Technical Summary: Attention Sinks in Massively Multilingual Neural Machine Translation

Problem Statement

This paper identifies a systematic artifact in the cross-attention analysis of NLLB-200, Meta's 600-million-parameter massively multilingual Neural Machine Translation (NMT) model. The authors discovered that non-content tokens—specifically end-of-sequence tokens (</s>), language identifier tags (e.g., swh_Latn), and punctuation—absorb the vast majority of cross-attention mass (83–91%).

This phenomenon, termed "attention sinks," distorts interpretability analyses. Because these tokens are structurally ubiquitous (present in every sentence) rather than semantically relevant, they dominate aggregate statistics. Consequently, raw cross-attention similarity metrics severely underestimate true content-level similarity by nearly a factor of two (e.g., 36.7% raw vs. 70.7% filtered). This artifact masks genuine linguistic signals, rendering uncorrected analyses of word alignment, structural correspondence, and cross-lingual generalization unreliable.

Unlike prior work on attention sinks in large autoregressive language models (LLMs) which attribute the phenomenon to position-based bias (e.g., the initial token), the authors identify the mechanism in NLLB-200 as vocabulary-based. The sinks arise from specific token types (language tags and punctuation) that are statistically ubiquitous due to tokenization and vocabulary design, rather than their position in the sequence.

Methodology

To address this artifact, the authors developed and validated a content-only filtering methodology. The pipeline operates as follows:

1. Token Classification: Source tokens are categorized into four types:
   • Language tokens: Matching the NLLB identifier pattern (e.g., [a-z]{3}_[A-Z][a-z]+).
   • Punctuation: A curated set of 32 marks and symbols.
   • Special tokens: Structural tokens (<s>, </s>, <pad>, <unk>).
   • Content tokens: All remaining tokens.
2. Filtering and Renormalization: The attention weights of non-content tokens are zeroed out. The remaining attention distribution is renormalized so that content-token weights sum to 1.0 for each decoder step.
   • Formally, for an attention vector $a$ and a content mask $m$, the filtered vector $a^*$ is calculated as: $a^* = \frac{a \odot m}{\sum_i a_i m_i}$.
3. Implementation: The authors implemented an efficient pipeline using chunked HDF5 reading to process large attention files (1,000 sentences per language) on standard CPU workstations, taking approximately 3–5 minutes per language.

Key Contributions

The paper makes four primary contributions:

1. Discovery and Characterization: A systematic documentation of attention sinks in NLLB-200, quantifying the disproportionate attention absorbed by </s> (78–87%), language tags (1.5–2%), and punctuation, leaving only 9–17% for content tokens across diverse languages.
2. Filtering Methodology: A principled pipeline for removing non-content tokens and renormalizing attention distributions, applicable to the large datasets produced by NLLB-200 inference.
3. Corrected Linguistic Analysis: Application of the filter to four African languages (Swahili, Kikuyu, Somali, Luo) and validation on four non-African languages (German, Turkish, Chinese, Hindi), revealing substantive signals previously masked.
4. Open Toolkit: The release of a filtering toolkit and corrected attention datasets to support reproducible interpretability research.

Key Results

Applying the content-only filter revealed significant shifts in the model's behavior and performance metrics:

• Restored Similarity Metrics: Filtering nearly doubled the absolute similarity values. Teacher-forcing (TF) similarity rose from 36.7% to 70.7%, and generation (Gen) similarity rose from 28.7% to 53.8%.
• Teacher-Forcing vs. Generation Gap: The gap between TF and Gen modes increased from 8.0 percentage points (pp) in raw data to 16.9 pp in filtered data. This indicates that generation mode produces significantly more diffuse and uncertain attention distributions than TF mode, a nuance previously obscured.
• Language Family Clustering: Corrected data revealed distinct patterns based on language families:
  • Bantu (Swahili, Kikuyu): Showed moderate entropy and high local bias, consistent with monotonic alignment due to shared SVO word order with English.
  • Nilotic (Luo): Exhibited the lowest peak attention and local bias, indicating the most distributed patterns.
  • The Somali Paradox: Somali (Cushitic, SOV word order) presented a contradictory pattern masked in raw data: it exhibited the highest entropy (most distributed attention) yet the highest local bias and tied-highest peak attention. The authors hypothesize this reflects a strategy where the model maintains a monotonic core (attending strongly to local tokens) while spreading residual attention broadly to handle the SVO-to-SOV reordering challenge.
• Universality: The sink effect was confirmed across 8 languages spanning 7 language families, 3 scripts, and both SVO and SOV word orders, with non-content tokens consistently absorbing 80–91% of attention mass.

Significance and Claims

The authors assert that their findings fundamentally alter the reliability of prior interpretability studies on NLLB-200 and similar models. They claim that any study computing aggregate attention statistics or alignment metrics without filtering non-content tokens is reporting values inflated by approximately 95% relative to content-only baselines.

The paper argues that content-only filtering must become a standard preprocessing step for NMT interpretability. By removing the "noise" of structural tokens, researchers can recover genuine linguistic signals, such as the specific challenges of SOV word order reordering and the sensitivity of attention mechanisms to decoding errors. The authors conclude that their toolkit enables a more accurate theoretical account of cross-lingual generalization and failure modes in massively multilingual models.