How do Large Language Models Handle Multilingualism?

Summary

This paper delves deeper into how LLMs handle multilingualism. The authors hypothesized a three-stage multilingual workflow called MWork: understanding, task-solving, and generating and which language(s) become essential in each stage. To verify their proposed workflow, they experimentally identify language-specific parameters and selectively deactivate them within different structures so this allows to assess the functionality of corresponding structures and enables the hypothesis. To do so, the authors develop a novel approach called Parallel Language-specific Neuron Detection (PLND). Without requiring labeled data, PLND can identify the language specific neurons. Following the PLND, the authors successfully identify the language-specific neurons which account for only 0.13% of all neurons. Their extensive results report that, by deactivating those neurons, the multilingual task performance could significantly dropped. This paper tackles an important research question of LLMs: how do large language models handle multilingualism? To address this, they deliberately design a new approach of PLND and successfully identifies a language-specific neurons in the LLM models. Along with their MWork hypothesis, they carefully conducted experiments with multiple languages with different scrips and verified their hypothesized three-stage multilingual workflow. This paper develops a useful model analysis tool of PLND. Their hypothesis is well explored in multiple languages. Technically well sound.

Strengths

- well-organized paper. clear presentation. - Technically sound by extensive experiments with different natural language understanding tasks with diverse language sets

Weaknesses

- no major weakness though

Questions

- Regarding those language-specific neuron, are there any patterns or trends related to language family grouping or similarity? Such analysis or findings would be interesting to NLP researchers. - Have you ever tried this approach against bigger models with > 7B parameters? - Since language-specific neurons are identified via the proposed framework, can we do model compression by preserving those specific parameters without losing multilingual capability? If you have examined this perspective, can you please share the results?

Rating

7

Confidence

3

Soundness

3

Presentation

4

Contribution

4

Limitations

N/A

Summary

The paper presents two key contributions. One is Parallel Language-specific Neuron Detection, a method of identifying elements of multilingual LLMs that are responsible for handling particular languages; the method only requires unlabeled text data in each language in order to detect these neurons, which makes it cheap and efficient. The second contribution is an insight into the workflow of multilingual LLMs (referred to in the paper as MWork), claiming that LLMs process input in three steps: understanding, task solving and output generation; task solving further splits into retrieval and thinking. Also, the thinking step is shown to be happening in English, while understanding, knowledge retrieval and generation are handled multilingually / specifically to the language of the input. Authors use PLND to verify that the models chosen for the experiments actually follow this workflow and conduct thorough experiments on two models about a dozen languages other than English, including high-resource and low-resource examples.

Strengths

- Both key contributions (PLND and MWork) are of fundamental relevance to the field of LLM interpretability and analysis - Presented experiments are rigorous and thorough - The paper is written clearly and although the subject is complex and multi-faceted, the authors do a good job of introducing the necessary concepts and definitions and explaining the rationale behind their work as well as the conducted experiments and analysis of their results.

Weaknesses

W1. Most importantly, the claims are very broad ("LLMs are this / LLMs do that"), however this is verified on two very similar models: the included Mistral and Vicuna models both have 7 billion parameters, which puts them on the modest end of modern LLMs. How can we be sure that models with more parameters (13 / 30 / 70 / ...) behave in a similar manner without testing?, what about the influence of the context length on the performance? This is even mentioned in the limitations section of the paper: I am not suggesting that the authors run an additional set of expensive experiments -- I am suggesting that the paper text should be adjusted to reflect the actual findings of the presented experiments and would avoid bold claims about "all LLMs". W2. Another weakly supported claim is that 400 documents was enough for language-specific neuron tuning, simply because tuning on 800 documents did not yield better results. I find that this is a single instance and without running more comparisons and more thorough comparisons this claim does not hold and should be adjusted in the text. This is especially important since improvements of 2.3% (low-resource) and 3.6% (high-resource) are small-scaled improvements and bigger amounts of language-specific data could be expected to yield bigger improvements. Again, rather than running tons of more experiments for a single paper I instead suggest that the paper text should be adjusted (e.g. calling it a "pilot study" or "preliminary indications/results" -- PLND and MWork themselves are already worthy contributions). W3. Not a strong weakness, but minor typos: - row 101: "With a set of n corpus" --> "With a set of n corpora", same on row 103 ('corpus' in plural is 'corpora') - row 319: "one brunch of work" --> "one branch of work"

Questions

Q1. There seems to be a leap from the experimental results to conclusions about how LLMs behave. Are the presented conclusions (about understanding->thinking->generating) the only possible explanation of the obtained results, or could there be other explanations? Q2. In your work the language-specificity is defined via that language only (importance is higher than a threshold). Would you expect a stronger performance of PLND if the language-specificity were defined contrastively, that is high impact on one language and low impact on other languages?, in other words, could a neuron be important to several languages?, and what would be the repercussions of such neurons on your work. Q3. In this work you focus on the "English / non-English" setting. Is this driven by the fact that English is the most abundant language in the training data?, and what would you expect the behavior of the models to be if another language were the most frequent, would the thinking be done in that language?, what about a more balanced setting where more than one language are dominant in the data (for example 40% English, 40% Chinese and other less frequent languages)? Q4. Concerning language-specific tuning, hypothetically, is multilingual performance of LLMs bound by English-centerdness? that is, can one achieve better performance of an LLM in a non-English language if that language's data is included into pre-training from the start, rather than being English-centered and tuned to a non-English language? Q5. Modern language models are often trained with programming code added to natural language texts. What would you expect your analysis to show in case the input is programming code or comments? Q6. Could full fine-tuning to a non-English language achieve better results than just tuning that language's language-specific neurons?

Rating

6

Confidence

4

Soundness

3

Presentation

4

Contribution

4

Limitations

All good.

Summary

This paper proposes MWork, a workflow to study how large language models handle multilingual inputs. The main idea is to detect English and non-English neurons by probing their value differences and performance gaps. Based on the results, they argue that there are three steps in the workflow: understanding, task solving, and generating. They also use experiments to support their hypothesis.

Strengths

- They propose an interesting way to understand the mechanism of large language models.

Weaknesses

- The authors classify neurons based on English and non-English, which is not very convincing to me. I believe that, beyond these two categories, there should be some neurons that are helpful for **all** languages. These neurons can be used for general understanding and logical reasoning. However, this paper ignores this aspect. See more explanations below. - For neuron detection, when computing the impact, I am curious why you do not remove the overlapping neurons. In this way, you can detect the real language-specific ones. Specifically, in Appendix C, although the authors argue that the intersection with English from other languages is relatively limited based on rows, if we look at the columns, it seems that all the language-specific neurons are covered by English ones, suggesting that they are actually not language-specific. It is very likely that these neurons are essential for general understanding and are language-agnostic. - It is not clear how English and non-English tokens are detected in Figure 1. Do you determine this by applying a decoder to decode from the hidden representations? - The authors argue that large language models will use English tokens for task solving with multilingual inputs based on the interpretation of Figure 1. However, Figure 1 also shows that models will use non-English tokens for task solving with English. I don’t think this is reasonable. From my perspective, the way to classify tokens or neurons is not accurate, and the ignorance of language-agnostic neurons makes this figure unreasonable. - From the experimental results, deactivating language-specific neurons sometimes causes a performance drop for both English and non-English, suggesting the existence of general neurons.

Questions

See above.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes.

Summary

This paper examines how LLMs handle multilingualism. The author proposes a hypothetic workflow (MWork), which suggests that LLMs understand the multilingual query, think in English, and than generate results in the input language. A neuron detection method is proposed to detect language specific neurons. By deactivation some of the neurons, the authors validates the above workflow, and show that the ability for a specific language could be improved by finetuning only language specific neurons.

Strengths

The paper presents an interesting hypothesized workflow and validate it with neuron analysis. The analysis with neuron deactivation is sound enough.

Weaknesses

The analysis of this paper is based on the understanding of layers. However, it might be possible that different tasks or even different instances may have different splitting of layers for different stages in the workflow. How would this affect the analysis? It is still strange to know that 400 documents could improve the language ability of a low-resource languages, especially when 800 documents will not be more helpful.

Questions

See the weakness part.

Rating

7

Confidence

5

Soundness

3

Presentation

3

Contribution

3

Limitations

N/A