xCoT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning (2401.07037v1)

Abstract: Chain-of-thought (CoT) has emerged as a powerful technique to elicit reasoning in LLMs and improve a variety of downstream tasks. CoT mainly demonstrates excellent performance in English, but its usage in low-resource languages is constrained due to poor language generalization. To bridge the gap among different languages, we propose a cross-lingual instruction fine-tuning framework (xCOT) to transfer knowledge from high-resource languages to low-resource languages. Specifically, the multilingual instruction training data (xCOT-INSTRUCT) is created to encourage the semantic alignment of multiple languages. We introduce cross-lingual in-context few-shot learning (xICL)) to accelerate multilingual agreement in instruction tuning, where some fragments of source languages in examples are randomly substituted by their counterpart translations of target languages. During multilingual instruction tuning, we adopt the randomly online CoT strategy to enhance the multilingual reasoning ability of the LLM by first translating the query to another language and then answering in English. To further facilitate the language transfer, we leverage the high-resource CoT to supervise the training of low-resource languages with cross-lingual distillation. Experimental results on previous benchmarks demonstrate the superior performance of xCoT in reducing the gap among different languages, highlighting its potential to reduce the cross-lingual gap.

• The paper introduces xCoT, integrating cross-lingual instruction tuning with chain-of-thought reasoning to enhance multilingual NLP.
• It employs a novel Random-CoT strategy and cross-lingual in-context few-shot learning to bridge high-resource and low-resource languages.
• Experimental results show up to a 15% performance improvement on benchmarks like MGSM, underscoring its superior reasoning capabilities.

xCoT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning

The paper "xCoT: Cross-lingual Instruction Tuning for Cross-lingual Chain-of-Thought Reasoning" introduces a novel approach to employ chain-of-thought (CoT) reasoning across multiple languages using cross-lingual instruction fine-tuning. This technique seeks to bridge the performance gap between high-resource languages, such as English, and low-resource languages, enhancing multilingual NLP tasks.

Introduction to Cross-lingual CoT

Chain-of-thought prompts have proven effective in eliciting reasoning within LLMs to solve complex tasks. However, CoT primarily exhibits success in high-resource languages, neglecting multilingual settings where performance significantly drops. This paper proposes xCoT, a framework designed to harness cross-lingual instruction tuning (xICL) for generating coherent language representations and facilitating knowledge transfer from resource-rich to resource-poor languages.

Figure 1: Illustration of xCoT. The cross-lingual instruction tuning is used to align representations of different languages.

xCoT Framework

The xCoT framework leverages cross-lingual instruction data, referred to as xCoT-Instruct, encouraging semantic alignment among multiple languages. It comprises a unique processing approach for multilingual alignment: cross-lingual in-context few-shot learning (xICL). This technique mixes language tokens within example queries to enhance multilingual reasoning capabilities. Furthermore, the framework employs a randomly online CoT strategy (Random-CoT) that translates queries into other languages, promoting multilingual reasoning and enabling answers in English.

Figure 2: Overview of xCoT. The cross-lingual in-context few-shot learning (xICL) encourages multilingual alignment in instruction tuning, where the query in the example is mixed with different language tokens. During multilingual instruction tuning, the randomly online CoT strategy (Random-CoT) is used to promote the multilingual reasoning ability of LLM and then answer in English. Finally, we leverage the high-resource CoT to supervise the training of low-resource languages with cross-lingual distillation.

Methodology

Cross-lingual Instruction Tuning

The xCoT-Instruct dataset is developed by translating English instruction data into various target languages, supplemented by cross-lingual in-context examples. The approach involves swapping snippets of the source language in examples with translated tokens of target languages. During multilingual instruction tuning, Random-CoT asks the model to convert queries into different languages prior to generating responses in English. The enhancement of reasoning in non-English queries is further supported through cross-lingual distillation of high-resource CoT outcomes.

Implementation Details

The paper implements xCoT using Llama-2 and Bloom models, experimenting with a variety of embeddings and language datasets. The framework requires fine-tuning of LLMs on a mix of xCoT-Instruct data, leveraging the described Random-CoT strategy for consistent linguistic performance across languages.

Evaluation and Results

xCoT's performance is evaluated on multilingual benchmarks like MGSM and MSVAMP, involving languages like German, Chinese, and Spanish. The experimental results demonstrate state-of-the-art performance with significant improvements over baseline models, achieving an average margin enhancement of 15%. The strategically interactive fine-tuning showcased xCoT's capacity to elevate low-resource language understanding via strategic multilingual instruction data synthesis and execution.

Figure 3: (a) and (b) are representations of Llama-7B and our method from the last decoder layer. Each color denotes one language (11 languages in MGSM).

Analysis and Future Directions

Analysis of the results emphasizes the effectiveness of cross-lingual transfer through xCoT, proving valuable across different linguistic scenarios. Evaluations reflect the model's adaptation and success when handling multiple languages with distinct reasoning paths. Future directions could explore more robust training datasets and the integration of additional resource-rich languages to enhance the practical applications of xCoT in global communicative interfaces.

Conclusion

This paper presents xCoT, an innovative cross-lingual framework that integrates instruction tuning and CoT reasoning, successfully reducing performance disparities across multilingual NLP frameworks. The results confirm the advantage of adopting a fine-tuned multilingual strategy, propelling LLMs towards superior cross-linguistic reasoning capabilities. The framework lays a strong foundation for advancing multilingual AI beyond traditional language limitations.