We present an effective recipe to train strong long-context LLMs that are capable of utilizing massive context windows of up to 32,000 tokens. Our models are built through continual pretraining from Llama 2 checkpoints with longer text sequences and on a dataset where long texts are upsampled. We perform extensive evaluation using language modeling, synthetic context probing tasks, and a wide range of downstream benchmarks. Across all evaluations, our models achieve consistent improvements on most regular-context tasks and significant improvements on long-context tasks over Llama 2. Moreover, with a cost-effective instruction tuning procedure that is free of expensive annotation, the presented models can already surpass gpt-3.5-turbo-16k‘s overall performance on long-context benchmarks. Alongside these results, we provide an in-depth analysis on each individual component of our method. We delve into Llama’s position encodings and discuss its key limitation in modeling long data. We examine the impact of various design choices in the pretraining process, including the data mix and the training curriculum of sequence lengths – ablation results suggest that having abundant long texts in the pretrain dataset is not the key to achieving strong performance, and we empirically verify that long context continual pretraining is more efficient and similarly effective compared to pretraining from scratch with long sequences.

Entity linking methods based on dense retrieval are widely adopted in large-scale applications for their efficiency, but they can fall short of generative models, as they are sensitive to the structure of the embedding space. To address this issue, this paper introduces DUCK, an approach to infusing structural information in the space of entity representations, using prior knowledge of entity types. Inspired by duck typing in programming languages, we define the type of an entity based on its relations with other entities in a knowledge graph. Then, porting the concept of box embeddings to spherical polar coordinates, we represent relations as boxes on the hypersphere. We optimize the model to place entities inside the boxes corresponding to their relations, thereby clustering together entities of similar type. Our experiments show that our method sets new state-of-the-art results on standard entity-disambiguation benchmarks. It improves the performance of the model by up to 7.9 F1 points, outperforms other type-aware approaches, and matches the results of generative models with 18 times more parameters.

Embedding representations of text are useful for downstream natural language processing tasks. Several universal sentence representation methods have been proposed with a particular focus on self-supervised pre-training approaches to leverage the vast quantities of unlabelled data. However, there are two challenges for generating rich embedding representations for a new document. 1) The latest rich embedding generators are based on very large costly transformer-based architectures. 2) The rich embedding representation of a new document is limited to only the information provided without access to any explicit contextual and temporal information that could potentially further enrich the representation. We propose efficient retrieval-augmented text embeddings (ERATE) that tackles the first issue and offers a method to tackle the second issue. To the best of our knowledge, we are the first to incorporate retrieval to general purpose embeddings as a new paradigm, which we apply to the semantic similarity tasks of SentEval. Despite not reaching state-of-the-art performance, ERATE offers key insights that encourages future work into investigating the potential of retrieval-based embeddings.

Mixture of Experts layers (MoEs) enable efficient scaling of language models through conditional computation. This paper presents a detailed empirical study of how autoregressive MoE language models scale in comparison with dense models in a wide range of settings: in- and out-of-domain language modeling, zero- and few-shot priming, and full-shot fine-tuning. With the exception of fine-tuning, we find MoEs to be substantially more compute efficient. At more modest training budgets, MoEs can match the performance of dense models using ~4 times less compute. This gap narrows at scale, but our largest MoE model (1.1T parameters) consistently outperforms a compute-equivalent dense model (6.7B parameters). Overall, this performance gap varies greatly across tasks and domains, suggesting that MoE and dense models generalize differently in ways that are worthy of future study. We make our code and models publicly available for research use.

Progress in sentence simplification has been hindered by a lack of labeled parallel simplification data, particularly in languages other than English. We introduce MUSS, a Multilingual Unsupervised Sentence Simplification system that does not require labeled simplification data. MUSS uses a novel approach to sentence simplification that trains strong models using sentence-level paraphrase data instead of proper simplification data. These models leverage unsupervised pretraining and controllable generation mechanisms to flexibly adjust attributes such as length and lexical complexity at inference time. We further present a method to mine such paraphrase data in any language from Common Crawl using semantic sentence embeddings, thus removing the need for labeled data. We evaluate our approach on English, French, and Spanish simplification benchmarks and closely match or outperform the previous best supervised results, despite not using any labeled simplification data. We push the state of the art further by incorporating labeled simplification data.

Text simplification aims at making a text easier to read and understand by simplifying grammar and structure while keeping the underlying information identical. It is often considered an all-purpose generic task where the same simplification is suitable for all; however multiple audiences can benefit from simplified text in different ways. We adapt a discrete parametrization mechanism that provides explicit control on simplification systems based on Sequence-to-Sequence models. As a result, users can condition the simplifications returned by a model on attributes such as length, amount of paraphrasing, lexical complexity and syntactic complexity. We also show that carefully chosen values of these attributes allow out-of-the-box Sequence-to-Sequence models to outperform their standard counterparts on simplification benchmarks. Our model, which we call ACCESS (as shorthand for AudienCe-CEntric Sentence Simplification), establishes the state of the art at 41.87 SARI on the WikiLarge test set, a +1.42 improvement over the best previously reported score.

In order to simplify a sentence, human editors perform multiple rewriting transformations: they split it into several shorter sentences, paraphrase words (i.e. replacing complex words or phrases by simpler synonyms), reorder components, and/or delete information deemed unnecessary. Despite these varied range of possible text alterations, current models for automatic sentence simplification are evaluated using datasets that are focused on a single transformation, such as lexical paraphrasing or splitting. This makes it impossible to understand the ability of simplification models in more realistic settings. To alleviate this limitation, this paper introduces ASSET, a new dataset for assessing sentence simplification in English. ASSET is a crowdsourced multi-reference corpus where each simplification was produced by executing several rewriting transformations. Through quantitative and qualitative experiments, we show that simplifications in ASSET are better at capturing characteristics of simplicity when compared to other standard evaluation datasets for the task. Furthermore, we motivate the need for developing better methods for automatic evaluation using ASSET, since we show that current popular metrics may not be suitable when multiple simplification transformations are performed.

Pretrained language models are now ubiquitous in Natural Language Processing. Despite their success, most available models have either been trained on English data or on the concatenation of data in multiple languages. This makes practical use of such models –in all languages except English– very limited. In this paper, we investigate the feasibility of training monolingual Transformer-based language models for other languages, taking French as an example and evaluating our language models on part-of-speech tagging, dependency parsing, named entity recognition and natural language inference tasks. We show that the use of web crawled data is preferable to the use of Wikipedia data. More surprisingly, we show that a relatively small web crawled dataset (4GB) leads to results that are as good as those obtained using larger datasets (130+GB). Our best performing model CamemBERT reaches or improves the state of the art in all four downstream tasks.

Les modèles de langue neuronaux contextuels sont désormais omniprésents en traitement automatique des langues. Jusqu’à récemment, la plupart des modèles disponibles ont été entraînés soit sur des données en anglais, soit sur la concaténation de données dans plusieurs langues. L’utilisation pratique de ces modèles — dans toutes les langues sauf l’anglais — était donc limitée. La sortie récente de plusieurs modèles monolingues fondés sur BERT (Devlin et al., 2019), notamment pour le français, a démontré l’intérêt de ces modèles en améliorant l’état de l’art pour toutes les tâches évaluées. Dans cet article, à partir d’expériences menées sur CamemBERT (Martin et al., 2019), nous montrons que l’utilisation de données à haute variabilité est préférable à des données plus uniformes. De façon plus surprenante, nous montrons que l’utilisation d’un ensemble relativement petit de données issues du web (4Go) donne des résultats aussi bons que ceux obtenus à partir d’ensembles de données plus grands de deux ordres de grandeurs (138Go).

We introduce EASSE, a Python package aiming to facilitate and standardise automatic evaluation and comparison of Sentence Simplification (SS) systems. EASSE provides a single access point to a broad range of evaluation resources: standard automatic metrics for assessing SS outputs (e.g. SARI), word-level accuracy scores for certain simplification transformations, reference-independent quality estimation features (e.g. compression ratio), and standard test data for SS evaluation (e.g. TurkCorpus). Finally, EASSE generates easy-to-visualise reports on the various metrics and features above and on how a particular SS output fares against reference simplifications. Through experiments, we show that these functionalities allow for better comparison and understanding of the performance of SS systems.

In this paper, we present the details of the neural dependency parser and the neural tagger submitted by our team ‘ParisNLP’ to the CoNLL 2018 Shared Task on parsing from raw text to Universal Dependencies. We augment the deep Biaffine (BiAF) parser (Dozat and Manning, 2016) with novel features to perform competitively: we utilize an indomain version of ELMo features (Peters et al., 2018) which provide context-dependent word representations; we utilize disambiguated, embedded, morphosyntactic features from lexicons (Sagot, 2018), which complements the existing feature set. Henceforth, we call our system ‘ELMoLex’. In addition to incorporating character embeddings, ELMoLex benefits from pre-trained word vectors, ELMo and morphosyntactic features (whenever available) to correctly handle rare or unknown words which are prevalent in languages with complex morphology. ELMoLex ranked 11th by Labeled Attachment Score metric (70.64%), Morphology-aware LAS metric (55.74%) and ranked 9th by Bilexical dependency metric (60.70%).