Standardized benchmarks are central to evaluating and comparing model performance in Natural Language Processing (NLP). However, Large Language Models (LLMs) have exposed shortcomings in existing benchmarks, and so far there is no clear solution. In this paper, we survey a wide scope of benchmarking issues, and provide an overview of solutions as they are suggested in the literature. We observe that these solutions often tackle a limited number of issues, neglecting other facets. Therefore, we propose concrete checklists to cover all aspects of benchmarking issues, both for benchmark creation and usage. We illustrate the use of our checklists by applying them to three popular NLP benchmarks (i.e., SuperGLUE, WinoGrande, and ARC-AGI). Additionally, we discuss the potential advantages of adding minimal-sized test-suites to benchmarking, which would ensure downstream applicability on real-world use cases.

In this paper, we present our results of the classification of Corporate Social Responsibility (CSR) Themes and Topics shared task, which encompasses cross-lingual multi-class classification and monolingual multi-label classification. We examine the performance of multiple machine learning (ML) models, ranging from classical models to pre-trained large language models (LLMs), and assess the effectiveness of Data Augmentation (DA), Data Translation (DT), and Contrastive Learning (CL). We find that state-of-the-art generative LLMs in a zero-shot setup still fall behind on more complex classification tasks compared to fine-tuning local models with enhanced datasets and additional training objectives. Our work provides a wide array of comparisons and highlights the relevance of utilizing smaller language models for more complex classification tasks.

This study introduces a new method for distance-based unsupervised topical text classification using contextual embeddings. The method applies and tailors sentence embeddings for distance-based topical text classification. This is achieved by leveraging the semantic similarity between topic labels and text content, and reinforcing the relationship between them in a shared semantic space. The proposed method outperforms a wide range of existing sentence embeddings on average by 35%. Presenting an alternative to the commonly used transformer-based zero-shot general-purpose classifiers for multiclass text classification, the method demonstrates significant advantages in terms of computational efficiency and flexibility, while maintaining comparable or improved classification results.