Phanxuan Phuc

fakerphan

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

3 mins read

Published by: FAIR

CRISS

CRISS stands for “Cross-lingual Retrieval for Iterative Self-Supervised Training” which was created by FacebookAI in 2020 and published in this paper: “Cross-lingual Retrieval for Iterative Self-Supervised Training. The official code for this paper can be found in the fairseq GitHub repository: criss.

CRISS is a novel self-supervised training technique that iteratively trains a multilingual NMT model on mined bi-text data with the help a cross-lingual model. The cross-lingual language model used here is a pre-trained mBART model as shown below:

As see from the previous figure, CRISS consists of different components:

  • Cross-lingual language model (pre-trained mBART).

  • Bi-text Mined data.

  • Multilingual NMT (standard Seq2Seq Transformer Model).

Pre-trained mBART

mBART is used with CRISS because it’s able to form a language-agnostic sentence representation. To understand the language agnosticity of mBART, they performed the following study on the TED58 dataset which contains multi-way translations of TED talks in 58 languages.

For each language pair, they encoded both the source and target sentences using mBART to obtain two sentence representations. They did that for the whole dataset. Then, for each sentence in the source language, then found the closest sentence in the target language using cosine similarity. And they found out that mBART was able to retrieve the correct translation with 57% accuracy on average as shown in the following figure:

The high retrieval accuracy suggests that the pre-trained mBART model is able to generate language agnostic representation that are aligned at the semantic level in the vector space. Moreover this cross-lingual representation can be improved by a relatively small amount of parallel data of just one direction. As we can see from the following figure, the sentence retrieval accuracy was improved for all language directions by 27% (absolute) on average after fine-tuning it on just Japanese-English parallel dataset (223, 000 sentence pairs in the IWSLT17).

Note:
This is a very important discover! This means that in order to train a multilingual model that can translate between N languages, we don’t need to fine-tune it all $N(N - 1)$ language pairs, but only a subset of them.

Bi-text Mining

To mine bi-text data without supervised signals, they proposed the following algorithm which uses monolingual data of two languages as an input and a pre-trained MNMT model. The full algorithms can be seen below:

As we can see, the algorithm is pretty self-explanatory with three additional functions used. Let’s understand them one-by-one:

  • Embed():
    This function takes a sentence and a model and returns the sentence representation of the given sentence resulting from the give model.

  • KNN():
    This is the K-Nearest-Neighbor algorithm. This function takes as input three variables: the vector representation of the source sentence, the dataset of the target language, and an integer k. And it returns the nearest k vectors from the given dataset to the given sentence representation. FAISS tool is used here for fast searching.

  • score():
    This function takes as an input two sentence representations and returns the score based on the margin function which is considered cosine score normalized by average distances. The margin scoring function is defined below:

\[\text{score}\left( x,y \right) = \frac{\cos\left( x,y \right)}{\sum_{z \in N_{x}}^{}\frac{\cos\left( x,z \right)}{2k} + \sum_{z \in N_{y}}^{}\frac{\cos\left( z,y \right)}{2k}}\]

Using this algorithm enables us to create bi-text data in an unsupervised manner. Now, let’s get to how to use this mined bi-text data to train MNMT model.

Multilingual Training

The following algorithm shows how to perform a multilingual training. First, we simply augment each mined pair $\left( x,y \right)$ of sentences by adding a target language token at the beginning of y to form a target language token augmented pair $\left( x,y’ \right)$. Then, we aggregate all mined pairs $\left\{ \left( x,y \right)’ \right\}$ of the mined language pairs into a single data set to train a standard seq2seq machine translation transformer models from the pre-trained mBART model.

Unsupervised MT

CRISS was evaluated on unsupervised neural machine translation benchmarks that cover both low resource and high resource language directions. For English-French, they used WMT’14, for English-German and English-Romanian, they used WMT’16 test data, and for English-Nepali and English-Sinhala, they used Flores test set.

As shown in the following table, the CRISS model overperforms state-of-the-art in 9 out of 10 language directions: