Phanxuan Phuc

fakerphan

I’m a AI Engineer in Vin Bigdata. And I summarize research papers in these topics:

4 mins read

Published by: Beihang University & Microsoft & Tencent Inc.

ALM: Alternating Language Modeling

ALM stands for “Alternating Language Modeling” which is a novel cross-lingual pre-training method proposed by Microsoft in 2021 and published in their paper: Alternating Language Modeling for Cross-Lingual Pre-Training. The official code for this paper can be found in the following repository: ALM.

The idea behind ALM is to predict one language conditioned on the context of the other language which can minor the gap between the embeddings of languages and that’s very beneficial for the cross-lingual setting. The following figure shows the difference between ALM and XLM. In ALM, the input sequence of ALM is mixed with different languages.

Code-Switching

Given a bilingual sentence pair $\left( X,Y \right)$ with the source sentence $X = \left\{ x_{1},\ x_{2},\ …x_{N} \right\}$ and the target translation $Y = \left\{ y_{1},\ y_{2},\ …y_{M} \right\}$, where $N$ and $M$ are the lengths of the source and target sentences respectively, a code-switched sentence $U = \left\{ u_{1},\ u_{2},\ …x_{L} \right\}$ of length $L$ is composed of phrases from $X$ and $Y$.

For each phrase $u_{\left\lbrack i:j \right\rbrack}$, it comes from either the source sentence $x_{\left\lbrack a:b \right\rbrack},\ 1 \leq a \leq b \leq N$ or the target sentence $y_{\left\lbrack c:d \right\rbrack},\ 1 \leq c \leq d \leq N$ where the constraint is that these two phrases are the linguistic translation counterpart in the parallel sentence $\left( X,Y \right)$. The following figure shows different examples of code-switched data between English and Chinese:

When most of the code-switched sequence $U$ is derived from $X$, this sample is called major-source-language sample. When most of the code-switched sequence $U$ is derived from $Y$, this sample is called major-target-language.

Due to the scarcity of natural code-switched sentences, the researchers had to construct them from bilingual sentence pairs. They did that by following these steps:

  • First, they performed word alignment with the GIZA toolkit between the parallel sentences, and extract a bilingual phrase table using statistical machine translation techniques.

  • Then, for each sentence pair in training corpus, they created the major-source-language samples by substituting some phrases in source sentence with the corresponding target phrases with highest probabilities in phrase table. And same for major-target-language, they substituted some phrases in target sentence with the corresponding source phrases.

  • Each phrase is limited to less than 5 words for both source language and target language. And the substituted words are less than 30% of the total words in the sentence.

  • Each bilingual sentence pair is used to create multiple alternating language sentences by randomly choosing the substituted phrases as shown in the following figure:

Pre-training

ALM uses the encoder-part of the Transformer-base architecture with 1024 embedding and hidden units, 8 heads and learned positional embeddings. During training, they used BPE with 95K sub-word tokens. They used Adam optimizer with parameters of $\beta_{1} = 0.9$ and $\beta_{2} = 0.98$ accompanied with a dropout rate of $0.1$. They set the inverse-square-root learning rate schedule with a linear warmup where the number of warmup step is $4000$ and a learning rate of $0.0005$. The batch size was set to $8192$ tokens. During decoding, they sued a beam size of 8.

MLM was done like the following; Given a parallel sentence pair, they combined two sentences from different languages into a single code-switched sequence as described above. Then, they mask out a certain percentage of words in the sequences and feed them into Transformer model to learn to predict the words being masked out as shown in the following figure:

Data used in pre-training was monolingual data from Wikipedia and bilingual data from IWSLT14 German-to-English translation dataset. The input sentences to the ALM model were $10\%$ from the source sentence (without code-switching), $10\%$ from the target sentence (without code-switching), and the rest $80\%$ were code-switched. MLM was applied by randomly masking $15\%$ of the tokens using:

  • The [MASK] token $80\%$ of the time.

  • A random token $10\%$ of the time.

  • Keep them unchanged $10\%$ of the time.

Machine Translation

After pre-training ALM on the MLM objective, they fine-tuned it on the machine translation task. Using parallel data from WMT14 English-German dataset and IWSLT14 German-English dataset, they used ALM as the encoder of machine translation, and construct a Transformer-based decoder initialized from the ALM. In other words, they fed the source language into ALM, and generated the target language with decoder.

Fine-tuning was done with the Adam optimizer with a linear warmup as the pre-training. The learning rates was tuned based on the performance on the validation set which was $5 \times 10^{- 4}$ for IWSLT14 German-English dataset and $10^{- 3}$ for WMT14 English-German. Early stopping was used.

The following figure shows the performance on the two datasets; the one on the left shows the results of WMT14 and the one on the right shows the results of IWSLT14. In both datasets, we can see that ALM outperforms all other models: