Why do small language models underperform? Studying Language Model Saturation via the Softmax Bottleneck

Recent advances in language modeling consist in pretraining highly parameterized neural networks on extremely large web-mined text corpora. Training and inference with such models can be costly in practice, which incentivizes the use of smaller counterparts. However, it has been observed that smaller models can suffer from saturation, characterized as a drop in performance at some advanced point in training followed by a plateau. In this paper, we find that such saturation can be explained by a mismatch between the hidden dimension of smaller models and the high rank of the target contextual probability distribution. This mismatch affects the performance of the linear prediction head used in such models through the well-known softmax bottleneck phenomenon. We measure the effect of the softmax bottleneck in various settings and find that models based on less than 1000 hidden dimensions tend to adopt degenerate latent representations in late pretraining, which leads to reduced evaluation performance.

On the Scaling Laws of Geographical Representation in Language Models

Language models have long been shown to embed geographical information in their hidden representations. This line of work has recently been revisited by extending this result to Large Language Models (LLMs). In this paper, we propose to fill the gap between well-established and recent literature by observing how geographical knowledge evolves when scaling language models. We show that geographical knowledge is observable even for tiny models, and that it scales consistently as we increase the model size. Notably, we observe that larger language models cannot mitigate the geographical bias that is inherent to the training data.

Anisotropy Is Inherent to Self-Attention in Transformers

The representation degeneration problem is a phenomenon that is widely observed among self-supervised learning methods based on Transformers. In NLP, it takes the form of anisotropy, a singular property of hidden representations which makes them unexpectedly close to each other in terms of angular distance (cosine-similarity). Some recent works tend to show that anisotropy is a consequence of optimizing the cross-entropy loss on long-tailed distributions of tokens. We show in this paper that anisotropy can also be observed empirically in language models with specific objectives that should not suffer directly from the same consequences. We also show that the anisotropy problem extends to Transformers trained on other modalities. Our observations suggest that anisotropy is actually inherent to Transformers-based models.

Headless Language Models: Learning without Predicting with Contrastive Weight Tying

Self-supervised pre-training of language models usually consists in predicting probability distributions over extensive token vocabularies. In this study, we propose an innovative method that shifts away from probability prediction and instead focuses on reconstructing input embeddings in a contrastive fashion via Constrastive Weight Tying (CWT). We apply this approach to pretrain Headless Language Models in both monolingual and multilingual contexts. Our method offers practical advantages, substantially reducing training computational requirements by up to 20 times, while simultaneously enhancing downstream performance and data efficiency. We observe a significant +1.6 GLUE score increase and a notable +2.7 LAMBADA accuracy improvement compared to classical LMs within similar compute budgets.

Is Anisotropy Inherent to Transformers?

The representation degeneration problem is a phenomenon that is widely observed among self-supervised learning methods based on Transformers. In NLP, it takes the form of anisotropy, a singular property of hidden representations which makes them unexpectedly close to each other in terms of angular distance (cosine-similarity). Some recent works tend to show that anisotropy is a consequence of optimizing the cross-entropy loss on long-tailed distributions of tokens. We show in this paper that anisotropy can also be observed empirically in language models with specific objectives that should not suffer directly from the same consequences. We also show that the anisotropy problem extends to Transformers trained on other modalities. Our observations tend to demonstrate that anisotropy might actually be inherent to Transformers-based models.

MANTa: Efficient Gradient-Based Tokenization for Robust End-to-End Language Modeling

Static subword tokenization algorithms have been an essential component of recent works on language modeling. However, their static nature results in important flaws that degrade the models' downstream performance and robustness. In this work, we propose MANTa, a Module for Adaptive Neural TokenizAtion. MANTa is a differentiable tokenizer trained end-to-end with the language model. The resulting system offers a trade-off between the expressiveness of byte-level models and the speed of models trained using subword tokenization. In addition, our tokenizer is highly explainable since it produces an explicit segmentation of sequences into blocks. We evaluate our pre-trained model on several English datasets from different domains as well as on synthetic noise. We find that MANTa improves robustness to character perturbations and out-of-domain data. We then show that MANTa performs comparably to other models on the general-domain GLUE benchmark. Finally, we show that it is considerably faster than strictly byte-level models.

Clustering-based Automatic Construction of Legal Entity Knowledge Base from Contracts

In contract analysis and contract automation, a knowledge base (KB) of legal entities is fundamental for performing tasks such as contract verification, contract generation and contract analytic. However, such a KB does not always exist nor can be produced in a short time. In this paper, we propose a clustering-based approach to automatically generate a reliable knowledge base of legal entities from given contracts without any supplemental references. The proposed method is robust to different types of errors brought by pre-processing such as Optical Character Recognition (OCR) and Named Entity Recognition (NER), as well as editing errors such as typos. We evaluate our method on a dataset that consists of 800 real contracts with various qualities from 15 clients. Compared to the collected ground-truth data, our method is able to recall 84\% of the knowledge.

Multilingual Unsupervised Sentence Simplification

Progress in Sentence Simplification has been hindered by the lack of supervised data, particularly in languages other than English. Previous work has aligned sentences from original and simplified corpora such as English Wikipedia and Simple English Wikipedia, but this limits corpus size, domain, and language. In this work, we propose using unsupervised mining techniques to automatically create training corpora for simplification in multiple languages from raw Common Crawl web data. When coupled with a controllable generation mechanism that can flexibly adjust attributes such as length and lexical complexity, these mined paraphrase corpora can be used to train simplification systems in any language. We further incorporate multilingual unsupervised pretraining methods to create even stronger models and show that by training on mined data rather than supervised corpora, we outperform the previous best results. We evaluate our approach on English, French, and Spanish simplification benchmarks and reach state-of-the-art performance with a totally unsupervised approach. We will release our models and code to mine the data in any language included in Common Crawl.