SpanBERT: Improving Pre-training by Representing and Predicting Spans

07/24/2019 ∙ by Mandar Joshi, et al. ∙ Princeton University Facebook University of Washington 0

We present SpanBERT, a pre-training method that is designed to better represent and predict spans of text. Our approach extends BERT by (1) masking contiguous random spans, rather than random tokens, and (2) training the span boundary representations to predict the entire content of the masked span, without relying on the individual token representations within it. SpanBERT consistently outperforms BERT and our better-tuned baselines, with substantial gains on span selection tasks such as question answering and coreference resolution. In particular, with the same training data and model size as BERT-large, our single model obtains 94.6 respectively. We also achieve a new state of the art on the OntoNotes coreference resolution task (79.6 benchmark (70.8



There are no comments yet.


page 1

page 2

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1 Introduction

Pre-training methods like BERT  Devlin et al. (2019) have shown strong performance gains using self-supervised training that masks individual words or subword units. However, many NLP tasks involve reasoning about relationships between two or more spans of text. For example, in extractive question answering Rajpurkar et al. (2016), determining that the “Denver Broncos” is a type of “NFL team” is critical for answering the question “Which NFL team won Super Bowl 50?” Such spans provide a more challenging target for self supervision tasks, for example predicting “Denver Broncos” is much harder than predicting only “Denver” when you know the next word is “Broncos”. In this paper, we introduce a span-level pretraining approach that consistently outperforms BERT, with the largest gains on span selection tasks such as question answering and coreference resolution.

We present SpanBERT, a pre-training method that is designed to better represent and predict spans of text. Our method differs from BERT in both the masking scheme and the training objectives. First, we mask random contiguous spans, rather than random individual tokens. Second, we introduce a novel span-boundary objective (SBO) to train the model to predict the entire masked span from the observed tokens at its boundary. Span-based masking forces the model to predict entire spans solely using the context in which they appear. Furthermore, the span-boundary objective encourages the model to store this span-level information at the boundary tokens, which can be easily accessed during fine tuning. Figure 1 illustrates our approach.

Figure 1: An illustration of SpanBERT training. In this example, the span an American football game is masked. The span boundary objective then uses the boundary tokens was and to to predict each token in the masked span.

To implement SpanBERT, we build on a well-tuned replica of BERT, which already outperforms the original BERT. While building on our baseline, we find that pre-training on single segments, instead of two half-length segments with the next sentence prediction (NSP) objective, significantly improves performance on most downstream tasks. Therefore, we add our modifications on top of the tuned single-sequence BERT baseline.

Together, our pre-training process yields models that outperform all BERT baselines on a wide variety of tasks, and reach substantially better performance on span selection tasks in particular. Specifically, our method reaches 94.6% and 88.7% F1 on SQuAD 1.1 and 2.0 Rajpurkar et al. (2016, 2018), respectively. We also observe similar gains on five additional extractive question answering benchmarks (NewsQA, TriviaQA, SearchQA, HotpotQA, and Natural Questions).111We use the modified MRQA version of these datasets. See more details in Section 4.1.

SpanBERT also arrives at a new state of the art on the challenging CoNLL-2012 (“OntoNotes”) shared task for document-level coreference resolution, where we reach 79.6% F1, exceeding the previous top model by 6.6% absolute. Finally, we demonstrate that SpanBERT also helps on tasks that do not explicitly involve span selection, and show that our approach even improves performance on TACRED Zhang et al. (2017) and GLUE Wang et al. (2019).

While others show the benefits of adding more data Yang et al. (2019) and increasing model size Lample and Conneau (2019), this work demonstrates the importance of designing good pre-training tasks and objectives, which can also have a significant impact.

2 Background: BERT

BERT Devlin et al. (2019) is a self-supervised approach for pre-training a deep transformer encoder Vaswani et al. (2017), before fine-tuning it for a particular downstream task. BERT optimizes two training objectives – masked language modeling (MLM) and next sentence prediction (NSP) – which only require a large collection of unlabeled text.


Given a sequence of word or sub-word tokens

, BERT trains an encoder that produces a contextualized vector representation for each token:

. Since the encoder is implemented via a deep transformer, it uses positional embeddings to mark the absolute position of each token in the sequence.

Masked Language Modeling (MLM)

Also known as a cloze test, MLM is the task of predicting missing tokens in a sequence from their placeholders. Specifically, a subset of tokens is sampled and substituted with a different set of tokens. In BERT’s implementation, accounts for 15% of the tokens in ; of those, 80% are replaced with [MASK], 10% are replaced with a random token (according to the unigram distribution), and 10% are kept unchanged. The task is to predict the original tokens in from the modified input.

BERT selects each token in independently by randomly selecting a subset. In SpanBERT, we define by randomly selecting contiguous spans (Section 3.1).

Next Sentence Prediction (NSP)

The NSP task takes two sequences as input, and predicts whether is the direct continuation of . This is implemented in BERT by first reading from the corpus, and then (1) either reading from the point where ended, or (2) randomly sampling from a different point in the corpus. The two sequences are separated by a special [SEP]token. Additionally, a special [CLS]token is added to to form the input, where the target of [CLS]is whether indeed follows in the corpus.

In SpanBERT, we remove the NSP objective and sample a single full-length sequence (Section 3.3).

3 Model

We present SpanBERT, a self-supervised pre-training method designed to better represent and predict spans of text. Our approach is inspired by BERT Devlin et al. (2019), but deviates from its bi-text classification framework in three ways. First, we use a different random process to mask spans of tokens, rather than individual ones. We also introduce a novel auxiliary objective – the span boundary objective (SBO) – which tries to predict the entire masked span using only the representations of the tokens at the span’s boundary. Finally, SpanBERT samples a single contiguous segment of text for each training example (instead of two), and thus has no use for BERT’s next sentence prediction objective, which we omit.

3.1 Span Masking

Given a sequence of tokens , we select a subset of tokens by iteratively sampling spans of text until the masking budget (e.g. 15% of

) has been spent. At each iteration, we first sample the span’s length from a geometric distribution

, which is skewed towards shorter spans. We then randomly (uniformly) select the starting point for the span.

Following preliminary trials, we set , and also clip at . This yields a mean span length of . We also measure span length in complete words, not subword tokens, making the masked spans even longer. Figure 2 shows the distribution of span mask lengths.

As in BERT, we also mask 15% of the tokens in total: replacing 80% of the masked tokens with [MASK], 10% with random tokens and 10% with the original tokens. However, we perform this replacement at the span level and not for each token individually; i.e. all the tokens in a span are replaced with [MASK]or sampled tokens.

Figure 2: We sample random span lengths from a geometric distribution clipped at .

3.2 Span Boundary Objective (SBO)

Span selection models Lee et al. (2016, 2017); He et al. (2018) typically create a fixed-length representation of a span using its boundary tokens (start and end). To support such models, we would ideally like the representations for the end of the span to summarize as much of the internal span content as possible. We do so by introducing a span boundary objective that involves predicting each token of a masked span using only the representations of the observed tokens at the boundaries (Figure 1).

Formally, given a masked span , where indicates its start and end positions, we represent each token in the span using the encodings of the external boundary tokens and , as well as the positional embedding of the target token :

In this work, we implement the representation function as a 2-layer feed-forward network with GeLU activations Hendrycks and Gimpel (2016) and layer normalization Ba et al. (2016):

We then use the vector representation to predict and compute the cross-entropy loss exactly like the MLM objective.

SpanBERT sums the loss from both the span boundary and the regular masked language modeling objectives for each token in the masked span.

3.3 Single-Sequence Training

As described in Section 2, BERT’s examples contain two sequences of text , and an objective that trains the model to predict whether they are connected (NSP). We find that this setting is almost always worse than simply using a single sequence without the NSP objective (see Section 4.3 for further details). We conjecture that single-sequence training is superior to bi-sequence training with NSP because (a) the model benefits from longer full-length contexts, or (b) conditioning on context from another document adds noise to the masked language model. Therefore, in our approach, we remove both the NSP objective and the two-segment sampling procedure, and simply sample a single contiguous segment of up to tokens, rather than two half-segments that sum up to tokens together.

4 Experimental Setup

4.1 Tasks

We evaluate on a comprehensive suite of tasks, including seven question answering tasks, coreference resolution, nine tasks in the GLUE benchmark Wang et al. (2019), and relation extraction. We expect that the span selection tasks, question answering and coreference resolution, will particularly benefit from our span-based pre-training.

Extractive Question Answering

Given a short passage of text and a question as input, the task of extractive question answering is to select a contiguous span of text in the passage as the answer.

We first evaluate on SQuAD 1.1 and 2.0  Rajpurkar et al. (2016, 2018), which have served as major question answering benchmarks, particularly for pre-trained models  Peters et al. (2018); Devlin et al. (2019); Yang et al. (2019). We also evaluate on five more datasets from the MRQA shared task:222 MRQA changed the original datasets to unify them into the same format, e.g. all the contexts are truncated to a maximum of 800 tokens and only answerable questions are kept. NewsQA Trischler et al. (2017), SearchQA Dunn et al. (2017), TriviaQA Joshi et al. (2017), HotpotQA Yang et al. (2018) and Natural Questions (NaturalQA) Kwiatkowski et al. (2019). Because the MRQA shared task does not have a public test set, we split the development set in half to make new development and test sets. The datasets vary in both domain and collection methodology, making this collection a good testbed for evaluating whether our pre-trained models can generalize well across different data distributions.

Following BERT Devlin et al. (2019), we use the same QA model for all the datasets. We first convert the passage and question into a single sequence

, pass it to the pre-trained transformer encoder, and train two linear classifiers independently on top of it for predicting the answer span boundary (start and end). For the unanswerable questions in SQuAD 2.0, we simply set the answer span to be the special token

[CLS]for both training and testing.

Coreference Resolution

Coreference resolution is the task of clustering mentions in text which refer to the same real-world entities. We evaluate on the CoNLL-2012 shared task Pradhan et al. (2012) for document-level coreference resolution. The model is an augmentation of Lee et al.’s (2018) higher-order coreference model, which replaces the original LSTM-based encoders with BERT’s pre-trained transformer encoders.

Relation Extraction

TACRED Zhang et al. (2017) is a challenging relation extraction dataset. Given one sentence and two spans within it – subject and object – the task is to predict the relation between the spans from 42 pre-defined relation types, including no_relation. We follow the entity masking schema from Zhang et al. (2017) and replace the subject and object entities by their NER tags such as “[CLS][SUBJ-PER] was born in [OBJ-LOC] , Michigan, …”, and finally add a linear classifier on top of the [CLS]token to predict the relation type.


The General Language Understanding Evaluation (GLUE) benchmark Wang et al. (2019) consists of 9 sentence-level classification tasks: 2 single-sentence tasks: CoLA Warstadt et al. (2018), SST-2 Socher et al. (2013), 3 sentence similarity tasks: MRPC Dolan and Brockett (2005), STS-B Cer et al. (2017), QQP,333 and 4 natural language inference tasks: MNLI Williams et al. (2018), QNLI Rajpurkar et al. (2016), RTE Dagan et al. (2005); Bar-Haim et al. (2006); Giampiccolo et al. (2007) and WNLI Levesque et al. (2011). While recent work Liu et al. (2019a) has applied several task-specific strategies to increase performance on the individual GLUE tasks, we follow BERT’s single-task setting and add a linear classifier on top of the [CLS]token.

4.2 Implementation

We reimplemented BERT’s model and pre-training method in fairseq Ott et al. (2019). We used the model configuration of BERT-large as in devlin2018bert and also trained all our models on the same corpus: BooksCorpus and English Wikipedia using cased word-piece tokens.

The main difference in our implementation is that we use different masks at each epoch while BERT samples 10 different masks for each sequence during data processing. Additionally, the original BERT implementation samples shorter sequences with a small probability (0.1) while we always take sequences of up to 512 tokens until it reaches a document boundary.

444We refer the reader to RoBERTa Liu et al. (2019b) for further discussion on these modifications and their effects.

We also deviate from the optimization by running for 2.4M steps and using an epsilon of 1e-8 for Adam Kingma and Ba (2015), which converges to a better set of model parameters. The pre-training was done on 32 Volta V100 GPUs, and took 15 days to complete.

Fine-tuning is implemented based on HuggingFace’s codebase.555 Appendix A has more details.

4.3 Baselines

We compare SpanBERT to three baselines:

Google BERT

The pre-trained models released by devlin2018bert.666


Our reimplementation of BERT with improved data preprocessing and optimization (Section 4.2).

Our BERT-1seq

Our reimplementation of BERT trained on single full-length sequences without NSP (Section 3.3).

SQuAD 1.1 SQuAD 2.0
Human Perf. 82.3 91.2 86.8 89.4
Google BERT 84.3 91.3 80.0 83.3
Our BERT 86.5 92.6 82.8 85.9
Our BERT-1seq 87.5 93.3 83.8 86.6
SpanBERT 88.8 94.6 85.7 88.7
Table 1: Test results on SQuAD 1.1 and SQuAD 2.0.
NewsQA TriviaQA SearchQA HotpotQA NaturalQA (Avg)
Google BERT 68.8 77.5 81.7 78.3 79.9 77.3
Our BERT 71.0 79.0 81.8 80.5 80.5 78.6
Our BERT-1seq 71.9 80.4 84.0 80.3 81.8 79.7
SpanBERT 73.6 83.6 84.8 83.0 82.5 81.5
Table 2: Performance (F1) on the five MRQA extractive question answering tasks.
P R F1 P R F1 P R F1 Avg. F1
Prev. SotA: Lee et al. (2018) 81.4 79.5 80.4 72.2 69.5 70.8 68.2 67.1 67.6 73.0
Google BERT 84.9 82.5 83.7 76.7 74.2 75.4 74.6 70.1 72.3 77.1
Our BERT 85.1 83.5 84.3 77.3 75.5 76.4 75.0 71.9 73.9 78.3
Our BERT-1seq 85.5 84.1 84.8 77.8 76.7 77.2 75.3 73.5 74.4 78.8
SpanBERT 85.8 84.8 85.3 78.3 77.9 78.1 76.4 74.2 75.3 79.6
Table 3: Performance on the OntoNotes coreference resolution benchmark. The main evaluation is the average F1 of three metrics – MUC, , and on the test set.
Google BERT 59.3 95.2 88.5/84.3 86.4/88.0 71.2/89.0 86.1/85.7 93.0 71.1 80.4
Our BERT 58.6 93.9 90.1/86.6 88.4/89.1 71.8/89.3 87.2/86.6 93.0 74.7 81.1
Our BERT-1seq 63.5 94.8 91.2/87.8 89.0/88.4 72.1/89.5 88.0/87.4 93.0 72.1 81.7
SpanBERT 64.3 94.8 90.9/87.9 89.9/89.1 71.9/89.5 88.1/87.7 94.3 79.0 82.8
Table 4: Test set performance metrics on GLUE tasks. MRPC: F1/accuracy, STS-B: Pearson/Spearmanr correlation, QQP: F1/accuracy, MNLI: matched/mistached accuracies. WNLI (not shown) is always set to majority class (65.1% accuracy) and included in the average.
P R F1
Curr. SotA: Soares et al. (2019) - - 71.5
Google BERT 69.1 63.9 66.4
Our BERT 67.8 67.2 67.5
Our BERT-1seq 72.4 67.9 70.1
SpanBERT 70.8 70.9 70.8
Table 5: Test set performance on the TACRED relation extraction benchmark.

5 Results

We compare SpanBERT to the baselines per task, and draw conclusions based on the overall trends.

5.1 Per-Task Results

Extractive Question Answering

Table 1 shows the performance on both SQuAD 1.1 and 2.0. SpanBERT exceeds our BERT baseline by 2.0% (SQuAD 1.1) and 2.8% (SQuAD 2.0) F1. In SQuAD 1.1, this result accounts for over 27% error reduction, reaching 3.4% F1 above human performance.

Table 2 demonstrates that this trend goes beyond SQuAD, and is consistent in every MRQA dataset. On average, we see a 2.9% F1 improvement from our reimplementation of BERT. Although some gains are coming from single-sequence training (+1.1%), most of the improvement stems from span masking and the span boundary objective (+1.8%), with particularly large gains on TriviaQA (+3.2%) and HotpotQA (+2.7%).

Coreference Resolution

Table 3 shows the performance on the OntoNotes coreference resolution benchmark. Our BERT reimplementation improves the Google BERT model by 1.2% on the average F1 metric and single-sequence training brings another 0.5% gain. Finally, SpanBERT significantly improves on top of that, achieving a new state of the art of 79.6% F1 (previous best result is 73.0%).

Relation Extraction

Table 5 shows the performance on TACRED. SpanBERT achieves close to the current state of the art  Soares et al. (2019), and exceeds our reimplementation of BERT by 3.3% F1. Most of this gain (+2.6%) stems from single-sequence training although the contribution of span masking and the span boundary objective is still significant (+0.7%), resulting largely from higher recall. In addition to a different input encoding scheme, the current state of the art  Soares et al. (2019) pre-trains relation representations using distant supervision from entity-linked text.


Table 4 shows the performance on GLUE. For most tasks, the different models appear to perform similarly. Moving to single-sequence training without the NSP objective substantially improves CoLA, and yields smaller (but significant) improvements on MRPC and MNLI. The main gains from SpanBERT are in the SQuAD-based QNLI dataset (+1.3%) and in RTE (+6.9%), the latter accounting for most of the rise in SpanBERT’s GLUE average.

5.2 Overall Trends

We compared our approach to three BERT baselines on 17 benchmarks, and found that SpanBERT outperforms BERT on almost every task. In 14 tasks, SpanBERT performed better than all baselines. In 2 tasks (MRPC and QQP), it performed on-par with single-sequence trained BERT, but still outperformed the other baselines. In 1 task (SST-2), Google’s BERT baseline performed better than SpanBERT by 0.4% accuracy.

When considering the magnitude of the gains, it appears that SpanBERT is especially better at extractive question answering. In SQuAD 1.1, for example, we observe a solid gain of 2.0% F1 even though the baseline is already well above human performance. On MRQA, SpanBERT improves between 2.0% (NaturalQA) and 4.6% (TriviaQA) F1 on top of our BERT baseline.

Finally, we observe that single-sequence training works considerably better than bi-sequence training with next sentence prediction (NSP) for a wide variety of tasks. This is surprising because BERT’s ablations showed gains from the NSP objective Devlin et al. (2019). However, the ablation studies still involved bi-sequence data processing, i.e., the pre-training stage only controlled for the NSP objective while still sampling two half-length sequences.777We confirmed this fact with BERT’s authors. We hypothesize that bi-sequence training, as it is implemented in BERT (see Section 2), impedes the model from learning longer-range features, and consequently hurts performance on many downstream tasks.

6 Ablation Studies

SQuAD 2.0 NewsQA TriviaQA Coreference MNLI-m QNLI
Subword Tokens 83.8 72.0 76.3 77.7 86.7 92.5
Whole Words 84.3 72.8 77.1 76.6 86.3 92.8
Named Entities 84.8 72.7 78.7 75.6 86.0 93.1
Noun Phrases 85.0 73.0 77.7 76.7 86.5 93.2
Random Spans 85.4 73.0 78.8 76.4 87.0 93.3
Table 6: The effect of replacing BERT’s original masking scheme (Subword Tokens) with different masking schemes. Results are F1 scores for QA tasks and accuracy for MNLI and QNLI on the development sets. All the models are based on bi-sequence training with NSP.
SQuAD 2.0 NewsQA TriviaQA Coreference MNLI-m QNLI
Span Masking (2seq) + NSP 85.4 73.0 78.8 76.4 87.0 93.3
Span Masking (1seq) 86.7 73.4 80.0 76.3 87.3 93.8
Span Masking (1seq) + SBO 86.8 74.1 80.3 79.0 87.6 93.9
Table 7: The effects of different auxiliary objectives, given MLM over random spans as the primary objective.

We compare our random span masking scheme with linguistically-informed masking schemes, and find that masking random spans is a competitive and often better approach. We then study the impact of the span boundary objective (SBO), and contrast it with BERT’s next sentence prediction (NSP) objective.888To save time and resources, we use the checkpoints at 1.2M steps for all the ablation experiments.

6.1 Masking Schemes

Previous work  Sun et al. (2019) has shown improvements in downstream task performance by masking linguistically-informed spans during pre-training for Chinese. We compare our random span masking scheme with masking of linguistically-informed spans. Specifically, we train the following five baseline BERT models differing only in the way tokens are masked.

Subword Tokens

We sample random word-piece tokens, as in the original BERT.

Whole Words

We sample random words, and then mask all of the subword tokens in those words. The total number of masked subtokens is around 15%.

Named Entities

At 50% of the time, we sample from named entities in the text, and sample random whole words for the other 50%. The total number of masked subtokens is 15%. Specifically, we run spaCy’s named entity recognizer Honnibal and Montani (2017)999 on the corpus and select all the non-numerical named entity mentions as candidates.

Noun Phrases

Similar as Named Entities, we sample from noun phrases at 50% of the time. The noun phrases are extracted by running spaCy’s constituency parser.

Random Spans

We sample random spans from a geometric distribution, as in our SpanBERT (see Section 3.1).

Table 6 shows how different pre-training masking schemes affect performance on a selection of tasks. With the exception of coreference resolution, masking random spans is preferable to other strategies. Although linguistic masking schemes (named entities and noun phrases) are often competitive with random spans, their performance is not consistent; for instance, masking noun phrases achieves parity with random spans on NewsQA, but underperforms on TriviaQA (-1.1% F1).

On coreference resolution, we see that masking random subword tokens is preferable to any form of span masking. Nevertheless, we shall see in the following experiment that combining random span masking with the span boundary objective can significantly improve upon this result.

6.2 Auxiliary Objectives

In Section 4.3, we saw that bi-sequence training with the next sentence prediction (NSP) objective can hurt performance on downstream tasks, when compared to single-sequence training. We test whether this holds true for models pre-trained with span masking, and also evaluate the effect of replacing the NSP objective with the sentence boundary objective (SBO).

Table 7 confirms that single-sequence training typically improves performance. Adding SBO further improves performance, with a substantial gain on coreference resolution (+2.7% F1) over span masking alone. Unlike the NSP objective, SBO does not appear to have any adverse effects.

7 Related Work

Pre-trained contextualized word representations that can be trained from unlabeled text Dai and Le (2015); Melamud et al. (2016); Peters et al. (2018) have had immense impact on NLP lately, particularly as methods for initializing a large model before fine-tuning it for a specific task Howard and Ruder (2018); Radford et al. (2018); Devlin et al. (2019)

. Beyond differences in model hyperparameters and corpora, these methods mainly differ in their pre-training tasks and loss functions, with a considerable amount of contemporary literature proposing augmentations of BERT’s masked language modeling (MLM) objective.

While previous and concurrent work has looked at masking Sun et al. (2019) or dropping Song et al. (2019); Chan et al. (2019) multiple words from the input – particularly as pretraining for language generation tasks – SpanBERT pretrains span representations Lee et al. (2016), which are widely used for question answering, coreference resolution and a variety of other tasks. ERNIE Sun et al. (2019) shows improvements on Chinese NLP tasks using phrase and named entity masking. MASS Song et al. (2019) focuses on language generation tasks, and adopts the encoder-decoder framework to reconstruct a sentence fragment given the remaining part of the sentence. We attempt to more explicitly model spans using the SBO objective, and show that (geometrically distributed) random span masking works as well, and sometimes better than, masking linguistically-coherent spans. We evaluate on English benchmarks for question answering, relation extraction, and coreference resolution in addition to GLUE.

A different ERNIE Zhang et al. (2019) focuses on integrating structured knowledge bases with contextualized representations with an eye on knowledge-driven tasks like entity typing and relation classification. UNILM Dong et al. (2019) uses multiple language modeling objectives – unidirectional (both left-to-right and right-to-left), bidirectional, and sequence-to-sequence prediction – to aid generation tasks like summarization and question generation. XLM Lample and Conneau (2019) explores cross-lingual pre-training for multilingual tasks such as translation and cross-lingual classification. Kermit Chan et al. (2019), an insertion based approach, fills in missing tokens (instead of predicting masked ones) during pretraining; they show improvements on machine translation and zero-shot question answering.

Concurrent with our work, RoBERTa  Liu et al. (2019b) presents a replication study of BERT pre-training that measures the impact of many key hyperparameters and training data size. Also concurrent, XLNet Yang et al. (2019) combines an autoregressive loss and the Transformer-XL Dai et al. (2019) architecture with a more than an eight-fold increase in data to achieve current state-of-the-art results on multiple benchmarks. XLNet also masks spans (of 1-5 tokens) during pre-training, but predicts them autoregressively. Our model focuses on incorporating span-based pre-training, and as a side effect, we present a stronger BERT baseline while controlling for the corpus, architecture, and the number of parameters.

Related to our SBO objective, pair2vec Joshi et al. (2019) encodes word-pair relations using a negative sampling-based multivariate objective during pre-training. Later, the word-pair representations are injected into the attention-layer of downstream tasks, and thus encode limited downstream context. Unlike pair2vec, our SBO objective yields “pair” (start and end tokens of spans) representations which more fully encode the context during both pre-training and finetuning, and are thus more appropriately viewed as span representations. Stern et al. (2018) focus on improving language generation speed using a block-wise parallel decoding scheme; they make predictions for multiple time steps in parallel and then back off to the longest prefix validated by a scoring model. Also related are sentence representation methods Kiros et al. (2015); Logeswaran and Lee (2018) which focus on predicting surrounding contexts from sentence embeddings.

8 Conclusion

We presented a new method for span-based pre-training which extends BERT by (1) masking contiguous random spans, rather than random tokens, and (2) training the span boundary representations to predict the entire content of the masked span, without relying on the individual token representations within it. Together, our pre-training process yields models that outperform all BERT baselines on a variety of tasks, and reach substantially better performance on span selection tasks in particular.


We would like to thank Pranav Rajpurkar and Robin Jia for patiently helping us evaluate SpanBERT on SQuAD. We thank our colleagues at Facebook AI Research and the University of Washington for their insightful feedback.



A Fine-tuning Hyperparameters

We applied the following fine-tuning hyperparameters to all methods, including the baselines.

Extractive Question Answering

For all the question answering tasks, we use max_seq_length = 512 and a sliding window of size if the lengths are longer than 512. We choose learning rates from {5e-6, 1e-5, 2e-5, 3e-5, 5e-5} and batch sizes from {16, 32} and fine-tune 4 epochs for all the datasets.

Coreference Resolution

We divide the documents into multiple chunks of lengths up to max_seq_length and encode each chunk independently. We choose max_seq_length from {128, 256, 384, 512}, BERT learning rates from {1e-5, 2e-5}, task-specific learning rates from {1e-4, 2e-4, 3e-4} and fine-tune 20 epochs for all the datasets. We use batch size (one document) for all the experiments.

GLUE & Relation Extraction

We use max_seq_length = 128 and choose learning rates from {5e-6, 1e-5, 2e-5, 3e-5, 5e-5} and batch sizes from {16, 32} and fine-tuning 10 epochs for all the datasets. The only exception is CoLA, where we used 4 epochs (following devlin2018bert), because 10 epochs lead to severe overfitting.