Sentium: Sentiment Evaluation through Neurosymbolic Taxonomy - an Interpretable and Understandable Model
Abstract: Sentiment analysis has seen rapid progress driven by deep learning, but the opaque black-box nature of these models hinders trustworthy deployment in high-stakes domains where interpretability is crucial. We propose \textbf{Sentium} (\textbf{S}entiment \textbf{E}valuation through \textbf{N}eurosymbolic \textbf{T}axonomy, an \textbf{I}nterpretable and \textbf{U}nderstandable \textbf{M}odel), a cognitively-inspired architecture that closely emulates human sentiment comprehension processes. Sentium takes a hybrid approach by combining structured sentiment knowledge with neural models, achieving state-of-the-art performance while maintaining transparency through explicit compositional reasoning over semantic propositions. Compared to state-of-the-art financial language models, Sentium showed substantially lower misclassification rates for predicting true negatives as positive (Sentium=1.97\%; FLANG-BERT \citep{shah2022flue} =6.78\%, FinBERT \citep{araci2019finbert} =10.17\%). The code are available at: https://github.com/anonymous-submission
Paper Type: Long
Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics
Research Area Keywords: Interpretability and Analysis of Models for NLP, Linguistic Theories Cognitive Modeling and Psycholinguistics, Sentiment Analysis Stylistic Analysis and Argument Mining, NLP Applications, Ethics Bias and Fairness
Contribution Types: Model analysis & interpretability, Publicly available software and/or pre-trained models
Languages Studied: English
Submission Number: 4828
Loading