EMMA-X: An EM-like Multilingual Pre-training Algorithm for Cross-lingual Representation Learning

Expressing universal semantics common to all languages is helpful in understanding the meanings of complex and culture-specific sentences. The research theme underlying this scenario focuses on learning universal representations across languages with the usage of massive parallel corpora. However, due to the sparsity and scarcity of parallel data, there is still a big challenge in learning authentic ``universals'' for any two languages. In this paper, we propose EMMA-X: an EM-like Multilingual pre-training Algorithm, to learn (X)Cross-lingual universals with the aid of excessive multilingual non-parallel data. EMMA-X unifies the cross-lingual representation learning task and an extra semantic relation prediction task within an EM framework. Both the extra semantic classifier and the cross-lingual sentence encoder approximate the semantic relation of two sentences, and supervise each other until convergence. To evaluate EMMA-X, we conduct experiments on XRETE, a newly introduced benchmark containing 12 widely studied cross-lingual tasks that fully depend on sentence-level representations. Results reveal that EMMA-X achieves state-of-the-art performance. Further geometric analysis of the built representation space with three requirements demonstrates the superiority of EMMA-X over advanced models.

Paper

Similar papers

Peer review

Reviewer tN445/10 · confidence 5/52023-06-27

Summary

This paper proposed EM-like multilingual pre-training algorithm using both supervised parallel data and unsupervised monolingual data to train an encoder model such that it can produce language agnostic representations for multiple languages. That is, embedding space of sentences of similar meanings will be the same across languages. The method uses an EM procedures that employs an GMM classifier, which is pre-trained with parallel data to produce a similarity rank between 2 input sentence, and then train the encoder model both with parallel data and ranks produced by GMM with unsupervised data. The method achieves higher scores across many inference, similarity and retrieval tasks compared to unsupervised XLM-R.

Strengths

* The methodology seems original and new, especially for the new rank-based classifier. * The paper presentation and method description is clear and understandable.

Weaknesses

There are multiple concerns with the evaluation and comparison with the method. * Despite utilize large-scale parallel data from multiple languages heavily in the pretraining stage of encoder and GMM classififer, the paper compares its method with unsupervised models like XLM-R. This make the comparison in main table 1 invalid and incomparable. This method is more semi-supervised with supervised pretraining and unsupervised finetuning stages. * There are some comparison with supervised models like LaBSE, but apparently the proposed method do not score well with supervised models. Moreover, the writting seems to suggest the paper is comparing unsupervised vs supervised but in fact this is supervised vs supervised. * XLM-R is not designed for language agnostic representation, but there are notable baselines that achieve the same language-agnostic object with simpler methodology: [LASER](https://github.com/facebookresearch/LASER) which use parallel data and [CRISS](https://arxiv.org/abs/2006.09526) which does not use parallel data. Not only they are not compared, they are also not mentioned in the paper.

Questions

NA

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

No limitation mentioned.

Reviewer oxNi7/10 · confidence 4/52023-06-27

Summary

This paper proposes an EM-like pre-training algorithm called EMMA-X to learn cross-lingual representations. The authors unify semantic relation classification and universal representation learning into the framework. To this end, a GMM classifier and a cross-lingual encoder is jointly trained in the algorithm. Finally, the authors validate their method by conducting experiments on 12 cross-lingual tasks including inference, similarity, retrieval, and classification.

Strengths

- The paper is well-structured and easy to follow. - The EM-like pre-training, which includes a semantic relation rank task, is somewhat novel. - The experiments are extensive, and the results generally indicate the proposed method is effective. - The geometric analysis section is interesting and informative, where the authors visualize the representations from four different models.

Weaknesses

- Although the authors claim that EMMA-X can learn good representations with excessive non-parallel, the encoder, as far as I can see, is continue-trained on parallel sentence pairs with InfoNCE on parallel corpora for initialization. Therefore, this initialization weakens the claim. - The author does not verify the choice of a GMM classifier. At least, the author should give some simple intuition to the readers. - Semantic relation task is very important in the proposed framework. The authors claim that a binary separation is not good, but do not give further justification for the choices of N=4 semantic ranks.

Questions

- Consider removing the comma after each equation - Line 69: Why do we have to use L-2 normalization? There are multiple ways to perform normalizations, L-2 is just one of them. Therefore, I would rather say $f(\cdot)$ is a normalization, e.g., L-2. - Figure 1 in caption: "{y1, and y2, y3, ...}" -> {$y\_1, y\_2, y\_3, \cdots$} - In Figure 1, I assume $\gamma^{(x)} - \gamma^{(y_1)}$ is missing - The number of semantic ranks is set to 4. This sounds to me like a random selection. Did you do experiments to justify the choice? - The paper claims that the model can learn to learn cross-lingual universals with the aid of massive multilingual non-parallel data, but the cross-lingual encoder is continued trained on infoNCE on some parallel sentence pairs. I would assume this continue-training step is even more important than the method proposed. I would suggest the author justify the initialization by simply using the model weights of XLM-R. - Line 188-190: The number of sentence pairs is 3.2 billion, but what is the exact number of non-parallel sentences? That should also be mentioned in the main content of the paper. - Line 192: $\mathbb{R}^1$ -> $\mathbb{R}$ - Table 1: the authors should also use special characters to denote which results (BUCC and Tatoeba?) are obtained by zero-shot for EMMA-X. - Figure 2 and Table 4: according to the results, S-BERT performs very well (even better) than the proposed method, especially on long-tail languages. In this case, what would be the benefit of applying EMMA-X to those languages instead of simply using S-BERT?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The paper does not include a limitation or broader impact section, therefore not applicable.

Reviewer s5pr7/10 · confidence 4/52023-07-06

Summary

This paper proposes a new approach to learn cross-lingual representation learning as a universal alignment solution for any two languages without parallel data. The proposed approach resembles EM-like algorithm, which consists of a classifier to quantify the semantic similarity of two non-parallel sentences, as well as a sentence encoder that aligns the embedding representation of different languages. During the training process, the two components receive supervision from each other to reinforce the capability to recognize the accurate semantic similarity, where the paper also provides the theoretical justification to demonstrate the mutual influence being positive and effective. A new benchmark is also proposed that comprises four existing cross-lingual transfer tasks; the proposed approach is compared against other recent baselines, as well as ChatGPT. The empirical results show that new state-of-the-art performance can be obtained by this new approach with good margins.

Strengths

- A new approach is proposed based on a novel EM-like algorithm with mutual inference and supervision by two components. Especially, the approach proposes to rank the similarity of a sentence pair, rather than optimizing the traditional binary positive/negative contrastive objective, which is indeed interesting and effective shown by the results. - State-of-the-art performance is achieved by this approach on the new benchmark consisting different types of existing cross-lingual transfer tasks, comparing against strong baselines such as XLM-R. - A theoretical analysis is provided, in addition to the superior empirical results to justify the effectiveness of the proposed approach.

Weaknesses

- It might be over-claiming that the proposed approach operates to align the representation of different languages without parallel data, since Algorithm 1 clearly shows that parallel corpus is needed to bootstrap both the similarity classifier and the sentence encoder, and all experiments are based on this warm up that leverages parallel corpus. - There lacks more in-depth analysis regarding the ablation of the proposed approach. In particular: - There could be a simplified version of this approach without iterative optimization: we can first train a similarity classifier based on the proposed mixed-up method, and generate similarity rankings of sentence pairs on non-parallel data, which can then be used to optimize the contrastive objective directly for training the encoder. In this way, we can verify whether the improvement mostly comes from the fine-grained similarity ranking, and how much this iterative optimization mechanism contributes to the final performance. - There could also be another experiment that downgrades the similarity ranking to binary, as if the traditional positive/negative pair. Thus, it could show again that whether this EM-like algorithm is really necessary and how much this process could improve upon the plain contrastive objective.

Questions

What is the primary reason for adopting GMM as the similarity classifier? Can we use a normal classifier (e.g. using CLS on Transformers) and simply use their softmax probability for each rank?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

3 good

Contribution

3 good

Limitations

It would be good to discuss how much degradation could happen if there is not enough parallel corpus for the warm-up.

Reviewer FMvk6/10 · confidence 3/52023-07-07

Summary

The paper proposes EMMA-X, an EM-like approach to pretrain multilingual models. It learns the cross-lingual representation learning task and semantic relation prediction task within EM. They propose a new benchmark, XRETE, to evaluate the experiments with 12 cross-lingual tasks. The training involves two stages: 1) pretraining using parallel data (bitext) and 2) training with EM Framework with non-parallel data. The proposed approach outperforms the baselines.

Strengths

- The proposed method outperforms other existing pre-trained models, showing the solid multilingual / cross-lingual ability. - The proposed method is backed by a strong theoretical analysis - The paper proposes a new benchmark from existing datasets to evaluate the cross-linguality.

Weaknesses

- The main paper does not provide enough information about the model. I would suggest moving the model details from the Appendix to the main paper. - The benchmark does not provide the model parameters. The comparison may not be fair since the model sizes are different.

Questions

**Suggestion** - Figure 1 is not easily understandable. It would be great if there is a clear step-by-step procedure in the caption.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

No. The authors can have a section to address the limitations.

Reviewer LAMb6/10 · confidence 4/52023-07-07

Summary

This paper proposes to apply the EM framework to realize unified sentence representation learning with non-parallel multilingual data. This framework consists of two modules, a GMM classifier and a cross-language encoder, which are responsible for semantically related classification and cross-language unified representation. However, this is achieved on the premise that the parallel multilingual data initializes the models. In addition, the author conducted a theoretical analysis and forms a new benchmark, based on which the effectiveness of the framework was proved.

Strengths

The paper applies the EM algorithm to realize the optimization of the multilingual unified representation parallel-trained model under non-parallel data.

Weaknesses

1. Why choose the GMM model as the Semantic Relations model? 2. Why form a XRETE benchmark, what are its innovations and necessity? 3. This article introduces monolingual data from a continuous perspective, so what are its advantages compared with discrete methods(e.g. Back-translation)

Questions

1. Why choose the GMM model as the Semantic Relations model? 2. Why form a XRETE benchmark, what are its innovations and necessity? 3. This article introduces monolingual data from a continuous perspective, so what are its advantages compared with discrete methods(e.g. Back-translation)

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer s5pr2023-08-10

Thanks for your response. I am looking forward to the experiments that substitutes GMM with a normal classifier, which I think would provide good insights on this design choice during the EM iterations.

Reviewer tN442023-08-14

Response

Thanks the authors for the clarification. Given it is a fair comparison, the method is still not strong and outperforms consistently across benchmarks. Nonetheless I changed the scores to award the novelty of the work.

Reviewer oxNi2023-08-18

Thank you very much for your detailed response. All my questions have been appropriately answered now. I'll raise the recommendation score.

Authorsrebuttal2023-08-18

Thank you so much for your thoughtful message. I'm delighted to know that I could assist you in providing the information you needed. Your willingness to raise the recommendation score is truly generous and means a lot to me. Thanks again for your appreciation!

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC