Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Utsav Kumar Nareti, Suraj Kumar, Soumya Pandey, Soumi Chattopadhyay, Chandranath Adak

TL;DR

ProtoSiTex tackles the need for interpretable, fine-grained multi-label text classification by introducing adaptive prototypes learned through a dual-phase training regime and hierarchical supervision. It leverages an LM-based encoder to obtain subsentence representations, aligns them with learnable prototypes via multi-head attention, and derives multi-level predictions with explanations grounded in prototype content. A new hotel-review dataset with subsentence-level multi-label annotations supports evaluation of interpretability, while extensive experiments show ProtoSiTex achieving competitive accuracy alongside faithful, human-aligned explanations. The work demonstrates that semi-interpretable prototype-based models can match black-box performance and provide actionable insights for debugging and trust in real-world, multi-label NLP tasks.

Abstract

The rapid growth of user-generated text across digital platforms has intensified the need for interpretable models capable of fine-grained text classification and explanation. Existing prototype-based models offer intuitive explanations but typically operate at coarse granularity (sentence or document level) and fail to address the multi-label nature of real-world text classification. We propose ProtoSiTex, a semi-interpretable framework designed for fine-grained multi-label text classification. ProtoSiTex employs a dual-phase alternate training strategy: an unsupervised prototype discovery phase that learns semantically coherent and diverse prototypes, and a supervised classification phase that maps these prototypes to class labels. A hierarchical loss function enforces consistency across subsentence, sentence, and document levels, enhancing interpretability and alignment. Unlike prior approaches, ProtoSiTex captures overlapping and conflicting semantics using adaptive prototypes and multi-head attention. We also introduce a benchmark dataset of hotel reviews annotated at the subsentence level with multiple labels. Experiments on this dataset and two public benchmarks (binary and multi-class) show that ProtoSiTex achieves state-of-the-art performance while delivering faithful, human-aligned explanations, establishing it as a robust solution for semi-interpretable multi-label text classification.

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

TL;DR

ProtoSiTex tackles the need for interpretable, fine-grained multi-label text classification by introducing adaptive prototypes learned through a dual-phase training regime and hierarchical supervision. It leverages an LM-based encoder to obtain subsentence representations, aligns them with learnable prototypes via multi-head attention, and derives multi-level predictions with explanations grounded in prototype content. A new hotel-review dataset with subsentence-level multi-label annotations supports evaluation of interpretability, while extensive experiments show ProtoSiTex achieving competitive accuracy alongside faithful, human-aligned explanations. The work demonstrates that semi-interpretable prototype-based models can match black-box performance and provide actionable insights for debugging and trust in real-world, multi-label NLP tasks.

Abstract

The rapid growth of user-generated text across digital platforms has intensified the need for interpretable models capable of fine-grained text classification and explanation. Existing prototype-based models offer intuitive explanations but typically operate at coarse granularity (sentence or document level) and fail to address the multi-label nature of real-world text classification. We propose ProtoSiTex, a semi-interpretable framework designed for fine-grained multi-label text classification. ProtoSiTex employs a dual-phase alternate training strategy: an unsupervised prototype discovery phase that learns semantically coherent and diverse prototypes, and a supervised classification phase that maps these prototypes to class labels. A hierarchical loss function enforces consistency across subsentence, sentence, and document levels, enhancing interpretability and alignment. Unlike prior approaches, ProtoSiTex captures overlapping and conflicting semantics using adaptive prototypes and multi-head attention. We also introduce a benchmark dataset of hotel reviews annotated at the subsentence level with multiple labels. Experiments on this dataset and two public benchmarks (binary and multi-class) show that ProtoSiTex achieves state-of-the-art performance while delivering faithful, human-aligned explanations, establishing it as a robust solution for semi-interpretable multi-label text classification.
Paper Structure (32 sections, 8 equations, 7 figures, 3 tables)

This paper contains 32 sections, 8 equations, 7 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Proposed Methodology
  4. Problem Formulation
  5. ProtoSiTex Architecture
  6. Subsentence Representation via LM-based Encoder
  7. Dual Phase Learning
  8. Clustering-phase for Prototypes Learning
  9. Classification-phase for Learning Multi-granular Document Details
  10. Prototype Projection and Label Classification
  11. Training Strategy of ProtoSiTex
  12. Loss Functions for Classification Phase
  13. Loss Functions for Clustering Phase
  14. Proximity-based Losses
  15. Structure Regularization Losses
  16. ...and 17 more sections

Figures (7)

  • Figure 1: Proposed solution architecture: ProtoSiTex
  • Figure 2: Distribution of eight aspects in HR
  • Figure 3: Performance analysis of ProtoSiTex on HR: (a) with various prototype initializers; (b) without alternate training (w/o AT); (c) across different similarity measures; with a varying number of (d) prototypes, and (e) heads
  • Figure 4: Impact of hyperparameters $\alpha_i$ and $\lambda_i$ on HR
  • Figure 5: Aspect-wise performance of ProtoSiTex on HR
  • ...and 2 more figures