NILAS: Soft Lexical Constraints for Language Acquisition with Large Language Models : A new way of lexical constraint bridging AI and Second-Language Acquisition

Sammanfattning

The use of Large Language Models (LLMs) has drastically increased in the last few years, particularly in real-world applications such as the education sector. Although promising, significant concerns of hallucination, fabrication, & poor quality outputs exists, affecting trust & usability among users. One major area is language education, where AI, if consistent, could be used to make users learn efficiently & correctly. This thesis attempted to answer the problem of balancing natural, correct, & adequately constrained lexical generation for efficient language learning, utilizing a well-known method of acquisition called Comprehensible Input. The aim was to explore soft constrained decoding to allow strong signals of non-allowed words through, while encouraging constraints to balance naturalness & those constraints. The method developed utilized a per-word penalty of non-allowed words during decoding and a heterogeneous LLM-as-a-Judge to evaluate and iteratively improve naturalness & correctness on English, Swedish, Spanish, and Korean while being compared to an identical, lexically prompted model, and the state-of-the-art OpenAI ChatGPT 5.2. Results measured across a total of 36,000 samples showed a significant jump in lexical adherence performance in certain tests compared to the prompted version. They performed on par with ChatGPT 5.2 in others. Conclusions drew parallels of potential future improvements and the adoption of alternative methods for soft lexical constraint implementations to improve constraint stability, as well as exploration of larger models to test model limitations in such strict constraints presented in this thesis, and lastly introduce human evaluation to capture naturalness & correctness measures from real language learners.