Continual Pre-Training for Cross-Lingual LLM Adaptation: Enhancing Japanese Language Capabilities

Cross-lingual continual pre-training of large language models (LLMs) initially trained on English corpus allows us to leverage the vast amount of English language resources and reduce the pre-training cost. In this study, we constructed Swallow, an LLM with enhanced Japanese capability, by extending the vocabulary of Llama 2 to include Japanese characters and conducting continual pre-training on a large Japanese web corpus. Experimental results confirmed that the performance on Japanese tasks drastically improved through continual pre-training, and the performance monotonically increased with the amount of training data up to 100B tokens. Consequently, Swallow achieved superior performance compared to other LLMs that were trained from scratch in English and Japanese. An analysis of the effects of continual pre-training revealed that it was particularly effective for Japanese question answering tasks. Furthermore, to elucidate effective methodologies for cross-lingual continual pre-training from English to Japanese, we investigated the impact of vocabulary expansion and the effectiveness of incorporating parallel corpora. The results showed that the efficiency gained through vocabulary expansion had no negative impact on performance, except for the summarization task, and that the combined use of parallel corpora enhanced translation ability.

Paper

Similar papers

Reviewer ATKc6/10 · confidence 4/52024-05-10

Summary

This works presents a new version of LLAMA2, adapted for Japanese. First, LLAMA's vocabulary is expanded with a Japanese vocabulary, created from the new training data through MeCab, Unidic and BPE. Then, the model is trained by performing continual pre-training. Finally, the new model is tested on several downstream tasks, achieving satisfying results.

Rating

6

Confidence

4

Ethics flag

1

Reasons to accept

The work describes the construction of a new model and an analysis which could benefit the community.

Reasons to reject

As it is right now, the paper relies too much on the appendixes and is very hard to read and understand. Despite its importance, the experimental setup is missing. Similarly, the evaluation lacks context: It is mentioned which downstream task are being used for the evaluation, but there is no description even of which metrics and corpora are being used. Thus, result tables are not informative enough--we don't even know if greater values means better results or if the corpora could have already been seen at training time! While I find the work interesting, I believe it would benefit from a revision. The main details should be described on the document's body instead of relaying so much on the appendixes.

Reviewer rjny7/10 · confidence 4/52024-05-12

Summary

The authors present work that extends baseline Llama-2 large language models to include (more) Japanese-language data through continued pre-training on up to 100B additional tokens of Japanese data. By employing model vocabulary expansion in conjunction with the additional data, the authors improve generation efficiency by up to 78% due to the introduction of Japanese-specific vocabulary terms. Experimental results show marked improvements on Japanese performance on six evaluation tasks in addition to improvements in English-Japanese machine translation (MT). Some performance degradation is noted in English-based evaluation tasks as well as Japanese-English MT performance. Lastly, the authors compare their continued pre-training approach against similarly-sized Japanese language models trained from scratch.

Rating

7

Confidence

4

Ethics flag

1

Reasons to accept

The contributions in this paper are systematically explored to show effects of the two major contributions. Vocabulary expansion is demonstrated to increase generation efficiency and generally has little effect on task evaluation (with the exception of a minor performance regression on summarization.) Continued training with additional Japanese monolingual and Japanese-English parallel data improves performance on Japanese evaluation tasks. The authors also demonstrate that pretraining with a strong English baseline model outperforms similar-sized models that were trained from scratch. This pretraining approach is tested on the three common Llama-2 model sizes 7B, 13B and 70B to show that this tactic scales with differing model sizes.

Reasons to reject

Language-specific LLM pretraining has been explored in many other works, but the combination of examination of model size and effects of vocabulary expansion may help mitigate this fact.

Questions to authors

It would be interesting to see results of other ways of mitigating 'catastrophic forgetting' such as Elastic Weight Consolidation in place of Experience Replay, though the implementation details may be tricky.

Reviewer Uybi2024-06-02

Reviewer Uybi's response after rebuttal

Thank you to the authors for the explanations and clarifications. I re-read the paper while referencing to your rebuttal and related works. I feel that I have a better understanding of the contribution of this paper now. While most points are understood, I feel that I still have reserved opinions about the two statement below. **1.** Re: intact token v. byte-level tokens: whether the former provides more semantics when fed to an LLM. The original statement and the explanation are both centred around human interpretation of words/characters. After rebuttal, I feel unsure about whether/how "morphology" or "phonology" are used by an LLM if the tokens are mapped to indices and then vectors. LLM's embeddings (so-called semantics) are derived by contextualizing on the input itself and surrounding tokens. Note that I completely agree with the choice of expanding the vocabulary for extra language support. It improves *tokenizer fertility*. I find the current justification (relying on semantics) unsupported. **2.** I feel that the statement "impact of vocabulary expansion on performance is minor" is not accurate. I would say that if some changes are noticeably positive while some are negative (10% relative change is a lot), it actually means that the impact on test performance is significant rather than "minor"?

Authorsrebuttal2024-06-04

> I find the current justification (relying on semantics) unsupported. We appreciate the reviewer's insights and understanding of the concerns regarding the use of "morphology" and "phonology" in the context of LLMs where tokens are mapped to indices and vectors. Our justification for stating that byte-level tokens "do not provide any semantic meaning" is based on three primary reasons: semantic integrity, prevention of generating non-existent characters, and improved tokenization efficiency and contextualization. 1. Semantic Integrity: Byte-level tokens can result in shared tokens for characters with vastly different meanings. For instance, "猟" (hunt) is encoded as <0xE7> <0x8C> <0x9F>, and "猫" (cat) is encoded as <0xE7> <0x8C> <0xAB>, sharing with the identical first two bytes despite their different meanings. This overlap can dilute the semantic integrity of Japanese kanji letters, which are ideograms unlike most of other languages. It is possible to argue that English also suffers from the same problem, e.g., “cat” and “car” are totally different, but the treatment of English tokenization is much better than that of Japanese because the tokenizer of the base model was trained on a corpus with a lot of English text and little Japanese text. In other words, it is unlikely for the base tokenizer to split them into “cat” and “car” into <c> <a> <t> and <c> <a> <r>. While LLMs derive semantics from context, starting with more semantic Japanese tokens is reasonable because Japanese kanji letter are ideograms. 2. Prevention of Generating Non-existent Characters: Byte-level tokenization can produce byte sequences that do not correspond to any valid characters in the target language. For instance, [it is known](https://okumuralab.org/~okumura/misc/230611.html) in Japan that ChatGPT generated the non-existent Japanese word "視覴," allegedly caused by the impossible combination of byte tokens from "覚(perception)" and "聴(listen)." Such occurrences are less likely if we treat a Japanese kanji as a single token, which respects the boundaries of valid characters and words. This reduces the possibility of generating meaningless or incorrect text and thereby improves the reliability and quality of the model's output. 3. Improved Tokenization Efficiency and Contextualization: Representing a kanji letter as a single token provides a more consistent and semantically relevant input to the token embeddings, enhancing the model's ability to understand and generate coherent text. For example, "猫"(cat) = <0xE7> <0x8C> <0xAB> shares the first two tokens with 17 other kanji in byte-level tokenization, which potentially increases the semantic ambiguity and degrades its performance. By adding kanji letters to the vocabulary, we can ensure that each token has a clear and distinct meaning, which improves the model's ability to contextualize and generate accurate representations. > I feel that the statement "impact of vocabulary expansion on performance is minor" is not accurate. Thank you for your feedback. we would like to clarify our position regarding the statement, "the impact of vocabulary expansion on performance is minor." Excluding the XL-Sum task, the only dataset where we observed a relative change of more than 10% is JCQA for the Llama-2-JA-7b model. This kind of change was not observed in other model sizes or tasks, as we noted in §5.1.2. To provide more evidence, we present an additional analysis of the JCQA scores, isolated from the averaged numbers reported in Figure 7. The scores for 7b models’ JCQA with and without vocabulary expansion (VE and $\neg$VE, respectively) are as follows: |Num. of training tokens (B)|JCQA(VE)|JCQA($\neg$VE)| |---|---|---| |0|38.52|38.52| |20|39.23|45.76| |40|45.76|39.14| |60|42.18|49.15| |80|49.15|52.28| |100|48.08|54.33| With 40B tokens used for training, the model with vocabulary expansion (VE) outperforms the one without it ($\neg$VE). This indicates that the JCQA scores are not stable. In contrast, Figure 6 shows that the XL-Sum scores consistently show better performance for the model without vocabulary expansion, exhibiting a different trend. Furthermore, the average and standard deviation of JCQA scores after training 40B tokens (40,60,80,100 B Tokens) are: - Vocabulary Expansion (VE): 46.29 ± 3.09 - No Vocabulary Expansion ($\neg$VE): 48.73 ± 6.74 The observed 6.25-point (11.5%) difference of the JCQA scores between the vocabulary expansion (VE) and non-vocabulary expansion ($\neg$VE) of the Llama-2-JA-7b models does not indicate a substantial impact because of the instability and large standard deviation on the dataset. Therefore, considering these evaluation results and the observed instability, we regard that the impact of vocabulary expansion on overall performance is minor. In the camera-ready version, we will include these elaborations in the Appendix.

Authorsrebuttal2024-06-04

Initiating the discussion

We sincerely thank the reviewer for the valuable suggestions. We hope our response has thoroughly addressed all concerns and that the reviewer can consider improving their score based on our response. There is nothing we want to add to the response, but we are also happy to discuss with the reviewer if they still have a concern.

Reviewer ATKc2024-06-04

Thank you for your clarifications. I think that the extra page for the camera-ready will benefit the readability and understanding of the paper.

Reviewer rjny2024-06-05

Acknowledgement of Rebuttal

Thank you for addressing my review comments - as discussed by other reviewers, the paper may benefit from some portions of the appendix moved into the main paper if afforded an extra page upon acceptance.

Program Chairsdecision2024-07-10

Decision

Accept

© 2026 NYSGPT2525 LLC