This paper investigates good-enough parsing in Turkish by comparing human self-paced reading performance to the surprisal and attention patterns of three Turkish Large Language Models (LLMs), GPT-2-Base, GPT-2-Large, and LLaMA-3. The results show that Turkish speakers rely on good-enough parsing for implausible but grammatically permissible sentences (e.g., interpreting sentences such as ‘the man bit the dog’ as ‘the dog bit the man’). Although the smaller LLMs (e.g., GPT-2) were better predictors of human RTs, they seem to have relied more heavily on semantic plausibility than humans. Comparably, larger LLMs (e.g., LLaMA-3) tended to make more probabilistic parsing based on word order, exhibiting less good-enough parsing behavior. Therefore, we conclude that LLMs take syntactic and semantic constraints into account when processing thematic roles, but not to the same extent as human parsers.

This study investigates zero-shot and few-shot cross-lingual transfer effects in Part-of-Speech (POS) tagging and Named Entity Recognition (NER) for Hamshentsnag, an endangered Western Armenian dialect. We examine how different source languages, Western Armenian (contact cognate), Eastern Armenian (ancestral cognate), Turkish (substrate or contact-induced), and English (non-cognate), affect the task performance using multilingual BERT and BERTurk. Results show that cognate varieties improved POS tagging by 8% F1, while the substrate source enhanced NER by 15% F1. BERTurk outperformed mBERT on NER but not on POS. We attribute this to task-specific advantages of different source languages. We also used script conversion and phonetic alignment with the target for non-Latin scripts, which alleviated transfer.

Using Quantized Low Rank Adaptation and Parameter Efficient Fine Tuning, we fine-tuned Meta AI’s LLaMA-2-7B large language model as a research assistant in the field of economics for three different types of tasks: title generation, abstract classification, and question and answer. The model was fine-tuned on economics paper abstracts and syntheticically created question-answer dialogues based on the abstracts. For the title generation, the results of the experiment demonstrated that LLaMA-2-Econ (the fine-tuned model) surpassed the base model (7B and 13B) with few shot learning, and comparable models of similar size like Mistral-7B and Bloom-7B in the BLEU and ROUGE metrics. For abstract categorization, LLaMA-2-Econ outperformed different machine and deep learning algorithms in addition to state-of-the-art models like GPT 3.5 and GPT 4 with both single and representative few shot learning. We tested the fine-tuned Q&A model by comparing its output with the base LLaMA-2-7B-chat with a Retrieval Augmented Generation (RAG) pipeline with semantic search and dense vector indexing, and found that LLaMA-2 performed on a par with the base model with RAG.