💻 computer science

ProtoSiTex: Learning Semi-Interpretable Prototypes for Multi-label Text Classification

ProtoSiTex is a semi-interpretable framework that advances fine-grained multi-label text classification by employing a dual-phase training strategy and hierarchical loss to learn adaptive, overlapping prototypes, achieving state-of-the-art performance while providing human-aligned explanations.

Original authors: Utsav Kumar Nareti, Suraj Kumar, Soumya Pandey, Soumi Chattopadhyay, Chandranath Adak, Sankha Subhra Mullick

Published 2026-05-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Utsav Kumar Nareti, Suraj Kumar, Soumya Pandey, Soumi Chattopadhyay, Chandranath Adak, Sankha Subhra Mullick

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

The Big Problem: The "Black Box" Dilemma

Imagine you have a super-smart robot that can read thousands of hotel reviews and tell you if a guest was happy about the food, the room, or the staff. It gets the answers right almost every time. But, if you ask the robot why it thinks the food was good, it just says, "Because I said so." It's a "black box."

In the real world, we don't just want the answer; we want to know the reason. We need to see the specific sentence in the review that made the robot decide the food was great.

Existing "explainable" models are like a teacher who only grades your entire essay as "Good" or "Bad." They can't point to the specific sentence where you made a great point. Other models try to look at sentences, but they struggle when a review has mixed feelings (e.g., "The room was huge, but the bed was uncomfortable"). They get confused by the overlap.

The Solution: ProtoSiTex (The "Smart Librarian")

The authors created a new system called ProtoSiTex. Think of it as a Smart Librarian who doesn't just read books; they organize them into specific "theme boxes" (prototypes) and can explain exactly which page in a book matches which theme.

Here is how it works, step-by-step:

1. Breaking it Down (The Subsentence)

Most systems read a whole paragraph at once. ProtoSiTex is like a detective who uses a magnifying glass. It breaks a review down into tiny chunks called subsentences (like phrases separated by commas).

Analogy: Instead of reading a whole chapter to find a plot twist, it highlights the exact sentence where the twist happens.

2. The "Theme Boxes" (Adaptive Prototypes)

The system creates a set of "Theme Boxes" (called Prototypes). These aren't pre-written labels; the system learns them on its own.

Analogy: Imagine a librarian who creates a box labeled "Cozy Atmosphere." Inside this box, they don't just put the word "cozy." They put actual examples of sentences from past reviews that feel cozy (e.g., "The fireplace was crackling," "The blankets were soft").
The Magic: If a new review says, "The room was small but felt warm," the system can see that "felt warm" fits into the "Cozy" box, even if the word "cozy" isn't there.

3. The Two-Step Dance (Dual-Phase Training)

To make these Theme Boxes perfect, the system does a special two-step training dance:

Step A (The Discovery Phase): The system looks at thousands of reviews without any labels. It groups similar sentences together to build its "Theme Boxes." It makes sure the boxes are different from each other (so the "Food" box doesn't look like the "Room" box).
Step B (The Teaching Phase): Now, the system is shown the correct answers (e.g., "This sentence is about Food"). It adjusts its boxes to make sure they match the labels perfectly.
Analogy: First, the librarian sorts books by how they feel (Discovery). Then, a teacher comes in and says, "Actually, this book belongs in the 'History' section, not 'Fiction'." The librarian adjusts the shelves (Teaching).

4. The "Chain of Command" (Hierarchical Loss)

The system checks its work at three levels to ensure it's not making mistakes:

Tiny Level: Does this specific phrase match the Theme Box?
Medium Level: Do all the phrases in this sentence make sense together?
Big Level: Does the whole review make sense?

Analogy: It's like a manager checking a report. They check the spelling of individual words, the grammar of the sentences, and the overall logic of the story. If the story makes sense but a word is wrong, the system catches it.

Why is it "Semi-Interpretable"?

The paper calls this Semi-Interpretable.

The Good News: You can see exactly which sentence matched which Theme Box. If the system says "The staff was rude," you can see the specific sentence "The staff was rude" and the "Rude Staff" theme box it matched. It's transparent.
The "Semi" Part: The math inside the system (how it calculates the similarity between sentences and boxes) is still complex and hidden. It's not a simple "If-Then" rule you can read like a manual. It's a smart, learned intuition.

The Results: Did it Work?

The authors tested this on:

IMDb: Movie reviews (Good vs. Bad).
TweetEVAL: Tweets with emotions (Joy, Anger, etc.).
A New Dataset (HR): They created a new set of Hotel Reviews where every tiny phrase was labeled (e.g., "Room," "Food," "Price").

The Outcome:

Accuracy: It performed just as well as the most powerful, complex "black box" AI models (like the big language models everyone uses).
Explanation: Unlike those black boxes, ProtoSiTex could point to the exact reason for its decision.
Handling Conflicts: It was great at handling mixed reviews (e.g., "Great food, terrible service") because it could separate the "Food" theme from the "Service" theme at the sentence level.

Summary

ProtoSiTex is a new way to teach computers to read text. Instead of guessing the whole meaning at once, it breaks text into small pieces, matches them to learned "Theme Boxes," and checks its work at every level. It gives us the accuracy of a super-computer with the clarity of a human explaining their reasoning.

Technical Summary: ProtoSiTex

Problem Statement
The rapid proliferation of user-generated text necessitates models capable of fine-grained, multi-label classification with inherent interpretability. While transformer-based language models achieve state-of-the-art performance, they operate as "black boxes," offering limited transparency. Existing post-hoc explanation methods (e.g., LIME, SHAP) often lack faithfulness and consistency. Conversely, inherently interpretable Prototype-Based Networks (PBNs) typically operate at coarse granularities (sentence or document level) and are largely restricted to single-label tasks. This limitation prevents them from handling real-world scenarios where texts contain overlapping, conflicting, or multi-faceted sentiments (e.g., a review praising the view but criticizing the staff). Furthermore, there is a scarcity of datasets providing subsentence-level multi-label annotations, hindering the evaluation of fine-grained interpretability.

Methodology: ProtoSiTex
The authors propose ProtoSiTex, a semi-interpretable framework designed for fine-grained multi-label text classification. The term "semi-interpretable" denotes a model that provides faithful, prototype-based explanations for its predictions while relying on latent neural mechanisms (e.g., attention, embeddings) that are not fully transparent. The architecture consists of three core stages:

Subsentence Representation: The input document is segmented into sentences and further into subsentences using punctuation cues. A language model encoder (specifically RoBERTa-Large in experiments) generates context-aware embeddings for these subsentences.
Dual-Phase Alternate Training: The model employs a unique training strategy alternating between two phases:
- Unsupervised Prototype Discovery (Clustering Phase): A learnable prototype layer, coupled with Multi-Head Attention (MHA), discovers semantically coherent and diverse prototypes. This phase uses proximity-based losses (ensuring subsentences align with nearest prototypes) and structure regularization losses (enforcing sparsity and diversity among prototypes) to prevent prototype collapse and redundancy.
- Supervised Classification: The discovered prototypes are mapped to class labels. This phase utilizes a hierarchical loss function that enforces consistency across three levels: subsentence, sentence, and document. It dynamically constructs ground-truth labels at the prototype level by aggregating subsentence labels and propagates predictions upward through binary mapping matrices ( $M_1$ for document-to-sentence, $M_2$ for sentence-to-subsentence).
Inference and Explanation: During inference, the model predicts document-level labels via hierarchical aggregation. Crucially, it generates explanations by linking each subsentence to its closest prototype and retrieving the top- $K$ most similar training subsentences associated with that prototype. These exemplars are summarized (using an LLM) to provide human-readable justifications for the prediction.

Key Contributions

Semi-Interpretable Framework: ProtoSiTex is the first prototype-based framework to perform fine-grained multi-label classification at the subsentence level. It uses adaptive prototypes and multi-head attention to disentangle overlapping and conflicting semantics, a capability lacking in prior PBNs.
Dual-Phase Optimization & Hierarchical Supervision: The proposed alternate training strategy decouples unsupervised prototype discovery from supervised classification, improving generalization and reducing overfitting. The hierarchical loss function ensures semantic consistency across granularities, enabling structured label propagation.
New Benchmark Dataset (HR): The authors introduce a Hotel Reviews (HR) dataset containing 3,002 reviews annotated at the subsentence level with 8 multi-label aspects (e.g., Room & Amenities, Food, Location). This dataset includes ground-truth spans for interpretability evaluation, addressing a gap in existing benchmarks.
Adaptive Semantic Aggregation: The model establishes explicit alignment between subsentence evidence, adaptive prototypes, and document predictions through differentiable aggregation, facilitating traceable explanations without sacrificing predictive performance.

Experimental Results
Experiments were conducted on the proposed HR dataset, as well as public benchmarks IMDb (binary sentiment) and TweetEVAL (multi-class emotion).

Performance: ProtoSiTex achieved state-of-the-art performance across all datasets. On the HR dataset, it attained a Macro-F1 score of 87.22% and Balanced Accuracy of 83.68%, outperforming strong black-box baselines (e.g., RoBERTa-Large, DeBERTaV3) and existing prototype-based models (e.g., GAProtoNet, ProtoLens).
Comparison: While matching the accuracy of large black-box transformers on simpler tasks (IMDb), ProtoSiTex significantly outperformed them on the complex, multi-label HR dataset. It also surpassed closed-source and open-source LLMs (e.g., GPT, LLaMA) in multi-label settings, demonstrating superior handling of label dependencies and fine-grained semantics.
Ablation Studies: Analyses confirmed that the dual-phase alternate training, multi-head attention for similarity, and random prototype initialization were critical to performance. The study also noted that masked-based encoders (like RoBERTa) were more effective than autoregressive models for this specific architecture due to their bidirectional context and dense embedding spaces.

Significance and Claims
The paper claims that ProtoSiTex establishes a robust solution for semi-interpretable multi-label text classification. Its primary significance lies in bridging the gap between predictive accuracy and interpretability in complex, real-world scenarios. By operating at the subsentence level, the model can faithfully explain why a specific label was assigned, even when multiple conflicting sentiments exist within a single document. The authors assert that the framework offers "faithful, human-aligned explanations" while maintaining competitive performance, making it suitable for applications requiring trust, fairness, and regulatory compliance. The introduction of the HR dataset is highlighted as a critical contribution for enabling empirical benchmarking of fine-grained, interpretable multi-label models.