Multiple-choice questions (MCQs) are widely used in the evaluation of large language models (LLMs) due to their simplicity and efficiency. However, there are concerns about whether MCQs can truly measure LLM’s capabilities, particularly in knowledge-intensive scenarios where long-form generation (LFG) answers are required. The misalignment between the task and the evaluation method demands a thoughtful analysis of MCQ’s efficacy, which we undertake in this paper by evaluating nine LLMs on four question-answering (QA) datasets in two languages: Chinese and English. We identify a significant issue: LLMs exhibit an order sensitivity in bilingual MCQs, favoring answers located at specific positions, i.e., the first position. We further quantify the gap between MCQs and long-form generation questions (LFGQs) by comparing their direct outputs, token logits, and embeddings. Our results reveal a relatively low correlation between answers from MCQs and LFGQs for identical questions. Additionally, we propose two methods to quantify the consistency and confidence of LLMs’ output, which can be generalized to other QA evaluation benchmarks. Notably, our analysis challenges the idea that the higher the consistency, the greater the accuracy. We also find MCQs to be less reliable than LFGQs in terms of expected calibration error. Finally, the misalignment between MCQs and LFGQs is not only reflected in the evaluation performance but also in the embedding space. Our code and models can be accessed at https://github.com/Meetyou-AI-Lab/Can-MC-Evaluate-LLMs.

Since ChatGPT released its API for public use, the number of applications built on top of commercial large language models (LLMs) increase exponentially. One popular usage of such models is leveraging its in-context learning ability and generating responses given user queries leveraging knowledge obtained by retrieval augmentation. One problem of deploying commercial retrieval-augmented LLMs is the cost due to the additionally retrieved context that largely increases the input token size of the LLMs. To mitigate this, we propose a token compression scheme that includes two methods: summarization compression and semantic compression. The first method applies a T5-based model that is fine-tuned by datasets generated using self-instruct containing samples with varying lengths and reduce token size by doing summarization. The second method further compresses the token size by removing words with lower impact on the semantic. In order to adequately evaluate the effectiveness of the proposed methods, we propose and utilize a dataset called Food-Recommendation DB (FRDB) focusing on food recommendation for women around pregnancy period or infants. Our summarization compression can reduce 65% of the retrieval token size with further 0.3% improvement on the accuracy; semantic compression provides a more flexible way to trade-off the token size with performance, for which we can reduce the token size by 20% with only 1.6% of accuracy drop.

Despite the recent successes of transformer-based models in terms of effectiveness on a variety of tasks, their decisions often remain opaque to humans. Explanations are particularly important for tasks like offensive language or toxicity detection on social media because a manual appeal process is often in place to dispute automatically flagged content. In this work, we propose a technique to improve the interpretability of these models, based on a simple and powerful assumption: a post is at least as toxic as its most toxic span. We incorporate this assumption into transformer models by scoring a post based on the maximum toxicity of its spans and augmenting the training process to identify correct spans. We find this approach effective and can produce explanations that exceed the quality of those provided by Logistic Regression analysis (often regarded as a highly-interpretable model), according to a human study.

Offensive language detection is an important and challenging task in natural language processing. We present our submissions to the OffensEval 2020 shared task, which includes three English sub-tasks: identifying the presence of offensive language (Sub-task A), identifying the presence of target in offensive language (Sub-task B), and identifying the categories of the target (Sub-task C). Our experiments explore using a domain-tuned contextualized language model (namely, BERT) for this task. We also experiment with different components and configurations (e.g., a multi-view SVM) stacked upon BERT models for specific sub-tasks. Our submissions achieve F1 scores of 91.7% in Sub-task A, 66.5% in Sub-task B, and 63.2% in Sub-task C. We perform an ablation study which reveals that domain tuning considerably improves the classification performance. Furthermore, error analysis shows common misclassification errors made by our model and outlines research directions for future.