English Intermediate-Task Training Improves Zero-Shot Cross-Lingual Transfer Too

05/26/2020 ∙ by Jason Phang, et al. ∙ NYU college 0

Intermediate-task training has been shown to substantially improve pretrained model performance on many language understanding tasks, at least in monolingual English settings. Here, we investigate whether English intermediate-task training is still helpful on non-English target tasks in a zero-shot cross-lingual setting. Using a set of 7 intermediate language understanding tasks, we evaluate intermediate-task transfer in a zero-shot cross-lingual setting on 9 target tasks from the XTREME benchmark. Intermediate-task training yields large improvements on the BUCC and Tatoeba tasks that use model representations directly without training, and moderate improvements on question-answering target tasks. Using SQuAD for intermediate training achieves the best results across target tasks, with an average improvement of 8.4 points on development sets. Selecting the best intermediate task model for each target task, we obtain a 6.1 point improvement over XLM-R Large on the XTREME benchmark, setting a new state of the art. Finally, we show that neither multi-task intermediate-task training nor continuing multilingual MLM during intermediate-task training offer significant improvements.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 2

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

Zero-shot cross-lingual transfer involves training a language-encoding model on task data in one language, and evaluating the tuned model on the same task in other languages. This format evaluates the extent to which task-specific knowledge learned in one language generalizes across languages. Transformer models such as mBERT (Devlin et al., 2019a), XLM (Conneau and Lample, 2019) and XLM-R (Conneau et al., 2019) that have been pretrained with a masked language modeling (MLM) objective on large corpora of multilingual data have shown remarkably strong results on zero-shot cross-lingual transfer, and show promise as a way of facilitating the construction of massively multilingual language technologies.

Intermediate-task training (STILTs; Phang et al., 2018) is the simple strategy of fine-tuning a pretrained model on a data-rich intermediate task, ideally related to the target task, before fine-tuning a second time on the downstream target task. Despite its simplicity, this two-phase training setup has been shown to be beneficial across a range of Transformer models and tasks (Wang et al., 2019a; Pruksachatkun et al., 2020), at least with English intermediate and target tasks.

Figure 1: We investigate the benefit of injecting an additional phase of intermediate-task training on English language task data. We also consider variants using multi-task intermediate-task training, as well as continuing multilingual MLM during intermediate-task training. Best viewed in color.

In this work, we investigate whether intermediate training on English language tasks can also improve performance in a zero-shot cross-lingual transfer setting. Starting with a pretrained multilingual language encoder, we perform intermediate-task training on one or more English language tasks, then fine-tune on the target task in English, and finally evaluate zero-shot on the same task in other languages.

Intermediate-task training on English data introduces a potential issue: we train the pretrained multilingual model extensively on only English data before attempting to use it on non-English target task data, leaving open the possibility that the model will lose the knowledge of other languages that it acquired during pretraining (Kirkpatrick et al., 2017; Yogatama et al., 2019). To attempt to mitigate this, we experiment with mixing in multilingual MLM training updates during the intermediate-task training.

Concretely, we use the pretrained XLM-R (Conneau et al., 2019) as our multilingual language encoder as it currently achieves state-of-the-art performance on many zero-shot cross-lingual transfer tasks. We perform experiments on 9 target tasks from the recently introduced XTREME benchmark Hu et al. (2020), which aims to evaluate zero-shot cross-lingual performance across diverse target tasks across up to 40 languages each. We investigate how intermediate-task training on 7 different tasks, including question answering, sentence tagging/completion, paraphrase detection, and natural language inference, impacts zero-shot cross-lingual transfer performance. We find:

  • Applying intermediate-task training to BUCC and Tatoeba, the two sentence retrieval target tasks that have no training data, yields dramatic improvements with almost every intermediate training configuration.

  • The question-answering target tasks show consistent smaller improvement with many intermediate tasks.

  • Evaluating our best performing models for each target task on the XTREME benchmark yields an average improvement of 6.1 points, setting the state of the art as of writing.

  • Neither continuing multilingual MLM training during intermediate-task training nor multi-task training meaningfully bolster transfer performance.

2 Approach

We propose a three-phase approach to training, illustrated in Figure 1: (i) we use a publicly available model pretrained on raw multilingual text using MLM; (ii) we perform intermediate-task training on one or more English intermediate tasks; (iii) we fine-tune the model on English target task data, before evaluating it on task data in multiple languages.

During phase (ii), all our intermediate tasks have English labeled data only. We experiment with performing intermediate-task training on single tasks individually and as well as a multi-task format. Intermediate tasks are described in detail in Section 2.1. We use target tasks from the recent XTREME benchmark, whose goal is to evaluate zero-shot cross-lingual transfer, i.e. training on English target-task data and evaluating the model on target-task data in different languages. Target tasks are described in Section 2.2.

2.1 Intermediate Tasks

We study the effect of intermediate-task training (STILTs; Phang et al., 2018) on zero-shot cross-lingual transfer into multiple target tasks and languages. We experiment with seven different intermediate tasks, all of them with English labeled data, as illustrated in Table 1.

Name Train Dev Test Task Genre/Source
Intermediate tasks
ANLI 1,104,934 22,857 N/A natural language inference Misc.
QQP 363,846 40,430 N/A paraphrase detection Quora questions
SQuAD 87,599 34,726 N/A span extraction Wikipedia
HellaSwag 39,905 10,042 N/A sentence completion Video captions & Wikihow
CCG 38,015 5,484 N/A tagging Wall Street Journal
Cosmos QA 25,588 3,000 N/A question answering Blogs
CommonsenseQA 9,741 1,221 N/A question answering Crowdsourced responses
Target tasks (XTREME Benchmark)
XNLI 392,702 2,490 5,010 natural language inference Misc.
PAWS-X 49,401 2,000 2,000 paraphrase detection Wiki/Quora
POS 21,253 3,974 47–20,436 tagging Misc.
NER 20,000 10,000 1,000–10,000 named entity recognition Wikipedia
XQuAD 87,599 34,726 1,190 question answering Wikipedia
MLQA 87,599 34,726 4,517–11,590 question answering Wikipedia
TyDiQA-GoldP 3,696 634 323–2,719 question answering Wikipedia
BUCC 1,896–14,330 sentence retrieval Wiki / news
Tatoeba 1,000 sentence retrieval Misc.
Table 1: Overview of the intermediate tasks (top) and target tasks (bottom) in our experiments. EM is short for Exact Match. For target tasks, Train and Dev correspond to the English training and development sets, while Test shows the range of sizes for the target-language test sets for each task. XQuAD, TyDiQA and BUCC do not have separate held-out development sets.

Anli + Mnli + Snli (Anli)

The Adversarial Natural Language Inference dataset (Nie et al., 2019) is collected adversarially using a human and model in the loop as an extensions of the Stanford Natural Language Inference (SNLI; Bowman et al., 2015) and the Multi-genre Natural Language Inference (MNLI; Williams et al., 2018) corpora. We follow Nie et al. (2019) and use the concatenated ANLI, MNLI and SNLI training sets, which we refer to as ANLI.

Ccg

CCGbank (Hockenmaier and Steedman, 2007) is a translation of the Penn Treebank into Combinatory Categorial Grammar (CCG) derivations. The CCG supertagging task that we use consists of assigning lexical categories to individual word tokens which roughly determine a full parse.111If a word is tokenized into sub-word tokens, we use the representation of the first sub-word token for the tag prediction for that word.

CommonsenseQA

CommonsenseQA Talmor et al. (2019) is a multiple-choice QA dataset generated by crowdworkers based on clusters of concepts from ConceptNet (Speer et al., 2017).

Cosmos QA

Cosmos QA is multiple-choice commonsense-based reading comprehension dataset (Huang et al., 2019b) generated by crowdworkers, with a focus on the causes and effects of events.

HellaSwag

HellaSwag (Zellers et al., 2019) is a commonsense reasoning dataset framed as a four-way multiple choice task, where examples consist of a text and four choices of spans, one of which is a plausible continuation of the given scenario. It is built using adversarial filtering (Zellers et al., 2018; Bras et al., 2020) with BERT.

Qqp

The Quora Question Pairs222http://data.quora.com/First-Quora-DatasetRelease-Question-Pairs is a paraphrase detection dataset constructed from questions posted on Quora, a community question-answering website. Examples in the dataset consist of two questions, labeled for whether they are semantically equivalent.

SQuAD

Stanford Question Answering Dataset (Rajpurkar et al., 2016) is a question-answering dataset consisting of passages extracted from Wikipedia articles and crowd-sourced questions and answers. Each example consists of a context passage and a question, and the answer is a text span from the context. We use SQuAD version 1.1.

Task Selection Criteria

We choose these tasks based on both the diversity of task formats and evidence of positive transfer from literature. Pruksachatkun et al. (2020) shows that MNLI (which is a subset ANLI), CommonsenseQA, Cosmos QA and HellaSwag are good candidates for intermediate tasks with positive transfer to a range of downstream tasks. CCG involves token-wise prediction and is similar to the POS and NER target tasks. SQuAD is a widely-used question-answering task, while QQP is semantically similar to sentence retrieval target tasks (BUCC and Tatoeba) as well as PAWS-X, another paraphrase-detection task.

2.2 Target Tasks

We use the 9 target tasks from the XTREME benchmark Hu et al. (2020), which span 40 different languages (hereafter referred to as the target languages): The Cross-lingual Question Answering (XQuAD; Artetxe et al., 2019) dataset, a cross-lingual extension of the SQuAD dataset Rajpurkar et al. (2016); the Multilingual Question Answering (MLQA; Lewis et al., 2019) dataset; the Typologically Diverse Question Answering (TyDiQA-GoldP; Clark et al., 2020) dataset; the Cross-lingual Natural Language Inference dataset (XNLI; Conneau et al., 2018), a cross-lingual extension of MNLI Williams et al. (2018); the Cross-lingual Paraphrase Adversaries from Word Scrambling (PAWS-X; Yang et al., 2019) dataset; the Universal Dependencies v2.5 (Nivre et al., 2018) POS tagging dataset; the Wikiann NER dataset Pan et al. (2017); the BUCC dataset Zweigenbaum et al. (2017, 2018), which involves identifying parallel sentences from corpora of different languages; and the Tatoeba Artetxe and Schwenk (2019) dataset, which involves aligning pairs of sentences.

Among the 9 tasks, BUCC and Tatoeba are sentence retrieval tasks do not include training sets, and are scored based on the similarity of learned representations (see Appendix). XQuAD, TyDiQA and Tatoeba do not include development sets separate from the test sets.333UDPOS also does not include development set data for a small number of languages: Kazakh, Thai, Tagalog and Yoruba. For all XTREME tasks, we follow the training and evaluation protocol described in the benchmark (Hu et al., 2020) and their sample implementation.444https://github.com/google-research/xtreme Intermediate and target task statistics are shown in Table 1.

2.3 Multilingual Masked Language Modeling

Our setup requires that we train the pretrained multilingual model extensively on English data before using it on a non-English target task, which can lead to the catastrophic forgetting of other languages acquired during pretraining. We investigate whether continuing to train on the multilingual MLM pretraining objective while fine-tuning on an English intermediate task can prevent catastrophic forgetting of the target languages and improve downstream transfer performance.

We construct a multilingual corpus across the 40 languages covered by the XTREME benchmark using Wikipedia dumps from April 14, 2020 for each language and the MLM data creation scripts from the jiant library (Wang et al., 2019c). In total, we use 2 million sentences sampled across all languages according to the sampling ratio of Conneau and Lample (2019) with the .

3 Experiments and Results

3.1 Models

We use the pretrained XLM-R Large model (Conneau et al., 2019) as a starting point for all our experiments.555XLM-R Large (Conneau et al., 2019) is a 550m-parameter variant of the RoBERTa masked language model (Liu et al., 2019b) trained on a cleaned version of CommonCrawl on 100 languages. Details on the intermediate and target task training procedures can be found in the Appendix.

Xlm-R

As our baseline, we directly fine-tune the pretrained XLM-R model on each target task’s English training data (if available) and evaluate zero-shot on non-English data, following the protocol for the XTREME benchmark (Hu et al., 2020).

XLM-R + Intermediate-Task Training

We include an additional intermediate-task training phase, and first fine-tune the pretrained XLM-R on an English intermediate task before we train/evaluate on the target tasks as described above.

We also experiment with multi-task training on all available intermediate tasks but SQuAD.666Excluded in this draft due to implementation issues. We follow Raffel et al. (2019a)

and sample batches of examples for each task with probability

, where is the number of examples in task and is a size limit constant.

Target tasks
XNLI PAWS-X POS NER XQuAD MLQA TyDiQA BUCC Tatoeba Avg.
Metric acc. acc. F1 F1 F1 / EM F1 / EM F1 / EM F1 acc.
# langs. 15 7 33 40 11 7 9 5 37
XLM-R 80.1 86.5 75.7 62.8 76.1 / 60.0 70.1 / 51.5 75.7 / 61.0 71.5 31.0 67.2

No MLM

ANLI - 0.8 + 0.4 - 0.9 - 0.8 - 0.6 / - 0.1 - 0.6 / - 0.8 + 2.2 / + 3.1 +20.1 +49.8 + 7.7
QQP - 1.4 - 2.1 - 5.6 - 6.9 - 3.8 / - 3.8 - 3.9 / - 4.4 - 0.6 / - 0.2 +20.2 +51.7 + 5.3
SQuAD - 1.4 + 0.7 - 1.6 + 0.2 + 1.1 / + 1.3 + 1.9 / + 2.5 + 5.6 / + 7.4 +19.7 +46.9 + 8.3
HellaSwag - 0.3 + 0.8 - 0.7 - 1.0 - 0.3 / + 0.1 - 0.1 / + 0.2 + 1.9 / + 1.3 +20.4 +49.9 + 7.9
CCG - 2.6 - 3.4 - 1.5 - 0.7 - 1.5 / - 1.3 - 1.6 / - 1.5 + 0.4 / + 0.7 + 5.5 +38.9 + 3.7
CosmosQA - 2.9 + 1.5 - 1.2 - 0.9 + 0.2 / + 0.3 + 0.4 / + 0.5 + 2.7 / + 3.8 +13.2 +28.8 + 4.7
CSQA - 2.9 - 0.6 - 1.7 - 0.5 + 0.2 / + 0.4 + 1.6 / + 1.6 + 3.0 / + 4.1 +11.3 +33.1 + 4.9
Multi-task - 1.6 - 0.2 - 2.3 - 2.4 - 2.6 / - 3.1 - 1.4 / - 1.7 + 1.9 / + 1.9 +18.4 +48.3 + 6.4

With MLM

ANLI - 0.1 + 0.5 + 0.4 - 0.4 - 0.2 / - 0.4 - 0.2 / - 0.3 + 2.5 / + 3.3 +17.8 +44.7 + 7.3
QQP - 0.5 + 0.6 - 0.2 + 0.7 - 0.3 / - 0.2 + 0.0 / + 0.2 + 2.9 / + 3.4 +18.0 +42.3 + 5.1
SQuAD - 4.1 + 0.3 - 2.0 - 1.3 + 0.8 / + 1.2 + 0.9 / + 1.1 + 5.0 / + 6.5 +12.8 +23.5 + 4.1
HellaSwag + 0.7 + 0.0 - 0.7 - 1.2 - 0.8 / - 0.8 - 1.0 / - 1.4 + 2.6 / + 3.5 + 8.9 + 6.9 + 1.7
CCG - 1.0 - 0.6 - 0.6 - 1.9 - 1.2 / - 1.3 - 1.6 / - 0.8 + 2.2 / + 2.7 -10.1 +22.2 + 0.9
CosmosQA - 0.3 + 0.4 - 1.2 - 1.5 - 0.6 / - 0.3 - 0.4 / - 0.3 + 2.2 / + 2.0 +18.2 +42.7 + 6.6
CSQA + 0.1 + 0.5 - 1.4 + 0.1 + 0.1 / + 0.1 + 0.8 / + 0.6 + 2.8 / + 3.1 +11.6 +25.9 + 4.5
Multi-task + 0.4 + 0.6 - 2.1 - 1.3 - 0.8 / - 1.0 - 0.6 / - 0.4 + 2.9 / + 3.8 +16.4 +45.6 + 6.8
XTREME Benchmark Scores
XLM-R (Hu et al., 2020) 79.2 86.4 72.6 65.4 76.6 / 60.8 71.6 / 53.2 65.1 / 45.0 66.0 57.3 68.1
XLM-R (Ours) 79.5 86.2 74.0 62.6 76.1 / 60.0 70.2 / 51.2 75.5 / 61.0 64.5 31.0 66.1
Our Best Models 80.4 87.7 74.4 63.4 77.2 / 61.3 72.3 / 53.5 81.2 / 68.4 71.9 82.7 74.2
Human 92.8 97.5 97.0 - 91.2 / 82.3 91.2 / 82.3 90.1 / - - - -
Table 2: Single-task and multi-task intermediate-task training results, with and without MLM co-training during intermediate-task training. For each target task, we report the mean result across all target languages. Multi-task experiments use all intermediate tasks except SQuAD. We underline the best results per target task with and without intermediate MLM co-training, and bold-face the best overall scores for each target task. We highlight the best score for each target task, excluding human performance, for XTREME Benchmark scores. : Tasks XQuAD, TyDiQA and Tatoeba do not have held-out test data. : Results obtained with our best-performing model for each target task, selected based on the development set.

XLM-R + Intermediate-Task Training + MLM

We consider a variant of intermediate-task training where the pretrained XLM-R model is trained on both an intermediate task as well as multilingual MLM simultaneously. We treat multilingual MLM as an additional task in the intermediate training phase, and use the same multi-task sampling strategy as above.

3.2 Results

Intermediate-Task Training

As shown in Table 2, no single intermediate task yields positive transfer across all target tasks. The target tasks TyDiQA, BUCC and Tatoeba see consistent gains from most or all the intermediate tasks. In particular, BUCC and Tatoeba, the two sentence retrieval tasks with no training data, benefit universally from intermediate-task training. PAWS-X, XQuAD and MLQA also exhibit gains with the additional intermediate-training on some intermediate tasks. On the other hand, we find generally flat or negative transfer to XNLI, POS and NER target tasks.

Among the intermediate tasks, we find that SQuAD performs best; in addition to having positive transfer to BUCC and Tatoeba, it also leads to improved performance across the three question-answering target tasks. Additionally, CommonsenseQA and CosmosQA lead to slightly improved performance across the question-answering tasks. ANLIdoes not lead to improved performance on XNLI as we first expected, despite the additional training data. This is consistent with Nie et al. (2019), who showed that adding ANLI data to SNLI and MNLI training sets did not improve performance on NLI benchmarks. QQP significantly improves sentence retrieval tasks performance but has broadly negative transfer to the remaining target tasks. CCG also has relatively poor transfer performance, consistent with Pruksachatkun et al. (2020).

One might find it surprising that intermediate-task training on SQuAD improves XQuAD and MLQA performance, which also use SQuAD as training data. We follow the sample implementation for target task training in the XTREME benchmark, which trains on SQuAD for only 2 epochs. This may explain why an additional phase of SQuAD training can improve performance.

MLM and Multi-Task Training

Incorporating MLM during intermediate-task training shows no clear trend. It reduces negative transfer, as seen in the cases of CommonsenseQA and QQP, but it also tends to reduce positive transfer, for instance as seen with intermediate-training on SQuAD. Both TyDiQA and PAWS-X see a small improvement from incorporating multilingual MLM training–in particular, every combination of intermediate tasks and MLM has strictly positive transfer to TyDiQA with an improvement of at least 2 points. The multi-task intermediate-training transfer results are on par with the average of the single intermediate-task transfer results, showing no clear benefit from using the additional tasks.

XTREME Benchmark Results

At the bottom of Table 2, we show results obtained by XLM-R on the XTREME benchmark as reported by Hu et al. (2020), results obtained with our re-implementation of XLM-R (i.e. our baseline), and results obtained with our best models, which use intermediate-task training are selected according to development set performance on each target task. Scores on the XTREME benchmark are computed based on the respective target-task test sets where available, and based on development sets for target tasks without separate held-out test sets.777We compute these scores internally as the official leaderboard is not yet open for submissions as of writing. We are generally able to replicate the best reported XLM-R baseline results, except for: Tatoeba, where our implementation significantly underperforms the reported scores in Hu et al. (2020); and TyDiQA, where our implementation outperforms the reported scores. On the other hand, our best models show gains in 8 out of the 9 XTREME tasks relative to both baseline implementations, attaining an average score of 74.2 across target tasks, a 6.1 point improvement over the previous best reported average score of 68.1.

4 Related work

Sequential transfer learning using pretrained Transformer-based encoders

(Phang et al., 2018) has been shown to be effective for many text classification tasks. This setup generally involves fine-tuning on a single task (Pruksachatkun et al., 2020; Vu et al., 2020) or multiple tasks (Liu et al., 2019a; Wang et al., 2019b; Raffel et al., 2019b), occasionally referred to as the intermediate task(s), before fine-tuning on the target task. Our work builds upon this line of work, focusing on intermediate-task training for improving cross-lingual transfer.

Early work on cross-lingual transfer mostly relies on the availability of parallel data, where one can perform translation (Mayhew et al., 2017) or project annotations from one language into another (Hwa et al., 2005; Agić et al., 2016). When there is enough data in multiple languages with consistent annotations, training multilingual models becomes an option for cross-lingual transfer (Plank et al., 2016; Cotterell et al., 2017; Zeman et al., 2017).

For large-scale cross-lingual transfer, Johnson et al. (2017)

train a multilingual neural machine translation system and perform zero-shot translation without explicit bridging between the source and target languages. Recent works on extending pretrained Transformer-based encoders to multilingual settings show that these models are effective for cross-lingual tasks and competitive with strong monolingual models on the XNLI benchmark

(Devlin et al., 2019b; Conneau and Lample, 2019; Conneau et al., 2019; Huang et al., 2019a). More recently, Artetxe et al. (2020) showed that cross-lingual transfer performance can be sensitive to translation artifacts arising from a multilingual datasets’ creation procedure.

Finally, Pfeiffer et al. (2020) propose adapter modules that learn language and task representation for cross-lingual transfer, which also allow adaptation to languages that are not observed in the pretraining data.

5 Conclusion and Future work

In this paper, we conduct a large-scale study on the impact intermediate-task training has on zero-shot cross-lingual transfer. We evaluate 7 intermediate tasks and investigate how intermediate-task training impacts the zero-shot cross-lingual transfer to the 9 target tasks in the XTREME benchmark.

Overall, intermediate-task training significantly improves the performance on BUCC and Tatoeba, the two sentence retrieval target tasks in the XTREME benchmark, across almost every intermediate-task configuration. Our best models obtain 5.9 and 25.5 point gains on BUCC and Tatoeba, respectively, in the leaderboard when compared to the best available XLM-R baseline scores (Hu et al., 2020). We also observed gains in question-answering tasks, particularly using SQuAD as an intermediate task, with absolute gains of 0.6 F1 for XQuAD, 0.7 F1 for MLQA, and 5.7 for F1 TyDiQA, again over the best available baseline scores. We improve over XLM-R by 6.1 points on average on the XTREME benchmark, setting a new state of the art. However, we found that neither incorporating multilingual MLM into the intermediate-task training phase nor multi-task training led to improved transfer performance.

While we have explored the extent to which English intermediate-task training can improve cross-lingual transfer, an obvious next avenue of investigation for future work is whether multilingual or target-language intermediate-task training can further bolster performance, and how the relationship between the intermediate- and target-task languages influences transfer.

Acknowledgments

This project has benefited from support to SB by Eric and Wendy Schmidt (made by recommendation of the Schmidt Futures program), by Samsung Research (under the project

Improving Deep Learning using Latent Structure

), by Intuit, Inc., by NVIDIA Corporation (with the donation of a Titan V GPU), by Google (with the donation of Google Cloud credits). IC has received funding from the European Union’s Horizon 2020 research and innovation program under the Marie Skłodowska-Curie grant agreement No 838188. This project has benefited from direct support by the NYU IT High Performance Computing Center.

References

Appendix A Implementation details

a.1 Intermediate Tasks

For intermediate-task training, we use a learning rate of 1e-5 without MLM, and 5e-6 with MLM. Hyperparameters in the Table 

3 were chosen based on an initial search based on intermediate task validation performance. We use a warmup of 10% of the total number of steps, and perform early stopping based on the first 500 development set examples of each task. For CCG, where tags are assigned for each word, we use the representation of first sub-word token of each word for prediction.

Task Batch size # Epochs
ANLI 24 2
CCG 24 15
CommonsenseQA 4 10
Cosmos QA 4 15
HellaSwag 24 7
QQP 24 3
SQuAD 8 3
MLM 8 -
Multi-task Mixed 3
Table 3: Intermediate-task training configuration.

a.2 Target Tasks / XTREME Benchmark

We follow the sample implementation for the XTREME benchmark unless otherwise stated. We use a learning rate of 3e-6, and use the same optimization procedure as for intermediate tasks. Hyperparameters in the Table 4 follow the sample implementation. For POS and NER, we use the same strategy as for CCG for matching tags to tokens. For BUCC and Tatoeba, we extract the representations for each token from the 13th self-attention layer, and use the mean-pooled representation as the embedding for that example, as in the sample implementation. Similarly, we follow the sample implementation and set an optimal threshold for each language sub-task for BUCC as a similarity score cut-off for extracting parallel sentences based on the development set and applied to the test set.

We randomly initialize the corresponding output heads for each task, regardless of the similarity between intermediate and target tasks (e.g. even if both the intermediate and target tasks train on SQuAD, we randomly initialize the output head in between phases).

Task Batch size # Epochs
XNLI (MNLI) 4 2
PAWS-X 32 5
XQuAD (SQuAD) 16 2
MLQA (SQuAD) 16 2
TyDiQA 16 2
POS 32 10
NER 32 10
BUCC - -
Tatoeba - -
Table 4: Target-task training configuration.

a.3 Software

Experiments are run using the NLP Runners888https://github.com/zphang/nlprunners

library, based on PyTorch

(Paszke et al., 2019) and Transformers (Wolf et al., 2019).