# deberta_decodingenhanced_bert_with_disentangled_attention__f153554f.pdf

Published as a conference paper at ICLR 2021

DEBERTA: DECODING-ENHANCED BERT WITH DISENTANGLED ATTENTION

Pengcheng He1, Xiaodong Liu2, Jianfeng Gao2, Weizhu Chen1

1 Microsoft Dynamics 365 AI 2 Microsoft Research {penhe,xiaodl,jfgao,wzchen}@microsoft.com

Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. In this paper we propose a new model architecture De BERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and Ro BERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively. Second, an enhanced mask decoder is used to incorporate absolute positions in the decoding layer to predict the masked tokens in model pre-training. In addition, a new virtual adversarial training method is used for fine-tuning to improve models generalization. We show that these techniques significantly improve the efficiency of model pre-training and the performance of both natural language understand (NLU) and natural langauge generation (NLG) downstream tasks. Compared to Ro BERTa-Large, a De BERTa model trained on half of the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9% (90.2% vs. 91.1%), on SQu AD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). Notably, we scale up De BERTa by training a larger version that consists of 48 Transform layers with 1.5 billion parameters. The significant performance boost makes the single De BERTa model surpass the human performance on the Super GLUE benchmark (Wang et al., 2019a) for the first time in terms of macro-average score (89.9 versus 89.8), and the ensemble De BERTa model sits atop the Super GLUE leaderboard as of January 6, 2021, outperforming the human baseline by a decent margin (90.3 versus 89.8). The pre-trained De BERTa models and the source code were released at: https://github.com/microsoft/De BERTa1.

1 INTRODUCTION

The Transformer has become the most effective neural network architecture for neural language modeling. Unlike recurrent neural networks (RNNs) that process text in sequence, Transformers apply self-attention to compute in parallel every word from the input text an attention weight that gauges the influence each word has on another, thus allowing for much more parallelization than RNNs for large-scale model training (Vaswani et al., 2017). Since 2018, we have seen the rise of a set of large-scale Transformer-based Pre-trained Language Models (PLMs), such as GPT (Radford et al., 2019; Brown et al., 2020), BERT (Devlin et al., 2019), Ro BERTa (Liu et al., 2019c), XLNet (Yang et al., 2019), Uni LM (Dong et al., 2019), ELECTRA (Clark et al., 2020), T5 (Raffel et al., 2020), ALUM (Liu et al., 2020), Struct BERT (Wang et al., 2019c) and ERINE (Sun et al., 2019) . These PLMs have been fine-tuned using task-specific labels and created new state of the art in many downstream natural language processing (NLP) tasks (Liu et al., 2019b; Minaee et al., 2020; Jiang et al., 2020; He et al., 2019a;b; Shen et al., 2020).

1Our code and models are also available at Hugging Face Transformers: https://github.com/ huggingface/transformers, https://huggingface.co/models?filter=deberta

Published as a conference paper at ICLR 2021

In this paper, we propose a new Transformer-based neural language model De BERTa (Decodingenhanced BERT with disentangled attention), which improves previous state-of-the-art PLMs using two novel techniques: a disentangled attention mechanism, and an enhanced mask decoder.

Disentangled attention. Unlike BERT where each word in the input layer is represented using a vector which is the sum of its word (content) embedding and position embedding, each word in De BERTa is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices based on their contents and relative positions, respectively. This is motivated by the observation that the attention weight of a word pair depends on not only their contents but their relative positions. For example, the dependency between the words deep and learning is much stronger when they occur next to each other than when they occur in different sentences.

Enhanced mask decoder. Like BERT, De BERTa is pre-trained using masked language modeling (MLM). MLM is a fill-in-the-blank task, where a model is taught to use the words surrounding a mask token to predict what the masked word should be. De BERTa uses the content and position information of the context words for MLM. The disentangled attention mechanism already considers the contents and relative positions of the context words, but not the absolute positions of these words, which in many cases are crucial for the prediction. Consider the sentence a new store opened beside the new mall with the italicized words store and mall masked for prediction. Although the local contexts of the two words are similar, they play different syntactic roles in the sentence. (Here, the subject of the sentence is store not mall, for example.) These syntactical nuances depend, to a large degree, upon the words absolute positions in the sentence, and so it is important to account for a word s absolute position in the language modeling process. De BERTa incorporates absolute word position embeddings right before the softmax layer where the model decodes the masked words based on the aggregated contextual embeddings of word contents and positions.

In addition, we propose a new virtual adversarial training method for fine-tuning PLMs to downstream NLP tasks. The method is effective in improving models generalization.

We show through a comprehensive empirical study that these techniques substantially improve the efficiency of pre-training and the performance of downstream tasks. In the NLU tasks, compared to Ro BERTa-Large, a De BERTa model trained on half the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9% (90.2% vs. 91.1%), on SQu AD v2.0 by +2.3%(88.4% vs. 90.7%), and RACE by +3.6% (83.2% vs. 86.8%). In the NLG tasks, De BERTa reduces the perplexity from 21.6 to 19.5 on the Wikitext-103 dataset. We further scale up De BERTa by pre-training a larger model that consists of 48 Transformer layers with 1.5 billion parameters. The single 1.5B-parameter De BERTa model substantially outperforms T5 with 11 billion parameters on the Super GLUE benchmark (Wang et al., 2019a) by 0.6%(89.3% vs. 89.9%), and surpasses the human baseline (89.9 vs. 89.8) for the first time. The ensemble De BERTa model sits atop the Super GLUE leaderboard as of January 6, 2021, outperforming the human baseline by a decent margin (90.3 versus 89.8).

2 BACKGROUND

2.1 TRANSFORMER

A Transformer-based language model is composed of stacked Transformer blocks (Vaswani et al., 2017). Each block contains a multi-head self-attention layer followed by a fully connected positional feed-forward network. The standard self-attention mechanism lacks a natural way to encode word position information. Thus, existing approaches add a positional bias to each input word embedding so that each input word is represented by a vector whose value depends on its content and position. The positional bias can be implemented using absolute position embedding (Vaswani et al., 2017; Radford et al., 2019; Devlin et al., 2019) or relative position embedding (Huang et al., 2018; Yang et al., 2019). It has been shown that relative position representations are more effective for natural language understanding and generation tasks (Dai et al., 2019; Shaw et al., 2018). The proposed disentangled attention mechanism differs from all existing approaches in that we represent each input word using two separate vectors that encode a word s content and position, respectively, and

Published as a conference paper at ICLR 2021

attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively.

2.2 MASKED LANGUAGE MODEL

Large-scale Transformer-based PLMs are typically pre-trained on large amounts of text to learn contextual word representations using a self-supervision objective, known as Masked Language Model (MLM) (Devlin et al., 2019). Specifically, given a sequence X txiu, we corrupt it into X by masking 15% of its tokens at random and then train a language model parameterized by θ to reconstruct X by predicting the masked tokens x conditioned on X:

max θ log pθp X| Xq max θ

i PC log pθp xi xi| Xq (1)

where C is the index set of the masked tokens in the sequence. The authors of BERT propose to keep 10% of the masked tokens unchanged, another 10% replaced with randomly picked tokens and the rest replaced with the [MASK] token.

3 THE DEBERTA ARCHITECTURE

3.1 DISENTANGLED ATTENTION: A TWO-VECTOR APPROACH TO CONTENT AND POSITION EMBEDDING

For a token at position i in a sequence, we represent it using two vectors, t Hiu and t Pi|ju, which represent its content and relative position with the token at position j, respectively. The calculation of the cross attention score between tokens i and j can be decomposed into four components as Ai,j t Hi, Pi|ju ˆ t Hj, Pj|iu

Hi H j Hi P j|i Pi|j H j Pi|j P j|i (2)

That is, the attention weight of a word pair can be computed as a sum of four attention scores using disentangled matrices on their contents and positions as content-to-content, content-to-position, position-to-content, and position-to-position 2.

Existing approaches to relative position encoding use a separate embedding matrix to compute the relative position bias in computing attention weights (Shaw et al., 2018; Huang et al., 2018). This is equivalent to computing the attention weights using only the content-to-content and content-toposition terms in equation 2. We argue that the position-to-content term is also important since the attention weight of a word pair depends not only on their contents but on their relative positions, which can only be fully modeled using both the content-to-position and position-to-content terms. Since we use relative position embedding, the position-to-position term does not provide much additional information and is removed from equation 2 in our implementation.

Taking single-head attention as an example, the standard self-attention operation (Vaswani et al., 2017) can be formulated as:

Q HWq, K HWk, V HWv, A QK

d Ho softmaxp Aq V

where H P RNˆd represents the input hidden vectors, Ho P RNˆd the output of self-attention, Wq, Wk, Wv P Rdˆd the projection matrices, A P RNˆN the attention matrix, N the length of the input sequence, and d the dimension of hidden states.

Denote k as the maximum relative distance, δpi, jq P r0, 2kq as the relative distance from token i to token j, which is defined as:

# 0 for i j ď k 2k 1 for i j ě k i j k others. (3)

2In this sense, our model shares some similarity to Tensor Product Representation (Smolensky, 1990; Schlag et al., 2019; Chen et al., 2019) where a word is represented using a tensor product of its filler (content) vector and its role (position) vector.

Published as a conference paper at ICLR 2021

We can represent the disentangled self-attention with relative position bias as equation 4, where Qc, Kc and Vc are the projected content vectors generated using projection matrices Wq,c, Wk,c, Wv,c P Rdˆd respectively, P P R2kˆd represents the relative position embedding vectors shared across all layers (i.e., staying fixed during forward propagation), and Qr and Kr are projected relative position vectors generated using projection matrices Wq,r, Wk,r P Rdˆd, respectively.

Qc HWq,c, Kc HWk,c, Vc HWv,c, Qr P Wq,r, Kr P Wk,r Ai,j Qc i Kc j looomooon

(a) content-to-content

Qc i Kr δpi,jq looooomooooon

(b) content-to-position

Kc j Qr δpj,iq looooomooooon

(c) position-to-content

Ho softmaxp A ?

Ai,j is the element of attention matrix A, representing the attention score from token i to token j. Qc i is the i-th row of Qc. Kc j is the j-th row of Kc. Kr δpi,jq is the δpi, jq-th row of Kr with regarding to relative distance δpi, jq. Qr δpj,iq is the δpj, iq-th row of Qr with regarding to relative distance δpj, iq. Note that we use δpj, iq rather than δpi, jq here. This is because for a given position i, position-to-content computes the attention weight of the key content at j with respect to the query position at i, thus the relative distance is δpj, iq. The position-to-content term is calculated as Kc j Qr δpj,iq . The content-to-position term is calculated in a similar way.

Finally, we apply a scaling factor of 1 ?

3d on A. The factor is important for stabilizing model training (Vaswani et al., 2017), especially for large-scale PLMs.

Algorithm 1 Disentangled Attention

Input: Hidden state H, relative distance embedding P , relative distance matrix δ. Content projection matrix Wk,c, Wq,c, Wv,c, position projection matrix Wk,r, Wq,r. 1: Kc HWk,c, Qc HWq,c, Vc HWv,c, Kr P Wk,r, Qr P Wq,r 2: AcÑc Qc K c 3: for i 0, ..., N 1 do 4: AcÑpri, :s Qcri, :s K r 5: end for 6: for i 0, ..., N 1 do 7: for j 0, ..., N 1 do 8: AcÑpri, js AcÑpri, δri, jss 9: end for 10: end for 11: for j 0, ..., N 1 do 12: ApÑcr:, js Kcrj, :s Q r 13: end for 14: for j 0, ..., N 1 do 15: for i 0, ..., N 1 do 16: ApÑcri, js ApÑcrδrj, is, js 17: end for 18: end for 19: A AcÑc AcÑp ApÑc 20: Ho softmaxp A ?

3dq Vc Output: Ho

3.1.1 EFFICIENT IMPLEMENTATION

For an input sequence of length N, it requires a space complexity of Op N 2dq (Shaw et al., 2018; Huang et al., 2018; Dai et al., 2019) to store the relative position embedding for each token. However, taking content-to-position as an example, we note that since δpi, jq P r0, 2kq and the embeddings

Published as a conference paper at ICLR 2021

of all possible relative positions are always a subset of Kr P R2kˆd, then we can reuse Kr in the attention calculation for all the queries.

In our experiments, we set the maximum relative distance k to 512 for pre-training. The disentangled attention weights can be computed efficiently using Algorithm 1. Let δ be the relative position matrix according to equation 3, i.e., δri, js δpi, jq. Instead of allocating a different relative position embedding matrix for each query, we multiply each query vector Qcri, :s by K r P Rdˆ2k, as in line 3 5. Then, we extract the attention weight using the relative position matrix δ as the index, as in line 6 10. To compute the position-to-content attention score, we calculate ApÑcr:, js, i.e., the column vector of the attention matrix ApÑc, by multiplying each key vector Kcrj, :s by Q r, as in line 11 13. Finally, we extract the corresponding attention score via the relative position matrix δ as the index, as in line 14 18. In this way, we do not need to allocate memory to store a relative position embedding for each query and thus reduce the space complexity to Opkdq (for storing Kr and Qr).

3.2 ENHANCED MASK DECODER ACCOUNTS FOR ABSOLUTE WORD POSITIONS

De BERTa is pretrained using MLM, where a model is trained to use the words surrounding a mask token to predict what the masked word should be. De BERTa uses the content and position information of the context words for MLM. The disentangled attention mechanism already considers the contents and relative positions of the context words, but not the absolute positions of these words, which in many cases are crucial for the prediction.

Given a sentence a new store opened beside the new mall with the words store and mall masked for prediction. Using only the local context (e.g., relative positions and surrounding words) is insufficient for the model to distinguish store and mall in this sentence, since both follow the word new with the same relative positions. To address this limitation, the model needs to take into account absolute positions, as complement information to the relative positions. For example, the subject of the sentence is store not mall . These syntactical nuances depend, to a large degree, upon the words absolute positions in the sentence.

There are two methods of incorporating absolute positions. The BERT model incorporates absolute positions in the input layer. In De BERTa, we incorporate them right after all the Transformer layers but before the softmax layer for masked token prediction, as shown in Figure 2. In this way, De BERTa captures the relative positions in all the Transformer layers and only uses absolute positions as complementary information when decoding the masked words. Thus, we call De BERTa s decoding component an Enhanced Mask Decoder (EMD). In the empirical study, we compare these two methods of incorporating absolute positions and observe that EMD works much better. We conjecture that the early incorporation of absolute positions used by BERT might undesirably hamper the model from learning sufficient information of relative positions. In addition, EMD also enables us to introduce other useful information, in addition to positions, for pre-training. We leave it to future work.

4 SCALE INVARIANT FINE-TUNING

This section presents a new virtual adversarial training algorithm, Scale-invariant-Fine-Tuning (Si FT), a variant to the algorithm described in Miyato et al. (2018); Jiang et al. (2020), for fine-tuning.

Virtual adversarial training is a regularization method for improving models generalization. It does so by improving a model s robustness to adversarial examples, which are created by making small perturbations to the input. The model is regularized so that when given a task-specific example, the model produces the same output distribution as it produces on an adversarial perturbation of that example.

For NLP tasks, the perturbation is applied to the word embedding instead of the original word sequence. However, the value ranges (norms) of the embedding vectors vary among different words and models. The variance gets larger for bigger models with billions of parameters, leading to some instability of adversarial training.

Published as a conference paper at ICLR 2021

Inspired by layer normalization (Ba et al., 2016), we propose the Si FT algorithm that improves the training stability by applying the perturbations to the normalized word embeddings. Specifically, when fine-tuning De BERTa to a downstream NLP task in our experiments, Si FT first normalizes the word embedding vectors into stochastic vectors, and then applies the perturbation to the normalized embedding vectors. We find that the normalization substantially improves the performance of the fine-tuned models. The improvement is more prominent for larger De BERTa models. Note that we only apply Si FT to De BERTa1.5B on Super GLUE tasks in our experiments and we will provide a more comprehensive study of Si FT in our future work.

5 EXPERIMENT

This section reports De BERTa results on various NLU tasks.

5.1 MAIN RESULTS ON NLU TASKS

Following previous studies of PLMs, we report results using large and base models.

5.1.1 PERFORMANCE ON LARGE MODELS

Model Co LA QQP MNLI-m/mm SST-2 STS-B QNLI RTE MRPC Avg. Mcc Acc Acc Acc Corr Acc Acc Acc

BERTlarge 60.6 91.3 86.6/- 93.2 90.0 92.3 70.4 88.0 84.05 Ro BERTalarge 68.0 92.2 90.2/90.2 96.4 92.4 93.9 86.6 90.9 88.82 XLNetlarge 69.0 92.3 90.8/90.8 97.0 92.5 94.9 85.9 90.8 89.15 ELECTRAlarge 69.1 92.4 90.9/- 96.9 92.6 95.0 88.0 90.8 89.46 De BERTalarge 70.5 92.3 91.1/91.1 96.8 92.8 95.3 88.3 91.9 90.00

Table 1: Comparison results on the GLUE development set.

We pre-train our large models following the setting of BERT (Devlin et al., 2019), except that we use the BPE vocabulary of Radford et al. (2019); Liu et al. (2019c). For training data, we use Wikipedia (English Wikipedia dump3; 12GB), Book Corpus (Zhu et al., 2015) (6GB), OPENWEBTEXT (public Reddit content (Gokaslan & Cohen, 2019); 38GB), and STORIES (a subset of Common Crawl (Trinh & Le, 2018); 31GB). The total data size after data deduplication (Shoeybi et al., 2019) is about 78G. Refer to Appendix A.2 for a detailed description of the pre-training dataset.

We use 6 DGX-2 machines (96 V100 GPUs) to train the models. A single model trained with 2K batch size and 1M steps takes about 20 days. Refer to Appendix A for the detailed hyperparamters.

We summarize the results on eight NLU tasks of GLUE (Wang et al., 2019b) in Table 1, where De BERTa is compared De BERTa with previous Transform-based PLMs of similar structures (i.e. 24 layers with hidden size of 1024) including BERT, Ro BERTa, XLNet, ALBERT and ELECTRA. Note that Ro BERTa, XLNet and ELECTRA are pre-trained on 160G training data while De BERTa is pretrained on 78G training data. Ro BERTa and XLNet are pre-trained for 500K steps with 8K samples in a step, which amounts to four billion training samples. De BERTa is pre-trained for one million steps with 2K samples in each step. This amounts to two billion training samples, approximately half of either Ro BERTa or XLNet. Table 1 shows that compared to BERT and Ro BERTa, De BERTa performs consistently better across all the tasks. Meanwhile, De BERTa outperforms XLNet in six out of eight tasks. Particularly, the improvements on MRPC (1.1% over XLNet and 1.0% over Ro BERTa), RTE (2.4% over XLNet and 1.7% over Ro BERTa) and Co LA (1.5% over XLNet and 2.5% over Ro BERTa) are significant. De BERTa also outperforms other SOTA PLMs, i.e., ELECTRAlarge and XLNetlarge, in terms of average GLUE score.

Among all GLUE tasks, MNLI is most often used as an indicative task to monitor the research progress of PLMs. De BERTa significantly outperforms all existing PLMs of similar size on MNLI and creates a new state of the art.

3https://dumps.wikimedia.org/enwiki/

Published as a conference paper at ICLR 2021

Model MNLI-m/mm SQu AD v1.1 SQu AD v2.0 RACE Re Co RD SWAG NER

Acc F1/EM F1/EM Acc F1/EM Acc F1

BERTlarge 86.6/- 90.9/84.1 81.8/79.0 72.0 - 86.6 92.8 ALBERTlarge 86.5/- 91.8/85.2 84.9/81.8 75.2 - - - Ro BERTalarge 90.2/90.2 94.6/88.9 89.4/86.5 83.2 90.6/90.0 89.9 93.4 XLNetlarge 90.8/90.8 95.1/89.7 90.6/87.9 85.4 - - - Megatron336M 89.7/90.0 94.2/88.0 88.1/84.8 83.0 - - - De BERTalarge 91.1/91.1 95.5/90.1 90.7/88.0 86.8 91.4/91.0 90.8 93.8 ALBERTxxlarge 90.8/- 94.8/89.3 90.2/87.4 86.5 - - - Megatron1.3B 90.9/91.0 94.9/89.1 90.2/87.1 87.3 - - - Megatron3.9B 91.4/91.4 95.5/90.0 91.2/88.5 89.5 - - -

Table 2: Results on MNLI in/out-domain, SQu AD v1.1, SQu AD v2.0, RACE, Re Co RD, SWAG, Co NLL 2003 NER development set. Note that missing results in literature are signified by - .

In addition to GLUE, De BERTa is evaluated on three categories of NLU benchmarks: (1) Question Answering: SQu AD v1.1 (Rajpurkar et al., 2016), SQu AD v2.0 (Rajpurkar et al., 2018), RACE (Lai et al., 2017), Re Co RD (Zhang et al., 2018) and SWAG (Zellers et al., 2018); (2) Natural Language Inference: MNLI (Williams et al., 2018); and (3) NER: Co NLL-2003. For comparison, we include ALBERTxxlarge (Lan et al., 2019) 4 and Megatron (Shoeybi et al., 2019) with three different model sizes, denoted as Megatron336M, Megatron1.3B and Megatron3.9B, respectively, which are trained using the same dataset as Ro BERTa. Note that Megatron336M has a similar model size as other models mentioned above5.

We summarize the results in Table 2. Compared to the previous SOTA PLMs with a similar model size (i.e., BERT, Ro BERTa, XLNet, ALBERTlarge, and Megatron336M), De BERTa shows superior performance in all seven tasks. Taking the RACE benchmark as an example, De BERTa significantly outperforms XLNet by +1.4% (86.8% vs. 85.4%). Although Megatron1.3B is three times larger than De BERTa, De BERTa outperforms it in three of the four benchmarks. We further report De BERTa on text generation tasks in Appendix A.4.

5.1.2 PERFORMANCE ON BASE MODELS

Our setting for base model pre-training is similar to that for large models. The base model structure follows that of the BERT base model, i.e., L 12, H 768, A 12. We use 4 DGX-2 with 64 V100 GPUs to train the base model. It takes 10 days to finish a single pre-training of 1M training steps with batch size 2048. We train De BERTa using the same 78G dataset, and compare it to Ro BERTa and XLNet trained on 160G text data.

We summarize the base model results in Table 3. Across all three tasks, De BERTa consistently outperforms Ro BERTa and XLNet by a larger margin than that in large models. For example, on MNLI-m, De BERTabase obtains +1.2% (88.8% vs. 87.6%) over Ro BERTabase, and +2% (88.8% vs. 86.8%) over XLNetbase.

Model MNLI-m/mm (Acc) SQu AD v1.1 (F1/EM) SQu AD v2.0 (F1/EM)

Ro BERTabase 87.6/- 91.5/84.6 83.7/80.5 XLNetbase 86.8/- -/- -/80.2 De BERTabase 88.8/88.5 93.1/87.2 86.2/83.1

Table 3: Results on MNLI in/out-domain (m/mm), SQu AD v1.1 and v2.0 development set.

4The hidden dimension of ALBERTxxlarge is 4 times of De BERTa and the computation cost is about 4 times of De BERTa. 5T5 (Raffel et al., 2020) has more parameters (11B). Raffel et al. (2020) only report the test results of T5 which are not comparable with other models.

Published as a conference paper at ICLR 2021

5.2 MODEL ANALYSIS

In this section, we first present an ablation study to quantify the relative contributions of different components introduced in De BERTa. Then, we study the convergence property to characterize the model training efficiency. We run experiments for analysis using the base model setting: a model is pre-trained using the Wikipedia + Bookcorpus dataset for 1M steps with batch size 256 in 7 days on a DGX-2 machine with 16 V-100 GPUs. Due to space limit, we visualize the different attention patterns of De BERTa and Ro BERTa in Appendix A.7.

5.2.1 ABLATION STUDY

To verify our experimental setting, we pre-train the Ro BERTa base model from scratch. The re-pretrained Ro BERTa model is denoted as Ro BERTa-Re Impbase. To investigate the relative contributions of different components in De BERTa, we develop three variations:

-EMD is the De BERTa base model without EMD. -C2P is the De BERTa base model without the content-to-position term ((c) in Eq. 4). -P2C is the De BERTa base model without the position-to-content term ((b) in Eq. 4). As XLNet also uses the relative position bias, this model is close to XLNet plus EMD.

Model MNLI-m/mm SQu AD v1.1 SQu AD v2.0 RACE Acc F1/EM F1/EM Acc

BERTbase Devlin et al. (2019) 84.3/84.7 88.5/81.0 76.3/73.7 65.0 Ro BERTabase Liu et al. (2019c) 84.7/- 90.6/- 79.7/- 65.6 XLNetbase Yang et al. (2019) 85.8/85.4 -/- 81.3/78.5 66.7 Ro BERTa-Re Impbase 84.9/85.1 91.1/84.8 79.5/76.0 66.8 De BERTabase 86.3/86.2 92.1/86.1 82.5/79.3 71.7 -EMD 86.1/86.1 91.8/85.8 81.3/78.0 70.3 -C2P 85.9/85.7 91.6/85.8 81.3/78.3 69.3 -P2C 86.0/85.8 91.7/85.7 80.8/77.6 69.6 -(EMD+C2P) 85.8/85.9 91.5/85.3 80.3/77.2 68.1 -(EMD+P2C) 85.8/85.8 91.3/85.1 80.2/77.1 68.5

Table 4: Ablation study of the De BERTa base model.

Table 4 summarizes the results on four benchmark datasets. First, Ro BERTa-Re Imp performs similarly to Ro BERTa across all benchmark datasets, verfiying that our setting is reasonable. Second, we see that removing any one component in De BERTa results in a sheer performance drop. For instance, removing EMD (-EMD) results in a loss of 1.4% (71.7% vs. 70.3%) on RACE, 0.3% (92.1% vs. 91.8%) on SQu AD v1.1, 1.2% (82.5% vs. 81.3%) on SQu AD v2.0, 0.2% (86.3% vs. 86.1%) and 0.1% (86.2% vs. 86.1%) on MNLI-m/mm, respectively. Similarly, removing either content-to-position or position-to-content leads to inferior performance in all the benchmarks. As expected, removing two components results in even more substantial loss in performance.

5.3 SCALE UP TO 1.5 BILLION PARAMETERS

Larger pre-trained models have shown better generalization results (Raffel et al., 2020; Brown et al., 2020; Shoeybi et al., 2019). Thus, we have built a larger version of De BERTa with 1.5 billion parameters, denoted as De BERTa1.5B. The model consists of 48 layers with a hidden size of 1,536 and 24 attention heads 6. De BERTa1.5B is trained on a pre-training dataset amounting to 160G, similar to that in Liu et al. (2019c), with a new vocabulary of size 128K constructed using the dataset.

To train De BERTa1.5B, we optimize the model architecture as follows. First, we share the projection matrices of relative position embedding Wk,r, Wq,r with Wk,c, Wq,c, respectively, in all attention layers to reduce the number of model parameters. Our ablation study in Table 13 on base models shows that the projection matrix sharing reduces the model size while retaining the model performance.

6See Table 8 in Appendix for the model hyperparameters.

Published as a conference paper at ICLR 2021

Second, a convolution layer is added aside the first Transformer layer to induce n-gram knowledge of sub-word encodings and their outputs are summed up before feeding to the next Transformer layer 7.

Table 5 reports the test results of Super GLUE (Wang et al., 2019a) which is one of the most popular NLU benchmarks. Super GLUE consists of a wide of NLU tasks, including Question Answering (Clark et al., 2019; Khashabi et al., 2018; Zhang et al., 2018), Natural Language Inference (Dagan et al., 2006; Bar-Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009), Word Sense Disambiguation (Pilehvar & Camacho-Collados, 2019), and Reasoning (Levesque et al., 2011; Roemmele et al., 2011). Since its release in 2019, top research teams around the world have been developing large-scale PLMs that have driven striking performance improvement on Super GLUE.

The significant performance boost due to scaling De BERTa to a larger model makes the single De BERTa1.5B surpass the human performance on Super GLUE for the first time in terms of macroaverage score (89.9 versus 89.8) as of December 29, 2020, and the ensemble De BERTa model (De BERTa Ensemble) sits atop the Super GLUE benchmark rankings as of January 6, 2021, outperforming the human baseline by a decent margin (90.3 versus 89.8). Compared to T5, which consists of 11 billion parameters, the 1.5-billion-parameter De BERTa is much more energy efficient to train and maintain, and it is easier to compress and deploy to apps of various settings.

Model Bool Q CB COPA Multi RC Re Co RD RTE Wi C WSC Average Acc F1/Acc Acc F1a/EM F1/EM Acc Acc Acc Score

Ro BERTalarge 87.1 90.5/95.2 90.6 84.4/52.5 90.6/90.0 88.2 69.9 89.0 84.6 NEXHA-Plus 87.8 94.4/96.0 93.6 84.6/55.1 90.1/89.6 89.1 74.6 93.2 86.7 T511B 91.2 93.9/96.8 94.8 88.1/63.3 94.1/93.4 92.5 76.9 93.8 89.3 T511B+Meena 91.3 95.8/97.6 97.4 88.3/63.0 94.2/93.5 92.7 77.9 95.9 90.2 Human 89.0 95.8/98.9 100.0 81.8/51.9 91.7/91.3 93.6 80.0 100.0 89.8 De BERTa1.5B+Si FT 90.4 94.9/97.2 96.8 88.2/63.7 94.5/94.1 93.2 76.4 95.9 89.9 De BERTa Ensemble 90.4 95.7/97.6 98.4 88.2/63.7 94.5/94.1 93.2 77.5 95.9 90.3

Table 5: Super GLUE test set results scored using the Super GLUE evaluation server. All the results are obtained from https://super.gluebenchmark.com on January 6, 2021.

6 CONCLUSIONS

This paper presents a new model architecture De BERTa (Decoding-enhanced BERT with disentangled attention) that improves the BERT and Ro BERTa models using two novel techniques. The first is the disentangled attention mechanism, where each word is represented using two vectors that encode its content and position, respectively, and the attention weights among words are computed using disentangled matrices on their contents and relative positions, respectively. The second is an enhanced mask decoder which incorporates absolute positions in the decoding layer to predict the masked tokens in model pre-training. In addition, a new virtual adversarial training method is used for fine-tuning to improve model s generalization on downstream tasks.

We show through a comprehensive empirical study that these techniques significantly improve the efficiency of model pre-training and the performance of downstream tasks. The De BERTa model with 1.5 billion parameters surpasses the human performance on the Super GLUE benchmark for the first time in terms of macro-average score.

De BERTa surpassing human performance on Super GLUE marks an important milestone toward general AI. Despite its promising results on Super GLUE, the model is by no means reaching the human-level intelligence of NLU. Humans are extremely good at leveraging the knowledge learned from different tasks to solve a new task with no or little task-specific demonstration. This is referred to as compositional generalization, the ability to generalize to novel compositions (new tasks) of familiar constituents (subtasks or basic problem-solving skills). Moving forward, it is worth exploring how to make De BERTa incorporate compositional structures in a more explicit manner, which could allow combining neural and symbolic computation of natural language similar to what humans do.

7Please refer to Table 12 in Appendix A.6 for the ablation study of different model sizes, and Table 13 in Appendix A.6 for the ablation study of new modifications.

Published as a conference paper at ICLR 2021

7 ACKNOWLEDGMENTS

We thank Jade Huang and Nikos Karampatziakis for proofreading the paper and providing insightful comments. We thank Yoyo Liang, Saksham Singhal, Xia Song, and Saurabh Tiwary for their help with large-scale model training. We also thank the anonymous reviewers for valuable discussions.

Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. ar Xiv preprint ar Xiv:1607.06450, 2016.

Roy Bar-Haim, Ido Dagan, Bill Dolan, Lisa Ferro, and Danilo Giampiccolo. The second PASCAL recognising textual entailment challenge. In Proceedings of the Second PASCAL Challenges Workshop on Recognising Textual Entailment, 01 2006.

Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. ar Xiv preprint ar Xiv:2004.05150, 2020.

Luisa Bentivogli, Ido Dagan, Hoa Trang Dang, Danilo Giampiccolo, and Bernardo Magnini. The fifth pascal recognizing textual entailment challenge. In In Proc Text Analysis Conference (TAC 09, 2009.

Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. ar Xiv preprint ar Xiv:2005.14165, 2020.

Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. ar Xiv preprint ar Xiv:1708.00055, 2017.

Kezhen Chen, Qiuyuan Huang, Hamid Palangi, Paul Smolensky, Kenneth D Forbus, and Jianfeng Gao. Natural-to formal-language generation using tensor product representations. ar Xiv preprint ar Xiv:1910.02339, 2019.

Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. ar Xiv preprint ar Xiv:1904.10509, 2019.

Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Bool Q: Exploring the surprising difficulty of natural yes/no questions. In Proceedings of NAACL-HLT 2019, 2019.

Kevin Clark, Minh-Thang Luong, Quoc V. Le, and Christopher D. Manning. ELECTRA: Pre-training text encoders as discriminators rather than generators. In ICLR, 2020.

Ido Dagan, Oren Glickman, and Bernardo Magnini. The pascal recognising textual entailment challenge. In Proceedings of the First International Conference on Machine Learning Challenges: Evaluating Predictive Uncertainty Visual Object Classification, and Recognizing Textual Entailment, MLCW 05, Berlin, Heidelberg, 2006.

Zihang Dai, Zhilin Yang, Yiming Yang, Jaime G Carbonell, Quoc Le, and Ruslan Salakhutdinov. Transformer-xl: Attentive language models beyond a fixed-length context. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 2978 2988, 2019.

Marie-Catherine De Marneffe, Mandy Simons, and Judith Tonhauser. The commitmentbank: Investigating projection in naturally occurring discourse. In proceedings of Sinn und Bedeutung, volume 23, pp. 107 124, 2019.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 4171 4186, 2019.

Published as a conference paper at ICLR 2021

William B Dolan and Chris Brockett. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP2005), 2005.

Li Dong, Nan Yang, Wenhui Wang, Furu Wei, Xiaodong Liu, Yu Wang, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Hon. Unified language model pre-training for natural language understanding and generation. In Advances in Neural Information Processing Systems, pp. 13042 13054, 2019.

Danilo Giampiccolo, Bernardo Magnini, Ido Dagan, and Bill Dolan. The third PASCAL recognizing textual entailment challenge. In Proceedings of the ACL-PASCAL Workshop on Textual Entailment and Paraphrasing, pp. 1 9, Prague, June 2007. Association for Computational Linguistics. URL https://www.aclweb.org/anthology/W07-1401.

Aaron Gokaslan and Vanya Cohen. Openwebtext corpus. http://Skylion007.github.io, 2019.

Pengcheng He, Xiaodong Liu, Weizhu Chen, and Jianfeng Gao. A hybrid neural network model for commonsense reasoning. ar Xiv preprint ar Xiv:1907.11983, 2019a.

Pengcheng He, Yi Mao, Kaushik Chakrabarti, and Weizhu Chen. X-sql: reinforce schema representation with context. ar Xiv preprint ar Xiv:1908.08113, 2019b.

Cheng-Zhi Anna Huang, Ashish Vaswani, Jakob Uszkoreit, Ian Simon, Curtis Hawthorne, Noam Shazeer, Andrew M Dai, Matthew D Hoffman, Monica Dinculescu, and Douglas Eck. Music transformer: Generating music with long-term structure. 2018.

Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Tuo Zhao. SMART: Robust and efficient fine-tuning for pre-trained natural language models through principled regularized optimization. In ACL, July 2020. doi: 10.18653/v1/2020.acl-main.197.

Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S Weld, Luke Zettlemoyer, and Omer Levy. Spanbert: Improving pre-training by representing and predicting spans. Transactions of the Association for Computational Linguistics, 8:64 77, 2020.

Daniel Khashabi, Snigdha Chaturvedi, Michael Roth, Shyam Upadhyay, and Dan Roth. Looking beyond the surface: A challenge set for reading comprehension over multiple sentences. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 252 262, 2018.

Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. ar Xiv preprint ar Xiv:1412.6980, 2014.

Nikita Kitaev, Lukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2019.

Guokun Lai, Qizhe Xie, Hanxiao Liu, Yiming Yang, and Eduard Hovy. Race: Large-scale reading comprehension dataset from examinations. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pp. 785 794, 2017.

Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. Albert: A lite bert for self-supervised learning of language representations. In International Conference on Learning Representations, 2019.

Hector Levesque, Ernest Davis, and Leora Morgenstern. The winograd schema challenge. In Thirteenth International Conference on the Principles of Knowledge Representation and Reasoning, 2012.

Hector J Levesque, Ernest Davis, and Leora Morgenstern. The Winograd schema challenge. In AAAI Spring Symposium: Logical Formalizations of Commonsense Reasoning, volume 46, pp. 47, 2011.

Liyuan Liu, Haoming Jiang, Pengcheng He, Weizhu Chen, Xiaodong Liu, Jianfeng Gao, and Jiawei Han. On the variance of the adaptive learning rate and beyond. In International Conference on Learning Representations, 2019a.

Published as a conference paper at ICLR 2021

Xiaodong Liu, Pengcheng He, Weizhu Chen, and Jianfeng Gao. Multi-task deep neural networks for natural language understanding. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pp. 4487 4496, Florence, Italy, July 2019b. Association for Computational Linguistics. URL https://www.aclweb.org/anthology/P19-1441.

Xiaodong Liu, Hao Cheng, Pengcheng He, Weizhu Chen, Yu Wang, Hoifung Poon, and Jianfeng Gao. Adversarial training for large neural language models. ar Xiv preprint ar Xiv:2004.08994, 2020.

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. ar Xiv preprint ar Xiv:1907.11692, 2019c.

Ilya Loshchilov and Frank Hutter. Fixing weight decay regularization in adam. 2018.

Stephen Merity, Caiming Xiong, James Bradbury, and Richard Socher. Pointer sentinel mixture models. ar Xiv, pp. ar Xiv 1609, 2016.

Shervin Minaee, Nal Kalchbrenner, Erik Cambria, Narjes Nikzad, Meysam Chenaghlu, and Jianfeng Gao. Deep learning based text classification: A comprehensive review. ar Xiv preprint ar Xiv:2004.03705, 2020.

Takeru Miyato, Shin-ichi Maeda, Masanori Koyama, and Shin Ishii. Virtual adversarial training: a regularization method for supervised and semi-supervised learning. IEEE transactions on pattern analysis and machine intelligence, 41(8):1979 1993, 2018.

Mohammad Taher Pilehvar and Jose Camacho-Collados. Wic: the word-in-context dataset for evaluating context-sensitive meaning representations. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pp. 1267 1273, 2019.

Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. Open AI Blog, 1(8), 2019.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1 67, 2020. URL http://jmlr.org/papers/v21/20-074.html.

Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. SQu AD: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, November 2016.

Pranav Rajpurkar, Robin Jia, and Percy Liang. Know what you don t know: Unanswerable questions for squad. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pp. 784 789, 2018.

Melissa Roemmele, Cosmin Adrian Bejan, and Andrew S. Gordon. Choice of plausible alternatives: An evaluation of commonsense causal reasoning. In 2011 AAAI Spring Symposium Series, 2011.

Imanol Schlag, Paul Smolensky, Roland Fernandez, Nebojsa Jojic, Jürgen Schmidhuber, and Jianfeng Gao. Enhancing the transformer with explicit relational encoding for math problem solving. ar Xiv preprint ar Xiv:1910.06611, 2019.

Peter Shaw, Jakob Uszkoreit, and Ashish Vaswani. Self-attention with relative position representations. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 2 (Short Papers), pp. 464 468, 2018.

Tao Shen, Yi Mao, Pengcheng He, Guodong Long, Adam Trischler, and Weizhu Chen. Exploiting structured knowledge in text via graph-guided representation learning. ar Xiv preprint ar Xiv:2004.14224, 2020.

Published as a conference paper at ICLR 2021

Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick Le Gresley, Jared Casper, and Bryan Catanzaro. Megatron-lm: Training multi-billion parameter language models using gpu model parallelism. ar Xiv preprint ar Xiv:1909.08053, 2019.

Paul Smolensky. Tensor product variable binding and the representation of symbolic structures in connectionist systems. Artificial intelligence, 46(1-2):159 216, 1990.

Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pp. 1631 1642, 2013.

Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, and Hua Wu. Ernie: Enhanced representation through knowledge integration. ar Xiv preprint ar Xiv:1904.09223, 2019.

Trieu H Trinh and Quoc V Le. A simple method for commonsense reasoning. ar Xiv preprint ar Xiv:1806.02847, 2018.

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in neural information processing systems, pp. 5998 6008, 2017.

Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Superglue: A stickier benchmark for general-purpose language understanding systems. In Advances in neural information processing systems, pp. 3266 3280, 2019a.

Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. In 7th International Conference on Learning Representations, ICLR 2019, 2019b.

Wei Wang, Bin Bi, Ming Yan, Chen Wu, Zuyi Bao, Liwei Peng, and Luo Si. Structbert: Incorporating language structures into pre-training for deep language understanding. ar Xiv preprint ar Xiv:1908.04577, 2019c.

Alex Warstadt, Amanpreet Singh, and Samuel R Bowman. Neural network acceptability judgments. ar Xiv preprint ar Xiv:1805.12471, 2018.

Adina Williams, Nikita Nangia, and Samuel Bowman. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers), pp. 1112 1122. Association for Computational Linguistics, 2018. URL http://aclweb.org/anthology/N18-1101.

Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems, pp. 5754 5764, 2019.

Rowan Zellers, Yonatan Bisk, Roy Schwartz, and Yejin Choi. Swag: A large-scale adversarial dataset for grounded commonsense inference. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 93 104, 2018.

Sheng Zhang, Xiaodong Liu, Jingjing Liu, Jianfeng Gao, Kevin Duh, and Benjamin Van Durme. Re Co RD: Bridging the gap between human and machine commonsense reading comprehension. ar Xiv preprint 1810.12885, 2018.

Yukun Zhu, Ryan Kiros, Rich Zemel, Ruslan Salakhutdinov, Raquel Urtasun, Antonio Torralba, and Sanja Fidler. Aligning books and movies: Towards story-like visual explanations by watching movies and reading books. In Proceedings of the IEEE international conference on computer vision, pp. 19 27, 2015.

Published as a conference paper at ICLR 2021

A.1 DATASET

Corpus Task #Train #Dev #Test #Label Metrics General Language Understanding Evaluation (GLUE) Co LA Acceptability 8.5k 1k 1k 2 Matthews corr SST Sentiment 67k 872 1.8k 2 Accuracy MNLI NLI 393k 20k 20k 3 Accuracy RTE NLI 2.5k 276 3k 2 Accuracy WNLI NLI 634 71 146 2 Accuracy QQP Paraphrase 364k 40k 391k 2 Accuracy/F1 MRPC Paraphrase 3.7k 408 1.7k 2 Accuracy/F1 QNLI QA/NLI 108k 5.7k 5.7k 2 Accuracy STS-B Similarity 7k 1.5k 1.4k 1 Pearson/Spearman corr Super GLUE WSC Coreference 554k 104 146 2 Accuracy Bool Q QA 9,427 3,270 3,245 2 Accuracy COPA QA 400k 100 500 2 Accuracy CB NLI 250 57 250 3 Accuracy/F1 RTE NLI 2.5k 276 3k 2 Accuracy Wi C WSD 2.5k 276 3k 2 Accuracy Re Co RD MRC 101k 10k 10k - Exact Match (EM)/F1 Multi RC Multiple choice 5,100 953 1,800 - Exact Match (EM)/F1 Question Answering SQu AD v1.1 MRC 87.6k 10.5k 9.5k - Exact Match (EM)/F1 SQu AD v2.0 MRC 130.3k 11.9k 8.9k - Exact Match (EM)/F1 RACE MRC 87,866 4,887 4,934 4 Accuracy SWAG Multiple choice 73.5k 20k 20k 4 Accuracy Token Classification Co NLL 2003 NER 14,987 3,466 3,684 8 F1

Table 6: Summary information of the NLP application benchmarks.

GLUE. The General Language Understanding Evaluation (GLUE) benchmark is a collection of nine natural language understanding (NLU) tasks. As shown in Table 6, it includes question answering (Rajpurkar et al., 2016), linguistic acceptability (Warstadt et al., 2018), sentiment analysis (Socher et al., 2013), text similarity (Cer et al., 2017), paraphrase detection (Dolan & Brockett, 2005), and natural language inference (NLI) (Dagan et al., 2006; Bar-Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009; Levesque et al., 2012; Williams et al., 2018). The diversity of the tasks makes GLUE very suitable for evaluating the generalization and robustness of NLU models.

Super GLUE. Super GLUE is an extension of the GLUE benchmark, but more difficult, which is a collection of eight NLU tasks. It covers a various of tasks including question answering (Zhang et al., 2018; Clark et al., 2019; Khashabi et al., 2018), natural language inference (Dagan et al., 2006; Bar-Haim et al., 2006; Giampiccolo et al., 2007; Bentivogli et al., 2009; De Marneffe et al., 2019), coreference resolution (Levesque et al., 2012) and word sense disambiguation (Pilehvar & Camacho-Collados, 2019).

RACE is a large-scale machine reading comprehension dataset, collected from English examinations in China, which are designed for middle school and high school students (Lai et al., 2017).

SQu AD v1.1/v2.0 is the Stanford Question Answering Dataset (SQu AD) v1.1 and v2.0 (Rajpurkar et al., 2016; 2018) are popular machine reading comprehension benchmarks. Their passages come from approximately 500 Wikipedia articles and the questions and answers are obtained by crowdsourcing. The SQu AD v2.0 dataset includes unanswerable questions about the same paragraphs.

Published as a conference paper at ICLR 2021

SWAG is a large-scale adversarial dataset for the task of grounded commonsense inference, which unifies natural language inference and physically grounded reasoning (Zellers et al., 2018). SWAG consists of 113k multiple choice questions about grounded situations.

Co NLL 2003 is an English dataset consisting of text from a wide variety of sources. It has 4 types of named entity.

A.2 PRE-TRAINING DATASET

For De BERTa pre-training, we use Wikipedia (English Wikipedia dump8; 12GB), Book Corpus (Zhu et al., 2015) 9 (6GB), OPENWEBTEXT (public Reddit content (Gokaslan & Cohen, 2019); 38GB) and STORIES10 (a subset of Common Crawl (Trinh & Le, 2018); 31GB). The total data size after data deduplication(Shoeybi et al., 2019) is about 78GB. For pre-training, we also sample 5% training data as the validation set to monitor the training process. Table 7 compares datasets used in different pre-trained models.

Model Wiki+Book Open Web Text Stories CC-News Giga5 Clue Web Common Crawl

16GB 38GB 31GB 76GB 16GB 19GB 110GB BERT XLNet Ro BERTa De BERTa De BERTa1.5B

Table 7: Comparison of the pre-training data.

A.3 IMPLEMENTATION DETAILS

Following Ro BERTa (Liu et al., 2019c), we adopt dynamic data batching. We also include span masking (Joshi et al., 2020) as an additional masking strategy with the span size up to three. We list the detailed hyperparameters of pre-training in Table 8. For pre-training, we use Adam (Kingma & Ba, 2014) as the optimizer with weight decay (Loshchilov & Hutter, 2018). For fine-tuning, even though we can get better and robust results with RAdam(Liu et al., 2019a) on some tasks, e.g. Co LA, RTE and RACE, we use Adam(Kingma & Ba, 2014) as the optimizer for a fair comparison. For fine-tuning, we train each task with a hyper-parameter search procedure, each run takes about 1-2 hours on a DGX-2 node. All the hyper-parameters are presented in Table 9. The model selection is based on the performance on the task-specific development sets.

Our code is implemented based on Huggingface Transformers11, Fair Seq12 and Megatron (Shoeybi et al., 2019)13.

A.3.1 PRE-TRAINING EFFICIENCY

To investigate the efficiency of model pre-training, we plot the performance of the fine-tuned model on downstream tasks as a function of the number of pre-training steps. As shown in Figure 1, for Ro BERTa-Re Impbase and De BERTabase, we dump a checkpoint every 150K pre-training steps, and then fine-tune the checkpoint on two representative downstream tasks, MNLI and SQu AD v2.0, and then report the accuracy and F1 score, respectively. As a reference, we also report the final model performance of both the original Ro BERTabase (Liu et al., 2019c) and XLNetbase (Yang et al., 2019). The results show that De BERTabase consistently outperforms Ro BERTa-Re Impbase during the course of pre-training.

8https://dumps.wikimedia.org/enwiki/ 9https://github.com/butsugiri/homemade_bookcorpus 10https://github.com/tensorflow/models/tree/master/research/lm_commonsense 11https://github.com/huggingface/transformers 12https://github.com/pytorch/fairseq 13https://github.com/NVIDIA/Megatron-LM

Published as a conference paper at ICLR 2021

Hyper-parameter De BERTa1.5B De BERTalarge De BERTabase De BERTabase ablation Number of Layers 48 24 12 12 Hidden size 1536 1024 768 768 FNN inner hidden size 6144 4096 3072 3072 Attention Heads 24 16 12 12 Attention Head size 64 64 64 64 Dropout 0.1 0.1 0.1 0.1 Warmup Steps 10k 10k 10k 10k Learning Rates 1.5e-4 2e-4 2e-4 1e-4 Batch Size 2k 2k 2k 256 Weight Decay 0.01 0.01 0.01 0.01 Max Steps 1M 1M 1M 1M Learning Rate Decay Linear Linear Linear Linear Adam ϵ 1e-6 1e-6 1e-6 1e-6 Adam β1 0.9 0.9 0.9 0.9 Adam β2 0.999 0.999 0.999 0.999 Gradient Clipping 1.0 1.0 1.0 1.0 Number of DGX-2 nodes 16 6 4 1 Training Time 30 days 20 days 10 days 7 days

Table 8: Hyper-parameters for pre-training De BERTa.

Hyper-parameter De BERTa1.5B De BERTalarge De BERTabase Dropout of task layer {0,0.15,0.3} {0,0.1,0.15} {0,0.1,0.15} Warmup Steps {50,100,500,1000} {50,100,500,1000} {50,100,500,1000} Learning Rates {1e-6, 3e-6, 5e-6} {5e-6, 8e-6, 9e-6, 1e-5} {1.5e-5,2e-5, 3e-5, 4e-5} Batch Size {16,32,64} {16,32,48,64} {16,32,48,64} Weight Decay 0.01 0.01 Maximun Training Epochs 10 10 10 Learning Rate Decay Linear Linear Linear Adam ϵ 1e-6 1e-6 1e-6 Adam β1 0.9 0.9 0.9 Adam β2 0.999 0.999 0.999 Gradient Clipping 1.0 1.0 1.0

Table 9: Hyper-parameters for fine-tuning De BERTa on down-streaming tasks.

(a) Results on MNLI development

(b) Results on SQu AD v2.0 development

Figure 1: Pre-training performance curve between De BERTa and its counterparts on the MNLI and SQu AD v2.0 development set.

Published as a conference paper at ICLR 2021

A.4 MAIN RESULTS ON GENERATION TASKS

In addition to NLU tasks, De BERTa can also be extended to handle NLG tasks. To allow De BERTa operating like an auto-regressive model for text generation, we use a triangular matrix for selfattention and set the upper triangular part of the self-attention mask to 8, following Dong et al. (2019).

We evaluate De BERTa on the task of auto-regressive language model (ARLM) using Wikitext103 (Merity et al., 2016). To do so, we train a new version of De BERTa, denoted as De BERTa-MT. It is jointly pre-trained using the MLM and ARLM tasks as in Uni LM (Dong et al., 2019). The pre-training hyper-parameters follows that of De BERTabase except that we use fewer training steps (200k). For comparison, we use Ro BERTa as baseline, and include GPT-2 and Transformer-XL as additional references. De BERTa-AP is a variant of De BERTa where absolute position embeddings are incorporated in the input layer as Ro BERTa. For a fair comparison, all these models are base models pre-trained in a similar setting.

Model Ro BERTa De BERTa-AP De BERTa De BERTa-MT GPT-2 Transformer-XL

Dev PPL 21.6 20.7 20.5 19.5 - 23.1 Test PPL 21.6 20.0 19.9 19.5 37.50 24

Table 10: Language model results in perplexity (lower is better) on Wikitext-103 .

Table 10 summarizes the results on Wikitext-103. We see that De BERTabase obtains lower perplexities on both dev and test data, and joint training using MLM and ARLM reduces perplexity further. That De BERTa-AP is inferior to De BERTa indicates that it is more effective to incorporate absolute position embeddings of words in the decoding layer as the EMD in De BERTa than in the input layer as Ro BERTa.

A.5 HANDLING LONG SEQUENCE INPUT

With relative position bias, we choose to truncate the maximum relative distance to k as in equation 3. Thus in each layer, each token can attend directly to at most 2pk 1q tokens and itself. By stacking Transformer layers, each token in the l th layer can attend to at most p2k 1ql tokens implicitly. Taking De BERTalarge as an example, where k 512, L 24, in theory, the maximum sequence length that can be handled is 24,528. This is a byproduct benefit of our design choice and we find it beneficial for the RACE task. A comparison of long sequence effect on the RACE task is shown in Table 11.

Sequence length Middle High Accuracy

512 88.8 85.0 86.3 768 88.7 86.3 86.8

Table 11: The effect of handling long sequence input for RACE task with De BERTa

Long sequence handling is an active research area. There have been a lot of studies where the Transformer architecture is extended for long sequence handling(Beltagy et al., 2020; Kitaev et al., 2019; Child et al., 2019; Dai et al., 2019). One of our future research directions is to extend De BERTa to deal with extremely long sequences.

A.6 PERFORMANCE IMPROVEMENTS OF DIFFERENT MODEL SCALES

In this subsection, we study the effect of different model sizes applied to large models on GLUE. Table 12 summarizes the results, showing that larger models can obtain a better result and Si FT also improves the model performance consistently.

Published as a conference paper at ICLR 2021

Model Co LA QQP MNLI-m/mm SST-2 STS-B QNLI RTE MRPC Avg. Mcc Acc Acc Acc Corr Acc Acc Acc

De BERTalarge 70.5 92.3 91.1/91.1 96.8 92.8 95.3 88.3 91.9 90.00 De BERTa900M 71.1 92.3 91.7/91.6 97.5 92.0 95.8 93.5 93.1 90.86 De BERTa1.5B 72.0 92.7 91.7/91.9 97.2 92.9 96.0 93.9 92.0 91.17 De BERTa1.5B+Si FT 73.5 93.0 92.0/92.1 97.5 93.2 96.5 96.5 93.2 91.93

Table 12: Comparison results of De BERTa models with different sizes on the GLUE development set.

Model Parameters MNLI-m/mm SQu AD v1.1 SQu AD v2.0 Acc F1/EM F1/EM Ro BERTa-Re Impbase 120M 84.9/85.1 91.1/84.8 79.5/76.0 De BERTabase 134M 86.3/86.2 92.1/86.1 82.5/79.3 + Share Projection 120M 86.3/86.3 92.2/86.2 82.3/79.5 + Conv 122M 86.3/86.5 92.5/86.4 82.5/79.7 + 128k Vocab 190M 86.7/86.9 93.1/86.8 83.0/80.1

Table 13: Ablation study of the additional modifications in De BERTa1.5B and De BERTa900M models. Note that we progressively add each component on the top of De BERTabase.

A.7 MODEL COMPLEXITY

With the disentangled attention mechanism, we introduce three additional sets of parameters Wq,r, Wk,r P Rdˆd and P P R2kˆd. The total increase in model parameters is 2L ˆ d2 2k ˆ d. For the large model pd 1024, L 24, k 512q, this amounts to about 49M additional parameters, an increase of 13%. For the base modelpd 768, L 12, k 512q, this amounts to 14M additional parameters, an increase of 12%. However, by sharing the projection matrix between content and position embedding, i.e. Wq,r Wq,c, Wk,r Wk,c, the number of parameters of De BERTa is the same as Ro BERTa. Our experiment on base model shows that the results are almost the same, as in Table 13.

The additional computational complexity is Op Nkdq due to the calculation of the additional positionto-content and content-to-position attention scores. Compared with BERT or Ro BERTa, this increases the computational cost by 30%. Compared with XLNet which also uses relative position embedding, the increase of computational cost is about 15%. A further optimization by fusing the attention computation kernel can significantly reduce this additional cost. For EMD, since the decoder in pre-training only reconstructs the masked tokens, it does not introduce additional computational cost for unmasked tokens. In the situation where 15% tokens are masked and we use only two decoder layers, the additional cost is 0.15 ˆ 2{L which results in an additional computational cost of only 3% for base model(L 12) and 2% for large model(L 24) in EMD.

A.8 ADDITIONAL DETAILS OF ENHANCED MASK DECODER

The structure of EMD is shown in Figure 2b. There are two inputs for EMD, (i.e., I, H). H denotes the hidden states from the previous Transformer layer, and I can be any necessary information for decoding, e.g., H, absolute position embedding or output from previous EMD layer. n denotes n stacked layers of EMD where the output of each EMD layer will be the input I for next EMD layer and the output of last EMD layer will be fed to the language model head directly. The n layers can share the same weight. In our experiment we share the same weight for n 2 layers to reduce the number of parameters and use absolute position embedding as I of the first EMD layer. When I H and n 1, EMD is the same as the BERT decoder layer. However, EMD is more general and flexible as it can take various types of input information for decoding.

A.9 ATTENTION PATTERNS

To visualize how De BERTa operates differently from Ro BERTa, we present in Figure 3 the attention patterns (taken in the last self-attention layers) of Ro BERTa, De BERTa and three De BERTa variants.

Published as a conference paper at ICLR 2021

Language Model Head

Transformer Layer

(a) BERT decoding layer

Language Model Head

Transformer Layer

(b) Enhanced Mask Decoder

Figure 2: Comparison of the decoding layer.

Figure 3: Comparison of attention patterns of the last layer among De BERTa, Ro BERTa and De BERTa variants (i.e., De BERTa without EMD, C2P and P2C respectively).

We observe two differences. First, Ro BERTa has a clear diagonal line effect for a token attending to itself. But this effect is not very visible in De BERTa. This can be attributed to the use of EMD, in which the absolute position embedding is added to the hidden state of content as the query vector, as verified by the attention pattern of De BERTa-EMD where the diagonal line effect is more visible than that of the original De BERTa. Second, we observe vertical strips in the attention patterns of Ro BERTa, which are mainly caused by high-frequent functional words or tokens (e.g., a , the , and punctuation). For De BERTa, the strip only appears in the first column, which represents the [CLS] token. We conjecture that a dominant emphasis on [CLS] is desirable since the feature vector of [CLS] is often used as a contextual representation of the entire input sequence in downstream tasks. We also observe that the vertical strip effect is quite obvious in the patterns of the three De BERTa variants.

We present three additional examples to illustrate the different attention patterns of De BERTa and Ro BERTa in Figures 4 and 5.

Published as a conference paper at ICLR 2021

Figure 4: Comparison on attention patterns of the last layer between De BERTa and Ro BERTa.

Published as a conference paper at ICLR 2021

Figure 5: Comparison on attention patterns of last layer between De BERTa and its variants (i.e. De BERTa without EMD, C2P and P2C respectively).

A.10 ACCOUNT FOR THE VARIANCE IN FINE-TUNING

Accounting for the variance of different runs of fine-tuning, in our experiments, we always follow Liu et al. (2019c) to report the results on downstream tasks by averaging over five runs with different random initialization seeds, and perform significance test when comparing results. As the examples shown in Table 14, De BERTabase significantly outperforms Ro BERTabase (p-value < 0.05).

Model MNLI-matched (Min/Max/Avg) SQu AD v1.1 (Min/Max/Avg) p-value Ro BERTabase 84.7/85.0/84.9 90.8/91.3/91.1 0.02 De BERTabase 86.1/86.5/86.3 91.8/92.2/92.1 0.01

Table 14: Comparison of De BERTa and Ro BERTa on MNLI-matched and SQu AD v1.1.