EVALUATING THE SEMANTIC AND SYNTACTIC EFFECTS OF STANDARDIZED TEXTUAL LLM PROMPTS

Sammanfattning

With the recent emergence of Large language models (LLMs) such as ChatGPT, a new technical environment has been introduced to an abundance of new users. As LLMs are known for being dependant on proper input to produce desired responses, different approaches have been proposed to thwart the impact of poorly structured input. By investigating the semantic and syntactic effects of applying known textual processing techniques such as lemmatization, stop word removal and tokenization, this thesis aims to find out whether such standardization could be one of these approaches. To this end, a grading system of four aspects drawing inspiration from human communication was created and used on input prompts of varying shots. The result suggests that standardization, while not always detrimental, potentially has a negative effect on the semantic and syntactic integrity of output. However, it underscores the possibility of utilizing tailored standardization processes to achieve a certain quality without incurring the negative effects