Phanxuan Phuc

fakerphan


XLM-R

XLM-R stands for “Cross-lingual Language Modeling-RoBERTa” which was created by FacebookAI in 2019 and published in this paper: “Unsupervised Cross-lingual Representation Learning at Scale”. The goal of paper is to improve the XLM model’s performance on both cross-lingual and monolingual understanding tasks. XLM-R follows the XLM model architecture with a changes that improve the performance at scale. And these changes are:

  • Maked Language Modeling:
    In XLM-R, they used the same MLM objective as the XLM model with only once change which is they removed the language embeddings which allows the model to better deal with code-switching.
  • Removed TLM:
    The focus of this paper is Unsupervised Cross-lingual Representation Learning. So, it made sense to remove the supervised objective in the XLM model which was the Translation Language Modeling objective or TLM for short.

  • Scaling to 100 languages:
    XLM-R is trained on 100 languages unlike XLM which was trained on just 15 languages. And for the first time, XLM-R shows that it is possible to have a single large model for all languages, without sacrificing per-language performance.

  • Bigger Vocabulary:
    XLM-R uses a shared vocabulary of 250k tokens while XLM uses 95k. The vocabulary consists of subwords encoded using Byte-Pair Encoding (BPE).

  • Validation Stopping Criterion:
    XLM uses the perplexity as a stopping criterion for pre-training. In XLM-R, they used downstream tasks as a stopping criterion as they observed that performance on downstream tasks continues to improve even after validation perplexity has plateaued.

  • Increasing the Training Data:
    They collected monolingual data about 100 languages using common crawl which was about 2.5 Terabytes. The following figure shows the difference between Wikipedia (the data used to train multilingual BERT or mBERT) and Common Crawl (the data used to train XLM-R).

And that is basically it! There are no novel parts in this paper. However, this paper provides a good analysis of multilingual models as we are going to see next.

Trade-offs

As I said earlier, this paper doesn’t provide any novel parts but it provides a good understanding and analysis to the multilingual models according to various factors:

Performance Vs # languages

In the paper, they called this the “the curse of multilinguality”. For a fixed sized model, the per-language capacity decreases as we increase the number of languages. While low-resource language performance can be improved by adding similar higher-resource languages. So, these two factors have to be traded off against each other.

Using XNLI as a metric, the authors trained different fixed-size XLM-R models in 7, 15, 30, 60 and 100 languages. And they found out that; initially, as we go from 7 to 15 languages, the model is able to take advantage of positive transfer which improves performance, especially on low resource languages. Beyond this point the curse of multilinguality kicks in and degrades performance across all languages.

The issue is even more prominent when the capacity of the model is small. To show this, they made the Transformer wider as they added more languages. They trained models in 7, 30, and 100 languages with hidden size of 768, 960, and 1152 respectively. As the following figure shows, the added capacity allows XLM-30 to be on par with XLM-7, thus overcoming the curse of multilinguality. But the added capacity for XLM-100, however, is not enough and it still lags behind.

Performance Vs Low / High-resource

Here, we study the effect of data sampling of high-resource (English and French) and low-resource (Swahili and Urdu) languages on the performance. Specifically, we investigate the impact of varying the $\alpha$ parameter which controls the exponential smoothing of the language sampling rate. Recap the shared vocabulary part from XLM if you can’t relate.

Models trained with higher values of $\alpha$, see batches of high-resource languages more often which increases the performance on high-resource languages. And when we decrease the value of , the model see batches of low-resource languages more often. So, we need to trade-off between these two factors.

They found that $\alpha = 0.3$ is the optimal value for α if we are considering both low-resource and high-resource languages.

Performance Vs Vocabulary

Previously, we showed the importance of scaling the model size as we increase the number of languages. Similarly, scaling the size of the shared vocabulary can improve the performance.

To illustrate this effect, they trained multiple models with different vocabulary sizes while keeping the overall number of parameters constant by adjusting the width of the transformer. And as shown in the following figure, we observe a 2.8% increase in XNLI average accuracy as we increase the vocabulary size from 32K to 256K:

Performance Vs Data Size

As you have probably figured out, the more data you have, the better the cross-lingual language model will be. They compared the same model trained on two different datasets: Wikipedia (60 Gigabytes) and Common Crawl (2.5 Terabytes).

Performance Vs Batch Size

Similarly for the batch sizes, the bigger the batch size is, the better the cross-lingual language model will be:

Results

Here, we are going to talk about the results of XLM-R model in comparison with all other state-of-the-art models on multiple tasks. In the paper, they used different sizes of XLM-R which are:

XLM-R BASE XLM-R
Transformer Blocks 12 24
Feed-Forward hidden neurons 768 1024
Attention Heads 12 16
Parameters 270 million 550 million

Multilingual Results

Here, we are going to compare the XLM-R model over cross-lingual understanding tasks such as:

  • XNLI:
    The following table shows the performance on the XNLI dataset which contains 15 different languages. In the table we can see that using the translate-train-all approach which leverages training sets from multiple languages, XLM-R obtains a new state of the art on XNLI of 83.6% average accuracy:
  • NER:
    The following table shows the results using CoNLL2002 and CoNLL-2003 datasets using the F1 metric:
  • Question Answering:
    The following table summarizes the results on MLQA question answering using the F1 and EM (exact match) scores for zero-shot classification where models are fine-tuned on the English Squad dataset and evaluated on the 7 languages of MLQA:

Monolingual Results

Here, we are going to compare the XLM-R model over monolingual understanding tasks such as:

  • GLUE:
    The following table shows the XLM-R performance on the the GLUE benchmark:

Important Finding

In this paper, they proved that the multilingual models obtain better performance over their monolingual counter-parts. They provided the first comprehensive study to assess this claim on the XNLI benchmark. For comparison, they used XLM models (called it XLM-7) and monolingual BERT models on 7 languages and compare performance as shown in the following table:

According to the former table, we can conclude the following:

  • Monolingual BERT models outperform XLM-7 for both Wikipedia and CC by 1.6% and 1.3% average accuracy.

  • XLM-7 outperforms BERT models when it leverages the training sets coming from multiple languages (translate-train-all).