💬 NLP

iBERT: Interpretable Embeddings via Sense Decomposition

The paper introduces iBERT, an interpretable encoder that represents tokens as sparse, non-negative mixtures of context-independent sense vectors to modularly expose and control discriminative linguistic cues, achieving superior performance in style-focused tasks while maintaining competitive results in authorship verification.

Original authors: Vishal Anand, Milad Alshomary, Kathleen McKeown

Published 2026-01-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Vishal Anand, Milad Alshomary, Kathleen McKeown

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a magical translator that doesn't just turn one language into another, but instead breaks every sentence down into its building blocks so clearly that you can see exactly why it sounds the way it does.

That is essentially what iBERT does.

The Problem: The "Black Box"

Most modern AI language models (like the ones powering chatbots or search engines) work like a black box. You feed them a sentence, and they spit out a single, dense number (a vector) that represents the meaning.

The Analogy: Imagine a smoothie. You know it tastes like "strawberry-banana," but you can't tell which part is the strawberry and which is the banana. If you want to remove the banana taste, you can't just pick it out; you have to throw away the whole smoothie and start over.
The Issue: In AI, this means we can't easily tell if a sentence sounds "sarcastic" or "formal" because those traits are mixed up with the actual meaning of the words. We can't edit the style without breaking the meaning.

The Solution: The "Lego" Approach

The researchers built iBERT (interpretable-BERT) to solve this. Instead of a smoothie, iBERT builds sentences out of Lego bricks.

Sense Vectors (The Bricks): When iBERT reads a word, it doesn't just give it one meaning. It breaks it down into a mix of 8 different "senses" (like 8 different colored Lego bricks).
- One brick might represent "emoji usage."
- Another might represent "sarcasm."
- Another might represent "formal grammar."
- Another might represent "using all caps."
Sparse Mixture (The Build): For any given word, iBERT only uses a few of these bricks. It says, "This word is 80% 'formal' and 20% 'noun,' but 0% 'sarcasm'."
The Result: The final sentence isn't a blurry smoothie; it's a clear structure where you can see exactly which bricks were used to build the "style" and which built the "meaning."

What Can You Do With This?

Because the AI built the sentence out of separate, labeled bricks, you can do things that were impossible before:

The "Style Edit": Imagine you have a sentence that is too informal. With iBERT, you can find the specific "informal" brick and swap it for a "formal" brick. The meaning of the sentence stays the same, but the tone changes. It's like changing the color of a specific Lego piece without taking the whole tower apart.
The "Style Detective": You can ask the AI, "Why did you think this text was sarcastic?" The AI can point directly to the specific "sarcasm" bricks it used, rather than giving a vague guess.
The "Authorship Check": The paper tested if iBERT could tell who wrote a text (authorship verification). It did just as well as the best existing models, proving that even though it separates style from meaning, it doesn't lose the ability to recognize a writer's unique "fingerprint."

How Well Does It Work?

The researchers tested iBERT on tasks specifically designed to measure how well a model understands style (like tone, formality, and humor) versus just meaning.

The Score: iBERT beat the previous best models by a significant margin (about 8 points) on these style tasks.
The Trade-off: It is slightly slower to run (about 2% slower) because it has to do the extra work of sorting the sentence into these different "senses," but the paper argues this small cost is worth it for the clarity and control it provides.

The Bottom Line

The paper claims that iBERT is a new kind of AI encoder that is transparent by design. It doesn't hide how it thinks; it exposes its thinking process as a mix of understandable parts.

It is not just a style model: The authors emphasize that while they tested it on style, the system is built to separate any distinct signals in language, whether they are about style, meaning, or something else entirely.
It is not magic: The paper admits that if you don't give it good training data (specifically data that clearly separates style from meaning), it won't work as well. It needs the right "ingredients" to learn how to build with its Lego bricks.

In short, iBERT turns the AI's "black box" into a glass box, letting us see, touch, and rearrange the specific parts that make language sound the way it does.

Technical Summary: iBERT: Interpretable Embeddings via Sense Decomposition

Problem Statement
Current neural encoders, such as SBERT and SimCSE, produce dense vector representations that excel at semantic retrieval but lack transparency. These models offer no clear mechanism to control or interpret how style and meaning are encoded, limiting their reliability in domains requiring representational transparency, such as authorship verification, tone-controlled generation, and stylistic moderation. Existing interpretability methods are often post-hoc (e.g., probing, saliency maps) and applied to representations not designed for interpretability, yielding incomplete or unfaithful explanations. Furthermore, prior stylistic representation learning often relies on proxy supervision (e.g., authorship identity), resulting in entangled embeddings where content and style are inextricably mixed. The paper asks: Can we design an encoder where semantic and stylistic representations are made explicit and controllable within the embedding space by design, rather than via post-hoc analysis?

Methodology
The authors propose iBERT (interpretable-BERT), an encoder architecture designed to produce inherently interpretable and controllable embeddings.

Architecture: iBERT encodes each input token as a sparse, non-negative mixture over $k$ $k$ context-independent sense vectors. Unlike standard dense embeddings, the model decomposes token representations into $k$ $k$ distinct "senses" (e.g., $k=8$ $k = 8$ in the experiments).
- Token Encoding: A token $x_i$ is mapped to a standard embedding, then projected into $k$ senses. A transformer encoder produces contextual states, which are projected into sense-specific queries and keys to compute mixture weights ( $\alpha_{\ell,i,j}$ ). The final token representation $o_i$ is a convex mixture of these context-independent sense vectors.
- Sentence Pooling: To generate sentence embeddings, the model applies structured pooling over token-level sense activations. The paper defines three variants:
  - iBERT-v1 (Mean Pooling): Uniform averaging of all token representations.
  - iBERT-v2 (Top-Sense Pooling): Identifies the dominant sense for a sequence and retains only that sense across all tokens (hard selection).
  - iBERT-v3 (Softmax Blend): Interpolates between v1 and v2 using a temperature parameter $\tau$ to control the sharpness of the sense selection (soft blending).
Training Pipeline:
- Stage 0 (MLM Pretraining): iBERT is pretrained from scratch on 5% of the FineWeb corpus (750B tokens) using Masked Language Modeling (MLM) to instantiate meaningful sense vectors. This yields a 171M-parameter encoder.
- Stage 1 (Style Representation Learning): The model is fine-tuned using contrastive triplets ( $\langle s, s^+, s^- \rangle$ ) targeting stylistic variation. The training data includes synthetic triplets from StyleSynth (Patel et al., 2025) with controlled lexical/syntactic variation and author-labeled triplets from Wegmann et al. (2022). The InfoNCE loss is used to encourage closeness between stylistically similar inputs while pushing apart dissimilar ones.
Baselines: To ensure a fair comparison, the authors train a standard BERT-base model from scratch on the same data and fine-tune it with the same contrastive objectives (creating BERT-SD, BERT-WG, and BERT-WG+SD) to serve as dense baselines.

Key Contributions

iBERT Architecture: The introduction of an encoder with controllable, decomposable representations that produce interpretable token and sentence embeddings by design.
Performance: Demonstration of strong performance across three benchmarks (STEL, SoC, and PAN), achieving up to +8% gains in Style Representation Effectiveness (STEL) over dense baselines while maintaining competitive performance on authorship verification.
Novel Interpretability: The ability to probe sense vectors for specific stylistic traits, ablate senses to identify attribution, and edit embeddings for controlled style transfer. The paper highlights that specific style attributes (e.g., emoji use, sarcasm, lexical choices) are assigned to specific sense vectors.

Results

STEL (Style Representation Effectiveness): iBERT models significantly outperform BERT baselines. Under direct-style supervision (StyleSynth), iBERT-v3-1 achieved 38.3% STEL (128 tokens), a +6.7 point gain over the BERT baseline. iBERT-v3 variants (soft blending) generally outperformed the hard-selection iBERT-v2, suggesting that soft sense weighting bridges the gap between uniform blending and strict sparsity.
SoC (Style-or-Content): iBERT-v3-1 achieved the highest separability of stylistic polarity (92.8% accuracy), demonstrating that the model can preserve polarity sensitivity while maintaining interpretability.
PAN (Authorship Verification): As a proxy-style task, iBERT variants matched or slightly outperformed BERT on 128-token inputs. However, on 512-token inputs, BERT slightly outperformed iBERT (61.92% vs. 61.07%), likely because BERT's tight coupling of semantic and stylistic features benefits more from extended context. iBERT maintained competitive performance even when supervision was noisy (author identity only), though it struggled more than BERT when style and content were heavily entangled without explicit style labels.
Interpretability Analysis: Probing revealed that individual senses specialize in distinct stylistic clusters (e.g., Sense $\ell=2$ aligns with affective tone like sarcasm; $\ell=0$ with surface markers like emojis). Ablation studies confirmed that removing a style-aligned sense drastically reduced the separability of that specific style (up to 66% distance reduction) while minimally affecting unrelated styles, validating the locality of the edits.

Significance and Claims
The paper claims that iBERT represents a shift from post-hoc interpretability to architecturally grounded interpretability. By design, iBERT decomposes discriminative signals (whether semantic or stylistic) into modular sense vectors.

Modularity: The model allows for targeted edits in the embedding space (e.g., changing formality or removing emojis) without disrupting unrelated attributes, a feat difficult in dense encoders.
Generalization: While evaluated on stylistic tasks, iBERT is not limited to style modeling. Its structural modularity is designed to interpretably decompose whichever discriminative signals are present in the data, enabling generalization even when supervision blends semantic and stylistic factors.
Trade-offs: The authors acknowledge a tradeoff: the modular structure can lead to slight performance gaps compared to dense encoders in tasks requiring deep contextual integration (e.g., long-form semantic drift). Additionally, sense specialization is sensitive to sequence length, and the current work focuses on English.

The authors conclude that iBERT offers a step toward embedding models that are not only powerful but inherently understandable, serving as a viable backbone for transparent, modular NLP pipelines.