities had a clear link to English. Fine-tuning a LLM to a target language (GEITje is fine-tuned on Dutch) reduces the number of linguistic issues and probably also the influence of English. We further find that using a more elaborate prompt leads to linguistically better results than a concise prompt. Finally, increasing the temperature for one of the models leads to lower linguistic quality but does not alter the influence of English.

This study, submitted to the BUCC2023 shared task on bilingual term alignment in comparable specialised corpora, introduces a supervised, feature-based classification approach. The approach employs both static cross-lingual embeddings and contextual multilingual embeddings, combined with surface-level indicators such as Levenshtein distance and term length, as well as linguistic information. Results exhibit improved performance over previous methodologies, illustrating the merit of integrating diverse features. However, the error analysis also reveals remaining challenges.

This contribution presents D-Terminer: an open access, online demo for monolingual and multilingual automatic term extraction from parallel corpora. The monolingual term extraction is based on a recurrent neural network, with a supervised methodology that relies on pretrained embeddings. Candidate terms can be tagged in their original context and there is no need for a large corpus, as the methodology will work even for single sentences. With the bilingual term extraction from parallel corpora, potentially equivalent candidate term pairs are extracted from translation memories and manual annotation of the results shows that good equivalents are found for most candidate terms. Accompanying the release of the demo is an updated version of the ACTER Annotated Corpora for Term Extraction Research (version 1.5).

The TermEval 2020 shared task provided a platform for researchers to work on automatic term extraction (ATE) with the same dataset: the Annotated Corpora for Term Extraction Research (ACTER). The dataset covers three languages (English, French, and Dutch) and four domains, of which the domain of heart failure was kept as a held-out test set on which final f1-scores were calculated. The aim was to provide a large, transparent, qualitatively annotated, and diverse dataset to the ATE research community, with the goal of promoting comparative research and thus identifying strengths and weaknesses of various state-of-the-art methodologies. The results show a lot of variation between different systems and illustrate how some methodologies reach higher precision or recall, how different systems extract different types of terms, how some are exceptionally good at finding rare terms, or are less impacted by term length. The current contribution offers an overview of the shared task with a comparative evaluation, which complements the individual papers by all participants.

Traditional approaches to automatic term extraction do not rely on machine learning (ML) and select the top n ranked candidate terms or candidate terms above a certain predefined cut-off point, based on a limited number of linguistic and statistical clues. However, supervised ML approaches are gaining interest. Relatively little is known about the impact of these supervised methodologies; evaluations are often limited to precision, and sometimes recall and f1-scores, without information about the nature of the extracted candidate terms. Therefore, the current paper presents a detailed and elaborate analysis and comparison of a traditional, state-of-the-art system (TermoStat) and a new, supervised ML approach (HAMLET), using the results obtained for the same, manually annotated, Dutch corpus about dressage.

We present the highlights of the now finished 4-year SCATE project. It was completed in February 2018 and funded by the Flemish Government IWT-SBO, project No. 130041.1