Rim Messaoudi, Achraf Louiza Rim and Francois Azelart, Akkodis Research, France

Text-based comments play a crucial role in providing feedback for various industries. However, effectively filtering and categorizing this feedback based on custom context-specific criteria requires sophisticated language modeling techniques. While traditional approaches have shown effectiveness, they often require a substantial amount of data to compensate for their modeling deficiencies. In this work, we focus on highlighting the performance and limitations of prompt-free few-shot text classification using open-source pre-trained sentence transformers. On the one hand, our research includes a comprehensive study across different benchmark datasets, encompassing 9 dimensions such as sentiment analysis, topic modeling, grammatical acceptance, and emotion classification. Also, we worked at making different experiences to test Prompt-Free Few-Shot Text Classification. On the other hand, we underline prompt-free few-shot classification limitations when the targeted criteria are complex. As an alternative approach, prompting an instruction-fine-tuned language model has demonstrated favorable outcomes, as proven by our application in the specific use case of "œIdentifying and extracting resolution results and actions from explanatory notes", achieving an accuracy rate of 80%.

Language models, Sentence transformers, SetFit, contrastive learning, distillation, intelligence compression, NLP, semantic similarity