Understanding the multilingual mechanisms of large language models (LLMs) provides insight into how they process different languages, yet this remains challenging. Existing studies often focus on individual neurons, but their polysemantic nature makes it difficult to isolate language-specific units from cross-lingual representations. To address this, we explore sparse autoencoders (SAEs) for their ability to learn monosemantic features that represent concrete and abstract concepts across languages in LLMs. While some of these features are language-independent, the presence of language-specific features remains underexplored. In this work, we introduce SAE-LAPE, a method based on feature activation probability, to identify language-specific features within the feed-forward network. We find that many such features predominantly appear in the middle to final layers of the model and are interpretable. These features influence the model's multilingual performance and language output and can be used for language identification with performance comparable to fastText along with more interpretability. Our code is available at https://github.com/LyzanderAndrylie/language-specific-features

Large Language Models (LLMs) have impressive multilingual capabilities, but they suffer from unexpected code-switching, also known as language mixing, which involves switching to unexpected languages in the model response. This problem leads to poor readability and degrades the usability of model responses. However, existing work on this issue lacks a mechanistic analysis and shows limited effectiveness. In this paper, we first provide an in-depth analysis of unexpected code-switching using sparse autoencoders and find that when LLMs switch to a language, the features of that language exhibit excessive pre-activation values. Based on our findings, we propose $\textbf{S}$parse $\textbf{A}$utoencoder-guided $\textbf{S}$upervised $\textbf{F}$ine$\textbf{t}$uning (SASFT), which teaches LLMs to maintain appropriate pre-activation values of specific language features during training. Experiments on five models across three languages demonstrate that SASFT consistently reduces unexpected code-switching by more than 50\% compared to standard supervised fine-tuning, with complete elimination in four cases. Moreover, SASFT maintains or even improves the models' performance on six multilingual benchmarks, showing its effectiveness in addressing code-switching while preserving multilingual capabilities.

The mechanisms behind multilingual capabilities in Large Language Models (LLMs) have been examined using neuron-based or internal-activation-based methods. However, these methods often face challenges such as superposition and layer-wise activation variance, which limit their reliability. Sparse Autoencoders (SAEs) offer a more nuanced analysis by decomposing the activations of LLMs into a sparse linear combination of SAE features. We introduce a novel metric to assess the monolinguality of features obtained from SAEs, discovering that some features are strongly related to specific languages. Additionally, we show that ablating these SAE features only significantly reduces abilities in one language of LLMs, leaving others almost unaffected. Interestingly, we find some languages have multiple synergistic SAE features, and ablating them together yields greater improvement than ablating individually. Moreover, we leverage these SAE-derived language-specific features to enhance steering vectors, achieving control over the language generated by LLMs. The code is publicly available at https://github.com/Aatrox103/multilingual-llm-features.

Dysarthria, a motor speech disorder, severely impacts voice quality, pronunciation, and prosody, leading to diminished speech intelligibility and reduced quality of life. Accurate assessment is crucial for effective treatment, but traditional perceptual assessments are limited by their subjectivity and resource intensity. To mitigate the limitations, automatic dysarthric speech assessment methods have been proposed to support clinicians on their decision-making. While these methods have shown promising results, most research has focused on monolingual environments. However, multilingual approaches are necessary to address the global burden of dysarthria and ensure equitable access to accurate diagnosis. This thesis proposes a novel multilingual dysarthria severity classification method, by analyzing three languages: English, Korean, and Tamil.

Recent advances in large language model (LLM) pruning have shown state-of-the-art compression results in post-training and retraining-free settings while maintaining high predictive performance. However, such research mainly considers calibrating pruning using English text, despite the multilingual nature of modern LLMs and their frequent uses in non-English languages. In this paper, we set out to explore effective strategies for calibrating the pruning of multilingual language models. We present the first comprehensive empirical study, comparing different calibration languages for pruning multilingual models across diverse tasks, models, and state-of-the-art pruning techniques. Our results present practical suggestions, for example, calibrating in the target language can efficiently yield lower perplexity, but does not necessarily benefit downstream tasks. Our further analysis experiments unveil that calibration in the target language mainly contributes to preserving language-specific features related to fluency and coherence, but might not contribute to capturing language-agnostic features such as language understanding and reasoning. Last, we provide practical recommendations for future practitioners.

Cross-lingual named entity recognition (NER) aims to train an NER model for the target language leveraging only labeled source language data and unlabeled target language data. Prior approaches either perform label projection on translated source language data or employ a source model to assign pseudo labels for target language data and train a target model on these pseudo-labeled data to generalize to the target language. However, these automatic labeling procedures inevitably introduce noisy labels, thus leading to a performance drop. In this paper, we propose a Global-Local Denoising framework (GLoDe) for cross-lingual NER. Specifically, GLoDe introduces a progressive denoising strategy to rectify incorrect pseudo labels by leveraging both global and local distribution information in the semantic space. The refined pseudo-labeled target language data significantly improves the model's generalization ability. Moreover, previous methods only consider improving the model with language-agnostic features, however, we argue that target language-specific features are also important and should never be ignored. To this end, we employ a simple auxiliary task to achieve this goal. Experimental results on two benchmark datasets with six target languages demonstrate that our proposed GLoDe significantly outperforms current state-of-the-art methods.

In the paper, we propose a novel way of improving named entity recognition in the Korean language using its language-specific features. While the field of named entity recognition has been studied extensively in recent years, the mechanism of efficiently recognizing named entities in Korean has hardly been explored. This is because the Korean language has distinct linguistic properties that prevent models from achieving their best performances. Therefore, an annotation scheme for {Korean corpora} by adopting the CoNLL-U format, which decomposes Korean words into morphemes and reduces the ambiguity of named entities in the original segmentation that may contain functional morphemes such as postpositions and particles, is proposed herein. We investigate how the named entity tags are best represented in this morpheme-based scheme and implement an algorithm to convert word-based {and syllable-based Korean corpora} with named entities into the proposed morpheme-based format. Analyses of the results of {statistical and neural} models reveal that the proposed morpheme-based format is feasible, and the {varied} performances of the models under the influence of various additional language-specific features are demonstrated. Extrinsic conditions were also considered to observe the variance of the performances of the proposed models, given different types of data, including the original segmentation and different types of tagging formats.