Large language models (LLMs) have emerged as tools for automatic dementia detection from language. This systematic review synthesizes 87 studies published between 1 January 2020 and 15 October 2025 that apply LLMs to spoken or written patient language for detecting Alzheimer’s disease, mild cognitive impairment, and related dementia. We group existing methods into three main families: finetuning, frozen LLMs used as embedding extractors, and prompting-based approaches, and compare them across datasets, languages, and task formulations. The analysis reveals a clear temporal shift from early dominance of finetuning and embedding-based pipelines to a recent surge in prompting methods, alongside heavy reliance on a small set of English speech corpora and limited work on handwriting as a linguistic modality. We also identify recurring methodological limitations, including scarce cross-dataset validation and limited attention to explainability. Taken together, the review outlines current capabilities and gaps, and proposes concrete directions for developing more robust, multimodal, and clinically interpretable LLM-based systems for dementia detection.

The Metaverse requires intelligent QA services for applications like digital twins and avatar assistants, yet assembling high-quality domain-specific data is challenging. We introduce a novel agent-oriented augmentation pipeline using open-source LLMs (LLaMA and DeepSeek) to autonomously generate and refine synthetic QA pairs. Agents leverage chain-of-thought prompts and feedback loops to iteratively validate and improve responses. By augmenting a limited Stack Overflow R-tag dataset (2,000 examples) with 4,500 synthetic items, our method boosts BERTScore by 11.2% and F1 by 8.0% over static zero-shot baselines. This transparent, cost-effective workflow lays groundwork for scalable QA augmentation in virtual environments.

Intent recognition is a fundamental component in task-oriented dialogue systems (TODS). Determining user intents and detecting whether an intent is Out-of-Scope (OOS) is crucial for TODS to provide reliable responses. However, traditional TODS require large amount of annotated data. In this work we propose a hybrid approach to combine BERT and LLMs in zero and few-shot scenarios to recognize intents and detect OOS utterances. Our approach leverages LLMs generalization power and BERT’s computational efficiency in such scenarios. We evaluate our method on multi-party conversation corpora and observe that sharing information from BERT outputs to LLMs lead to system performance improvement.

This paper explores the development of a first Neural Machine Translation for a low-resource language pair, Kabyle- French. Contrary to the idea that more sentences would enhance the translator's quality, our study suggests that using Back-translation on low-quality monolingual corpora can harm the translation performance. We deepen an analysis of the translator's performance to identify the optimal operational range, allowing us to use the Back-translation effectively and thus reduce the error rate during the generation of the parallel corpus. Our results highlight the central role played by the quality of the monolingual corpus in determining the effectiveness of Back-translation, highlighting the benefits of prioritizing data quality over quantity. Furthermore, we introduce a first method to evaluate the effectiveness of Back-translation, taking into account sentence length and part-of-speech tags. This evaluation demonstrates an improved performance when applying Back-translation to sentences of medium length.

While the fine-tuning process of extensive contextual language models often demands substantial computational capacity, utilizing generic pre-trained models in highly specialized domains can yield suboptimal results. This paper aims to explore an innovative approach to derive pertinent word embeddings tailored to a specific domain with limited computational resources (The introduced methodologies are tested within the domain of hepatic surgery, utilizing the French language.). This exploration takes place within a context where computational limitations prohibit the fine-tuning of large language models. A new embedding (referred to as FTW2V) that combines Word2Vec and FastText is introduced. This approach addresses the challenge of incorporating terms absent from Word2Vec’s vocabulary. Furthermore, a novel method is used to evaluate the significance of word embeddings within a specialized corpus. This evaluation involves comparing classification scores distributions of classifiers (Gradient Boosting) trained on word embeddings derived from benchmarked Natural Language Processing (NLP) models. As per this assessment technique, the FTW2V model, trained from scratch with limited computational resources, outperforms generic contextual models in terms of word embeddings quality. Additionally, a computationally efficient contextual model rooted in FTW2V is introduced. This modified model substitutes Gradient Boosting with a transformer and integrates Part Of Speech labels.

Integrating a source of structured prior knowledge, such as a knowledge graph, into transformer-based language models is an increasingly popular method for increasing data efficiency and adapting them to a target domain. However, most methods for integrating structured knowledge into language models require additional training in order to adapt the model to the non-textual modality. This process typically leads to some amount of catastrophic forgetting on the general domain. KnowBert is one such knowledge integration method which can incorporate information from a variety of knowledge graphs to enhance the capabilities of transformer-based language models such as BERT. We conduct a qualitative analysis of the results of KnowBert-UMLS, a biomedically specialized KnowBert model, on a variety of linguistic tasks. Our results reveal that its increased understanding of biomedical concepts comes at the cost, specifically, of general common-sense knowledge and understanding of casual speech.

Domain adaptation in pretrained language models usually comes at some cost, most notably out-of-domain performance. This type of specialization typically relies on pre-training over a large in-domain corpus, which has the side effect of causing catastrophic forgetting on general text. We seek to specialize a language model by incorporating information from a knowledge base into its contextualized representations, thus reducing its reliance on specialized text. We achieve this by following the KnowBert method, applied to the UMLS biomedical knowledge base. We evaluate our model on in-domain and out-of-domain tasks, comparing against BERT and other specialized models. We find that our performance on biomedical tasks is competitive with the state-of-the-art with virtually no loss of generality. Our results demonstrate the applicability of this knowledge integration technique to the biomedical domain as well as its shortcomings. The reduced risk of catastrophic forgetting displayed by this approach to domain adaptation broadens the scope of applicability of specialized language models.

Abstract This work focuses on the rapid development of linguistic annotation tools for low-resource languages (languages that have no labeled training data). We experiment with several cross-lingual annotation projection methods using recurrent neural networks (RNN) models. The distinctive feature of our approach is that our multilingual word representation requires only a parallel corpus between source and target languages. More precisely, our approach has the following characteristics: (a) it does not use word alignment information, (b) it does not assume any knowledge about target languages (one requirement is that the two languages (source and target) are not too syntactically divergent), which makes it applicable to a wide range of low-resource languages, (c) it provides authentic multilingual taggers (one tagger forNlanguages). We investigate both uni and bidirectional RNN models and propose a method to include external information (for instance, low-level information from part-of-speech tags) in the RNN to train higher level taggers (for instance, Super Sense taggers). We demonstrate the validity and genericity of our model by using parallel corpora (obtained by manual or automatic translation). Our experiments are conducted to induce cross-lingual part-of-speech and Super Sense taggers. We also use our approach in a weakly supervised context, and it shows an excellent potential for very low-resource settings (less than 1k training utterances).

Arabic is one of the Semitic languages written in Arabic script in its standard form. However, the recent rise of social media and new technologies has contributed considerably to the emergence of a new form of Arabic, namely Arabic written in Latin scripts, often called Romanized Arabic or Arabizi. While Romanized Arabic is an informal language, Berber or Tamazight uses Latin script in its standard form with some orthography differences depending on the country it is used in. Both these languages are under-resourced and unknown to the state-of-the-art language identiüers. In this paper, we present a language automatic identifier for both Romanized Arabic and Romanized Berber. We also describe the built linguistic resources (large dataset and lexicons) including a wide range of Arabic dialects (Algerian, Egyptian, Gulf, Iraqi, Levantine, Moroccan and Tunisian dialects) as well as the most popular Berber varieties (Kabyle, Tashelhit, Tarifit, Tachawit and Tamzabit). We use the Prediction by Partial Matching (PPM) and dictionary-based methods. The methods reach a macro-average F-Measure of 98.74% and 97.60% respectively.

This work focuses on the development of linguistic analysis tools for resource-poor languages. We use a parallel corpus to produce a multilingual word representation based only on sentence level alignment. This representation is combined with the annotated source side (resource-rich language) of the parallel corpus to train text analysis tools for resource-poor languages. Our approach is based on Recurrent Neural Networks (RNN) and has the following advantages: (a) it does not use word alignment information, (b) it does not assume any knowledge about foreign languages, which makes it applicable to a wide range of resource-poor languages, (c) it provides truly multilingual taggers. In a previous study, we proposed a method based on Simple RNN to automatically induce a Part-Of-Speech (POS) tagger. In this paper, we propose an improvement of our neural model. We investigate the Bidirectional RNN and the inclusion of external information (for instance low level information from Part-Of-Speech tags) in the RNN to train a more complex tagger (for instance, a multilingual super sense tagger). We demonstrate the validity and genericity of our method by using parallel corpora (obtained by manual or automatic translation). Our experiments are conducted to induce cross-lingual POS and super sense taggers.

Bilingual lexicons are central components of machine translation and cross-lingual information retrieval systems.Their manual construction requires strong expertise in both languages involved and is a costly process.Several automatic methods were proposed as an alternative but they often rely on resources available in a limited number of languages and their performances are still far behind the quality of manual translations.We introduce a novel approach to the creation of specific domain bilingual lexicon that relies on Wikipedia.This massively multilingual encyclopedia makes it possible to create lexicons for a large number of language pairs.Wikipedia is used to extract domains in each language, to link domains between languages and to create generic translation dictionaries.The approach is tested on four specialized domains and is compared to three state of the art approaches using two language pairs: French-English and Romanian-English.The newly introduced method compares favorably to existing methods in all configurations tested.

In the scope of Financial Watch project, several targeted events have been required by contacted users in banking and investment domains. Financial news are classified with respect of the list of desired events. In this paper, a conceptual approach for indexing short English news in the financial domain is presented. By using a supervised original learning approach, a categorization method is proposed. Experimentation of the method on a sample of news showed that in almost all cases, occurrences of the right events in each document have been recognized, with respect to a corpus on the financial domain.

Translation memories play a vital role in Computer-Aided Translation (CAT). Approaches to align sentences from parallel corpora are generally used to construct translation memories. In this paper, we present, on the one hand, a hybrid approach to align sentences combining different information sources (bilingual lexicon, sentence length and sentence position, semantic similarity), and on the other hand, a Computer-Aided Translation prototype based on a cross-language information retrieval approach. This approach consists in building a database of source and target sentences augmented with grammatical information of words and considering the sentence to translate as a "query" to that database. The Computer-Aided Translation prototype has been experimented on a set of examples and the experimental results we obtained are encouraging and outperform the current CAT tools.