Evaluating Interfaced LLM Bias

Evaluating Interfaced LLM Bias Kai-Ching Yeh National Taiwan University ykcmia@gmail.com Jou-An Chi National Taiwan University R11142005@ntu.edu.tw Da-Chen Lian National Taiwan University D08944091@ntu.edu.tw Shu-Kai Hsieh National Taiwan University shukaihsieh@ntu.edu.tw Abstract In this research, we comprehensively ana- lyze the potential biases inherent in Large Language Model, utilizing meticulously cu- rated input data to ascertain the extent to which such data sway machine-generated responses to yield prejudiced outcomes. Notwithstanding recent strides in mitigat- ing bias in LLM-based NLP, our findings underscore the continued susceptibility of these models to data-driven bias. We have integrated the PTT NTU board as our primary data source for this investigation. Moreover, our study elucidates that, in certain contexts, machines may manifest biases without supplementary prompts. However, they can be guided toward ren- dering impartial responses when provided with enhanced contextual nuances. Keywords: Bias, Natural Language Pro- cessing, LangChain 1 Introduction The evolution of Large Language Models (LLMs) has brought to the fore a series of ethi- cal concerns, one of the most pressing being im- plicit bias in these models (Zhou et al., 2023). Such biases can be attributed to the machine learning algorithms utilized in language mod- eling and the datasets chosen for training and fine-tuning. Training datasets culled from the Internet might predominantly mirror the char- acteristics of the most substantial user demo- graphic, which can be predominantly young and English-speaking (Bender et al., 2021). Moreover, the fine-tuning datasets reliant on manual annotation may inherently possess bi- ases stemming from the perspectives of the an- notators (Zhou et al., 2023). Consequently, LLMs can inadvertently perpetuate the biases in their training or fine-tuning datasets. Evaluating biases within LLMs is both cru- cial and an emergent domain of academic in- quiry. Some scholarly pursuits have centered on the capacity of the language model to dis- cern bias during data interpretation (Huang and Xiong, 2023; Parrish et al., 2021), whereas others have delved into the manifestation of bias when the language model engenders text for subsequent tasks (Dhingra et al., 2023; Huang et al., 2023). The challenge of biases engendered by LLMs has garnered substantial scholarly interest, cat- alyzing the advent of diverse methodologies to counteract these biases. Reinforcement learning with human feedback (RLHF), for in- stance, is a training paradigm that steers the model using human feedback (Bai et al., 2022; Christiano et al., 2017). This method aims to reconcile the model’s outputs with human norms and anticipations, curtailing detrimen- tal outcomes. Throughout its training phase, the model recalibrates its parameters in re- sponse to human feedback, thereby attenu- ating inherent biases. Beyond RLHF, there have been concerted efforts to rectify biases in fine-tuned instructional models via an ar- ray of strategies, one notable approach being the utilization of prompting techniques like the "chain-of-thought" (CoT) (Wei et al., 2022; Dige et al., 2023; Huang and Xiong, 2023; Gan- guli et al., 2023). Although a plethora of research has been dedicated to identifying bias in LLMs and for- mulating debiasing techniques, there remains an under-examined threat capable of directly …

Content truncated to 500 words