Towards Interpretable Neural Network-Based Sentiment Analysis Models: A Bottleneck-Based Latent Dimension Analysis
Year: 2026
Type: dissertation
Abstract: Neural network-based models, particularly those built on deep learning architectures, have achieved superior performance in sentiment analysis tasks compared with traditional sentiment analysis approaches. However, the deep learning-based models function as black boxes, making it difficult to understand how semantic meaning is internally represented.While prior work has focused on feature-level or attention-based explanations, less attention has been given to understanding how sentiment is represented within the internal latent space of these models. This dissertation addresses this gap by proposing a representation-level approach to interpreting neural network-based sentiment analysis models using a bottleneck layer architecture. The study introduces a neural network-based sentiment classification model with a bottleneck layer placed before the final classification layer. This design constrains high-dimensional representations into a small set of latent dimensions, allowing direct analysis of how sentiment-related information is encoded. It also supports examination of independence among dimensions. Multiple configurations with different bottleneck sizes are evaluated to assess the trade-off between predictive performance and dimension independence. The model is trained on Yelp restaurant reviews and compared with the baseline model. The results show that the bottleneck-based model maintains competitive performance. Based on the evaluation of performance and independence among latent dimensions, a configuration with ten latent dimensions is selected for detailed analysis. To interpret the learned representations, an integrated framework is developed that combines correlation analysis, contribution analysis, regression modeling, and qualitative examination of example reviews. These methods are used to examine how each dimension relates to model predictions and to identify the linguistic patterns associated with each dimension. The results show that sentiment prediction emerges from the interaction of multiple latent dimensions that capture distinct linguistic patterns. A small number of dimensions function as primary factors of sentiment polarity, exhibiting strong associations with classifier outputs and prediction decisions. Other dimensions capture discourse structure, narrative styles, and stylistic variations, operating as secondary signals that shape how evaluation is expressed, thereby influencing prediction outcomes. Specifically, in the ten-dimensional bottleneck model, one dimension captures overall evaluative polarity and has the strongest association with model prediction outcomes. Another dimension reflects complaint-related discourse structure and is strongly linked to negative sentiment. Additional dimensions capture experiential narrative, aggregated experiences, interaction with service staff, advice and recommendation language, and variation in expression style. These dimensions are not predefined but emerge from the model during training. The analysis shows that each dimension aligns with specific linguistic features. For example, the narrative dimension is associated with first-person pronouns and past-tense verbs, while the advice dimension is linked to imperative expressions and second-person pronouns. Other dimensions reflect differences in discourse organization and structural complexity. These findings indicate that sentiment prediction depends not only on evaluative or sentiment words but also on how information is structured and communicated. This study demonstrates that architectural design can improve interpretability and provides a framework for analyzing neural network-based sentiment analysis models at the representation level. The bottleneck-based framework introduced in the study is designed to be adaptable across domains and can be applied to other text classification tasks to examine how deep learning models internally organize linguistic information.
Author: Luo, Manman
Institution: University of Wisconsin–Milwaukee
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7164975232