Benjamin Clavié

Benjamin Clavié, Alexandru Ciceu, Frederick Naylor et al.

This case study investigates the task of job classification in a real-world setting, where the goal is to determine whether an English-language job posting is appropriate for a graduate or entry-level position. We explore multiple approaches to text classification, including supervised approaches such as traditional models like Support Vector Machines (SVMs) and state-of-the-art deep learning methods such as DeBERTa. We compare them with Large Language Models (LLMs) used in both few-shot and zero-shot classification settings. To accomplish this task, we employ prompt engineering, a technique that involves designing prompts to guide the LLMs towards the desired output. Specifically, we evaluate the performance of two commercially available state-of-the-art GPT-3.5-based language models, text-davinci-003 and gpt-3.5-turbo. We also conduct a detailed analysis of the impact of different aspects of prompt engineering on the model's performance. Our results show that, with a well-designed prompt, a zero-shot gpt-3.5-turbo classifier outperforms all other models, achieving a 6% increase in Precision@95% Recall compared to the best supervised approach. Furthermore, we observe that the wording of the prompt is a critical factor in eliciting the appropriate "reasoning" in the model, and that seemingly minor aspects of the prompt significantly affect the model's performance.

Benjamin Clavié, Antoine Chaffin, Griffin Adams

Over the last few years, multi-vector retrieval methods, spearheaded by ColBERT, have become an increasingly popular approach to Neural IR. By storing representations at the token level rather than at the document level, these methods have demonstrated very strong retrieval performance, especially in out-of-domain settings. However, the storage and memory requirements necessary to store the large number of associated vectors remain an important drawback, hindering practical adoption. In this paper, we introduce a simple clustering-based token pooling approach to aggressively reduce the number of vectors that need to be stored. This method can reduce the space & memory footprint of ColBERT indexes by 50% with virtually no retrieval performance degradation. This method also allows for further reductions, reducing the vector count by 66%-to-75% , with degradation remaining below 5% on a vast majority of datasets. Importantly, this approach requires no architectural change nor query-time processing, and can be used as a simple drop-in during indexation with any ColBERT-like model.

Benjamin Clavié, Guillaume Soulié

Understanding labour market dynamics requires accurately identifying the skills required for and possessed by the workforce. Automation techniques are increasingly being developed to support this effort. However, automatically extracting skills from job postings is challenging due to the vast number of existing skills. The ESCO (European Skills, Competences, Qualifications and Occupations) framework provides a useful reference, listing over 13,000 individual skills. However, skills extraction remains difficult and accurately matching job posts to the ESCO taxonomy is an open problem. In this work, we propose an end-to-end zero-shot system for skills extraction from job descriptions based on large language models (LLMs). We generate synthetic training data for the entirety of ESCO skills and train a classifier to extract skill mentions from job posts. We also employ a similarity retriever to generate skill candidates which are then re-ranked using a second LLM. Using synthetic data achieves an RP@10 score 10 points higher than previous distant supervision approaches. Adding GPT-4 re-ranking improves RP@10 by over 22 points over previous methods. We also show that Framing the task as mock programming when prompting the LLM can lead to better performance than natural language prompts, especially with weaker LLMs. We demonstrate the potential of integrating large language models at both ends of skills matching pipelines. Our approach requires no human annotations and achieve extremely promising results on skills extraction against ESCO.

Benjamin Clavié, Xianming Li, Antoine Chaffin et al.

Late interaction retrieval methods, pioneered by ColBERT, have emerged as a powerful alternative to single-vector neural IR. By leveraging fine-grained, token-level representations, they have been demonstrated to deliver strong generalisation and robustness, particularly in out-of-domain settings. They have recently been shown to be particularly well-suited for novel use cases, such as reasoning-based or cross-modality retrieval. At the same time, these models pose significant challenges of efficiency, usability, and integrations into fully fledged systems; as well as the natural difficulties encountered while researching novel application domains. Recent years have seen rapid advances across many of these areas, but research efforts remain fragmented across communities and frequently exclude practitioners. The purpose of this workshop is to create an environment where all aspects of late interaction can be discussed, with a focus on early research explorations, real-world outcomes, and negative or puzzling results to be freely shared and discussed. The aim of LIR is to provide a highly-interactive environment for researchers from various backgrounds and practitioners to freely discuss their experience, fostering further collaboration.

Benjamin Clavié, Florian Brand

Recent advancements in Large Vision-Language Models (VLMs), have greatly enhanced their capability to jointly process text and images. However, despite extensive benchmarks evaluating visual comprehension (e.g., diagrams, color schemes, OCR tasks...), there is limited assessment of VLMs' ability to read and reason about text-rich images effectively. To fill this gap, we introduce ReadBench, a multimodal benchmark specifically designed to evaluate the reading comprehension capabilities of VLMs. ReadBench transposes contexts from established text-only benchmarks into images of text while keeping textual prompts and questions intact. Evaluating leading VLMs with ReadBench, we find minimal-but-present performance degradation on short, text-image inputs, while performance sharply declines for longer, multi-page contexts. Our experiments further reveal that text resolution has negligible effects on multimodal performance. These findings highlight needed improvements in VLMs, particularly their reasoning over visually presented extensive textual content, a capability critical for practical applications. ReadBench is available at https://github.com/answerdotai/ReadBench .

Benjamin Clavié, Nathan Cooper, Benjamin Warner

While encoder-only models such as BERT and ModernBERT are ubiquitous in real-world NLP applications, their conventional reliance on task-specific classification heads can limit their applicability compared to decoder-based large language models (LLMs). In this work, we introduce ModernBERT-Large-Instruct, a 0.4B-parameter encoder model that leverages its masked language modelling (MLM) head for generative classification. Our approach employs an intentionally simple training loop and inference mechanism that requires no heavy pre-processing, heavily engineered prompting, or architectural modifications. ModernBERT-Large-Instruct exhibits strong zero-shot performance on both classification and knowledge-based tasks, outperforming similarly sized LLMs on MMLU and achieving 93% of Llama3-1B's MMLU performance with 60% less parameters. We also demonstrate that, when fine-tuned, the generative approach using the MLM head matches or even surpasses traditional classification-head methods across diverse NLU tasks.This capability emerges specifically in models trained on contemporary, diverse data mixes, with models trained on lower volume, less-diverse data yielding considerably weaker performance. Although preliminary, these results demonstrate the potential of using the original generative masked language modelling head over traditional task-specific heads for downstream tasks. Our work suggests that further exploration into this area is warranted, highlighting many avenues for future improvements.

Benjamin Clavié

As language-specific training data tends to be sparsely available compared to English, document retrieval in many languages has been largely relying on multilingual models. In Japanese, the best performing deep-learning based retrieval approaches rely on multilingual dense embedders, with Japanese-only models lagging far behind. However, multilingual models require considerably more compute and data to train and have higher computational and memory requirements while often missing out on culturally-relevant information. In this paper, we introduce JaColBERT, a family of multi-vector retrievers trained on two magnitudes fewer data than their multilingual counterparts while reaching competitive performance. Our strongest model largely outperform all existing monolingual Japanese retrievers on all dataset, as well as the strongest existing multilingual models on all out-of-domain tasks, highlighting the need for specialised models able to handle linguistic specificities. These results are achieved using a model with only 110 million parameters, considerably smaller than all multilingual models, and using only a limited Japanese-language. We believe our results show great promise to support Japanese retrieval-enhanced application pipelines in a wide variety of domains.

Benjamin Clavié, Sean Lee, Rikiya Takehi et al.

Multi-vector dense retrieval methods like ColBERT systematically use a single-layer linear projection to reduce the dimensionality of individual vectors. In this study, we explore the implications of the MaxSim operator on the gradient flows of the training of multi-vector models and show that such a simple linear projection has inherent, if non-critical, limitations in this setting. We then discuss the theoretical improvements that could result from replacing this single-layer projection with well-studied alternative feedforward linear networks (FFN), such as deeper, non-linear FFN blocks, GLU blocks, and skip-connections, could alleviate these limitations. Through the design and systematic evaluation of alternate projection blocks, we show that better-designed final projections positively impact the downstream performance of ColBERT models. We highlight that many projection variants outperform the original linear projections, with the best-performing variants increasing average performance on a range of retrieval benchmarks across domains by over 2 NDCG@10 points. We then conduct further exploration on the individual parameters of these projections block in order to understand what drives this empirical performance, highlighting the particular importance of upscaled intermediate projections and residual connections. As part of these ablation studies, we show that numerous suboptimal projection variants still outperform the traditional single-layer projection across multiple benchmarks, confirming our hypothesis. Finally, we observe that this effect is consistent across random seeds, further confirming that replacing the linear layer of ColBERT models is a robust, drop-in upgrade.

Benjamin Clavié, Marc Alphonsus

We aim to highlight an interesting trend to contribute to the ongoing debate around advances within legal Natural Language Processing. Recently, the focus for most legal text classification tasks has shifted towards large pre-trained deep learning models such as BERT. In this paper, we show that a more traditional approach based on Support Vector Machine classifiers reaches surprisingly competitive performance with BERT-based models on the classification tasks in the LexGLUE benchmark. We also highlight that error reduction obtained by using specialised BERT-based models over baselines is noticeably smaller in the legal domain when compared to general language tasks. We present and discuss three hypotheses as potential explanations for these results to support future discussions.

Benjamin Clavié, Akshita Gheewala, Paul Briton et al.

Large Transformer-based language models such as BERT have led to broad performance improvements on many NLP tasks. Domain-specific variants of these models have demonstrated excellent performance on a variety of specialised tasks. In legal NLP, BERT-based models have led to new state-of-the-art results on multiple tasks. The exploration of these models has demonstrated the importance of capturing the specificity of the legal language and its vocabulary. However, such approaches suffer from high computational costs, leading to a higher ecological impact and lower accessibility. Our findings, focusing on English language legal text, show that lightweight LSTM-based Language Models are able to capture enough information from a small legal text pretraining corpus and achieve excellent performance on short legal text classification tasks. This is achieved with a significantly reduced computational overhead compared to BERT-based models. However, our method also shows degraded performance on a more complex task, multi-label classification of longer documents, highlighting the limitations of this lightweight approach.

Benjamin Clavié, Kobi Gal

The use of large pretrained neural networks to create contextualized word embeddings has drastically improved performance on several natural language processing (NLP) tasks. These computationally expensive models have begun to be applied to domain-specific NLP tasks such as re-hospitalization prediction from clinical notes. This paper demonstrates that using large pretrained models produces excellent results on common learning analytics tasks. Pre-training deep language models using student forum data from a wide array of online courses improves performance beyond the state of the art on three text classification tasks. We also show that a smaller, distilled version of our model produces the best results on two of the three tasks while limiting computational cost. We make both models available to the research community at large.