Explore Better Relative Position Embeddings from Encoding Perspective for Transformer Models

Relative position embedding (RPE) is a successful method to explicitly and efficaciously encode position information into Transformer models. In this paper, we investigate the potential problems in Shaw-RPE and XL-RPE, which are the most representative and prevalent RPEs, and propose two novel RPEs called Low-level Fine-grained High-level Coarse-grained (LFHC) RPE and Gaussian Cumulative Distribution Function (GCDF) RPE. LFHC-RPE is an improvement of Shaw-RPE, which enhances the perception ability at medium and long relative positions. GCDF-RPE utilizes the excellent properties of the Gaussian function to amend the prior encoding mechanism in XL-RPE. Experimental results on nine authoritative datasets demonstrate the effectiveness of our methods empirically. Furthermore, GCDF-RPE achieves the best overall performance among five different RPEs.

Transformer-based Lexically Constrained Headline Generation

This paper explores a variant of automatic headline generation methods, where a generated headline is required to include a given phrase such as a company or a product name. Previous methods using Transformer-based models generate a headline including a given phrase by providing the encoder with additional information corresponding to the given phrase. However, these methods cannot always include the phrase in the generated headline. Inspired by previous RNN-based methods generating token sequences in backward and forward directions from the given phrase, we propose a simple Transformer-based method that guarantees to include the given phrase in the high-quality generated headline. We also consider a new headline generation strategy that takes advantage of the controllable generation order of Transformer. Our experiments with the Japanese News Corpus demonstrate that our methods, which are guaranteed to include the phrase in the generated headline, achieve ROUGE scores comparable to previous Transformer-based methods. We also show that our generation strategy performs better than previous strategies.

DiscoDVT: Generating Long Text with Discourse-Aware Discrete Variational Transformer

Despite the recent advances in applying pre-trained language models to generate high-quality texts, generating long passages that maintain long-range coherence is yet challenging for these models. In this paper, we propose DiscoDVT, a discourse-aware discrete variational Transformer to tackle the incoherence issue. DiscoDVT learns a discrete variable sequence that summarizes the global structure of the text and then applies it to guide the generation process at each decoding step. To further embed discourse-aware information into the discrete latent representations, we introduce an auxiliary objective to model the discourse relations within the text. We conduct extensive experiments on two open story generation datasets and demonstrate that the latent codes learn meaningful correspondence to the discourse structures that guide the model to generate long texts with better long-range coherence.

Do Transformer Modifications Transfer Across Implementations and Applications?

The research community has proposed copious modifications to the Transformer architecture since it was introduced over three years ago, relatively few of which have seen widespread adoption. In this paper, we comprehensively evaluate many of these modifications in a shared experimental setting that covers most of the common uses of the Transformer in natural language processing. Surprisingly, we find that most modifications do not meaningfully improve performance. Furthermore, most of the Transformer variants we found beneficial were either developed in the same codebase that we used or are relatively minor changes. We conjecture that performance improvements may strongly depend on implementation details and correspondingly make some recommendations for improving the generality of experimental results.

Transformer Feed-Forward Layers Are Key-Value Memories

Feed-forward layers constitute two-thirds of a transformer model's parameters, yet their role in the network remains under-explored. We show that feed-forward layers in transformer-based language models operate as key-value memories, where each key correlates with textual patterns in the training examples, and each value induces a distribution over the output vocabulary. Our experiments show that the learned patterns are human-interpretable, and that lower layers tend to capture shallow patterns, while upper layers learn more semantic ones. The values complement the keys' input patterns by inducing output distributions that concentrate probability mass on tokens likely to appear immediately after each pattern, particularly in the upper layers. Finally, we demonstrate that the output of a feed-forward layer is a composition of its memories, which is subsequently refined throughout the model's layers via residual connections to produce the final output distribution.

Attention Weights in Transformer NMT Fail Aligning Words Between Sequences but Largely Explain Model Predictions

This work proposes an extensive analysis of the Transformer architecture in the Neural Machine Translation (NMT) setting. Focusing on the encoder-decoder attention mechanism, we prove that attention weights systematically make alignment errors by relying mainly on uninformative tokens from the source sequence. However, we observe that NMT models assign attention to these tokens to regulate the contribution in the prediction of the two contexts, the source and the prefix of the target sequence. We provide evidence about the influence of wrong alignments on the model behavior, demonstrating that the encoder-decoder attention mechanism is well suited as an interpretability method for NMT. Finally, based on our analysis, we propose methods that largely reduce the word alignment error rate compared to standard induced alignments from attention weights.

Gated Transformer for Robust De-noised Sequence-to-Sequence Modelling

Robust sequence-to-sequence modelling is an essential task in the real world where the inputs are often noisy. Both user-generated and machine generated inputs contain various kinds of noises in the form of spelling mistakes, grammatical errors, character recognition errors, all of which impact downstream tasks and affect interpretability of texts. In this work, we devise a novel sequence-to-sequence architecture for detecting and correcting different real world and artificial noises (adversarial attacks) from English texts. Towards that we propose a modified Transformer-based encoder-decoder architecture that uses a gating mechanism to detect types of corrections required and accordingly corrects texts. Experimental results show that our gated architecture with pre-trained language models perform significantly better that the non-gated counterparts and other state-of-the-art error correction models in correcting spelling and grammatical errors. Extrinsic evaluation of our model on Machine Translation (MT) and Summarization tasks show the competitive performance of the model against other generative sequence-to-sequence models under noisy inputs.

On the Effects of Transformer Size on In- and Out-of-Domain Calibration

Large, pre-trained transformer language models, which are pervasive in natural language processing tasks, are notoriously expensive to train. To reduce the cost of training such large models, prior work has developed smaller, more compact models which achieves a significant speedup in training time while maintaining competitive accuracy to the original model on downstream tasks. Though these smaller pre-trained models have been widely adopted by the community, it is not known how well are they calibrated compared to their larger counterparts. In this paper, focusing on a wide range of tasks, we thoroughly investigate the calibration properties of pre-trained transformers, as a function of their size. We demonstrate that when evaluated in-domain, smaller models are able to achieve competitive, and often better, calibration compared to larger models, while achieving significant speedup in training time. Post-hoc calibration techniques further reduce calibration error for all models in-domain. However, when evaluated out-of-domain, larger models tend to be better calibrated, and label-smoothing instead is an effective strategy to calibrate models in this setting.

A Simple and Effective Positional Encoding for Transformers

Transformer models are permutation equivariant. To supply the order and type information of the input tokens, position and segment embeddings are usually added to the input. Recent works proposed variations of positional encodings with relative position encodings achieving better performance. Our analysis shows that the gain actually comes from moving positional information to attention layer from the input. Motivated by this, we introduce Decoupled Positional Attention for Transformers (DIET), a simple yet effective mechanism to encode position and segment information into the Transformer models. The proposed method has faster training and inference time, while achieving competitive performance on GLUE, XTREME and WMT benchmarks. We further generalize our method to long-range transformers and show performance gain.

fBERT: A Neural Transformer for Identifying Offensive Content

Transformer-based models such as BERT, XLNET, and XLM-R have achieved state-of-the-art performance across various NLP tasks including the identification of offensive language and hate speech, an important problem in social media. In this paper, we present fBERT, a BERT model retrained on SOLID, the largest English offensive language identification corpus available with over 1.4 million offensive instances. We evaluate fBERT's performance on identifying offensive content on multiple English datasets and we test several thresholds for selecting instances from SOLID. The fBERT model will be made freely available to the community.