Unicoder: Universal Encoder

Unicoder is an abbreviation of “Universal Encoder” which is a language encoder that is universal across different number of languages. The Unicoder model was proposed by Microsoft in 2019 and published in this paper: Unicoder: A Universal Language Encoder by Pre-training with Multiple Cross-lingual Tasks. The official code for this paper can be found on Microsoft’s official GitHub repository: microsoft/Unicoder.

A universal language encoder, or Unicoder, is a model that is able to generate a sentence representation embedding independent on the input language. From the name of the paper, we can see that they were able to create this universal language encoder by using pre-training of cross-lingual tasks. These cross-lingual tasks help Unicoder to learn the mappings of different languages. And that’s what we are going to talk about next.

Pre-training Objectives

In this part, we are going to discuss the five language modeling objectives used for the language modeling pre-training in the Unicoder. Since the XLM model outperformed mBERT in cross-lingual tasks, Unicoder uses the same structure as XLM and uses the same pre-training tasks along with three more novel pre-training tasks; all listed below:

MLM

MLM stands for “Masked Language Modeling” which is the first objective out of the five objectives used for learning the Unicoder. MLM is the same objective used for BERT where we randomly sample $15\%$ of the BPE tokens from the text streams, replace them by a $\lbrack MASK\rbrack$ token $80\%$ of the time, by a random token $10\%$ of the time, and we keep them unchanged $10\%$ of the time.

In fact, there are two differences between this approach and the one used with BERT:

  • Here, they used text streams of an arbitrary number of sentences (truncated at 256 tokens) instead of pairs of sentences.

  • To counter the imbalance between frequent tokens (e.g. punctuations or stop words) and rare tokens, they sub-sampled the frequent outputs according to a multinomial distribution, whose weights are proportional to the square root of their invert frequencies.

TLM

TLM stands for “Translation Language Modeling” which is the second objective used for pre-training the Unicoder. MLM objective is unsupervised and only requires monolingual data and cannot be used to leverage parallel data when it is available. So, this objective is used to leverage the existence of parallel data to improve cross-lingual pre-training.

TLM is an extension of MLM objective, where instead of considering monolingual text streams, we concatenate parallel sentences and perform MLM over this concatenation. For example, let’s assume that we have a parallel entry of two languages English and French as shown below:

To predict a word masked in an English sentence, the model can either attend to surrounding English words or to the French translation, encouraging the model to align the English and French representations. In particular, the model can leverage the French context if the English one is not sufficient to infer the masked English words. To facilitate the alignment, we also reset the positions of target sentences.

Word Recovery

This task aims to let the pre-trained model learn the underlying word alignments between two languages. This task is mainly motivated by the attention matrix used in the neural machine translation models. So, let’s first recap how that attention mechanism worked:

Given a bilingual sentence pair $\left( X,\ Y \right)$ of two languages $\left( s,\ t \right)$ of lengths $n$ and $m$ respectively, we obtain word embeddings $\left( X^{s},\ Y^{t} \right)$. The XLM way of doing that is by adding the language and position embeddings over this pair to obtain word embeddings:

\[X^{s} = \left\{ x_{1}^{s},...x_{n}^{s} \right\},\ \ \ \ \ \ \ Y^{t} = \left\{ y_{1}^{t},...y_{m}^{t} \right\}\]

Now that we have word embeddings, let’s apply the attention mechanism to obtain attentive embeddings $X^{t} = \left\{ x_{1}^{t},…x_{n}^{t} \right\}$. Using a trainable weight $W \in \mathbb{R}^{3 \times h}$ where $h$ is the word embedding dimension, the attention matrix $A_{\text{ij}} \in \mathbb{R}^{m \times n}$ will be:

\[A_{\text{ij}} = W\left\lbrack x_{i}^{s},\ y_{j}^{t},\ x_{i}^{s}\bigodot y_{j}^{s} \right\rbrack\]

Where:

  • $x_{i}^{s}$ is the $i^{\text{th}}$ word embedding of the source sentence $s$.

  • $y_{j}^{t}$ is the $j^{\text{th}}$ word embedding of the target sentence $t$.

  • $\bigodot$ is element-wise multiplication.

  • $\lbrack a,\ b\rbrack$ is the concatenation of two vectors; $a$ and $b$.

Now, the attentive embeddings $X^{t} = \left\{ x_{1}^{t},…x_{n}^{t} \right\}$ will be computed as:

\[x_{i}^{t} = \sum_{j = 1}^{n}{\text{softmax}\left( A_{\text{ij}} \right)y_{j}^{t}}\]

That’s how the attention mechanism works. Given that attentive embedding $X^{t}$ to this task, the model tries to predict the original sentence $X$ as shown in the following graph:

Paraphrase Classification

This task takes the concatenation of two sentences from different languages as input and classifies whether they are with the same meaning or not. This task is shown below in the following figure:

The dataset used in this task is created from machine translation datasets. Each bilingual sentence pair $\left( X,\ Y \right)$ servers as a positive sample. For negative samples, the most straight forward method is to replace $Y$ to a random sampled sentence from the target language. But this will make the classification task too easy.

To make the negative samples hard, they trained a light-weight paraphrase model with random negative samples. Then, they used this model to select sentence with high similarity score to $X$ but doesn’t equal to $Y$ as hard negative samples. They chose Deep Averaging Network (DAN) from this paper as the light-weight model. The dataset had $50\%$ positive samples and $50\%$ negative samples.

Cross-lingual Masked Language Model

Previous successful pre-training language models such as BERT and GPT, were conducted on document-level corpus rather than sentence-level corpus. Cross-lingual document is a sequence of sentences, and the sentences are written with different languages.

That’s why the researchers proposed cross-lingual masked language model task whose input come from cross-lingual document. One issue is that people won’t write cross-lingual document in most cases. To solve this issue, they decided to construct such the cross-lingual document-level corpus by replacing the sentences to their translation. Then, they masked this document-level corpus the same way as MLM as shown below:

In the paper, they replaced sentences with even indices to their translation in the other language separating them with $\left\lbrack /S \right\rbrack$ token. They truncated the cross-lingual document till it became $256$-tokens long, then fed it to the Unicoder.

Data

Unicoder is pre-trained on 15 languages. For MLM, they used the Wikipedia of these 15 languages. The other four tasks needed MT dataset, so they used the same dataset as the XLM paper which were collected from MultiUN, IIT Bombay corpus, OpenSubtitles 2018, EUbookshop corpus and GlobalVoices. The number of parallel data they used is reported at the following table:

Note:

  • When training, all language were sampled equally.

  • The max sequence length of MLM and cross-lingual language model is 256. For the other three tasks with two sentences as input, they set the max sequence length to 128 so the sum of them is 256.

Multilingual Fine-tuning

The fine-tuning task used in this paper is machine translation. There are three settings that people use when it comes to using machine translation for fine-tuning; and they are:

  • TRANSLATE-TRAIN: fine-tune the model on pseudo translation of target language training set and test it on target testing set.

  • TRANSLATE-TEST: fine-tune the model on source training set and test it on pseudo translation of source testing set.

  • TEST: fine-tune the model on source training set and test it on target testing set.

To further understand the difference , let’s assume that the source language is English and the target language is Chinese. Each dataset has training set and test set and the task is to evaluate on Chinese test, then the three settings will do as the following figure:

Note:
The source language and the target can be more than one language; but they can never intersect. In other words, a language can’t be in the source and the target at the same time.

Inspired by multi-task learning, the authors proposed a new fine-tuning strategy called “Multilingual Fine-tuning” where fine-tuning is done on training data of source language + pseudo translation of the source language as shown in the following figure:

Results

The Unicoder model is initialized from the XLM model, which means it is a $12$-layer transformer with $1024\ $hidden units, $16\ $heads, GELU activation. To build the vocabulary, they used BPE tokenization where the vocabulary size is $95,000$. To train Unicoder, they used Adam optimizer with learning rate that starts from $1e^{- 5}$ with invert square root decay. As a regularizer, they used a dropout of $0.1$. The batch size was set to $512$ in pre-training and $32$ in fine-tuning.

They fine-tuned Unicoder over these benchmarks:

  • XNLI: The following table shows test accuracy on the 15 XNLI languages. The source language here is English while the target language changes. The tables shows that Unicoder with Multilingual Fine-tuning achieves the best performance across all 15 languages.

    Using more languages in fine-tuning leads to better performance on the XNLI benchmark:

  • XQA: The following table shows test accuracy on the XQA. The average column is the average of “French (fr)” and “German (de)” result. And again, Unicoder with multilingual fine-tuning achieves the best performance across all three languages.