State-space models are a low-complexity alternative to transformers for encoding long sequences and capturing long-term dependencies. We propose LOCOST: an encoder-decoder architecture based on state-space models for conditional text generation with long context inputs. With a computational complexity of đť’Ş(L log L), this architecture can handle significantly longer sequences than state-of-the-art models that are based on sparse attention patterns. We evaluate our model on a series of long document abstractive summarization tasks. The model reaches a performance level that is 93-96% comparable to the top-performing sparse transformers of the same size while saving up to 50% memory during training and up to 87% during inference. Additionally, LOCOST effectively handles input texts exceeding 600K tokens at inference time, setting new state-of-the-art results on full-book summarization and opening new perspectives for long input processing.
This article focuses on large language models (LLMs) fine-tuning in the scarce data regime (also known as “few-shot learning setting”). We propose a method to increase the generalization capabilities of LLMs based on neural network subspaces. This optimization method, recently introduced in computer vision, aims to improve model generalization by identifying wider local optima through the joint optimization of an entire simplex of models in parameter space. Although this property would be highly beneficial in the context of training large language models in the “few-shot learning” setting, its adaptation to massive, pretrained transformers poses some challenges. First, their considerable number of parameters make it difficult to train several model jointly, and second, their deterministic parameter initialisation schemes make them unfit to the subspace method as originaly proposed. We show in this paper that its application to “Parameter Efficient Fine-Tuning” (PEFT) methods, however, is relatively natural, and we propose to apply it to prefix-tuning, by learning entire simplexes of continous prefixes. We test our method on a variant of the GLUE benchmark adapted to the few-shot learning setting, and show that both our contributions (learning prefix simplexes, and non-deterministic validation metric inference) jointly lead to a gain in average performances compared to state of the art methods.
State-of-the-art NLP models can adopt shallow heuristics that limit their generalization capability (McCoy et al., 2019). Such heuristics include lexical overlap with the training set in Named-Entity Recognition (Taille et al., 2020) and Event or Type heuristics in Relation Extraction (Rosenman et al., 2020). In the more realistic end-to-end RE setting, we can expect yet another heuristic: the mere retention of training relation triples. In this paper we propose two experiments confirming that retention of known facts is a key factor of performance on standard benchmarks. Furthermore, one experiment suggests that a pipeline model able to use intermediate type representations is less prone to over-rely on retention.
Despite efforts to distinguish three different evaluation setups (Bekoulis et al., 2018), numerous end-to-end Relation Extraction (RE) articles present unreliable performance comparison to previous work. In this paper, we first identify several patterns of invalid comparisons in published papers and describe them to avoid their propagation. We then propose a small empirical study to quantify the most common mistake’s impact and evaluate it leads to overestimating the final RE performance by around 5% on ACE05. We also seize this opportunity to study the unexplored ablations of two recent developments: the use of language model pretraining (specifically BERT) and span-level NER. This meta-analysis emphasizes the need for rigor in the report of both the evaluation setting and the dataset statistics. We finally call for unifying the evaluation setting in end-to-end RE.
Unsupervised relation extraction aims at extracting relations between entities in text. Previous unsupervised approaches are either generative or discriminative. In a supervised setting, discriminative approaches, such as deep neural network classifiers, have demonstrated substantial improvement. However, these models are hard to train without supervision, and the currently proposed solutions are unstable. To overcome this limitation, we introduce a skewness loss which encourages the classifier to predict a relation with confidence given a sentence, and a distribution distance loss enforcing that all relations are predicted in average. These losses improve the performance of discriminative based models, and enable us to train deep neural networks satisfactorily, surpassing current state of the art on three different datasets.