Abstract
This paper introduces a quantum-inspired approach to sentiment analysis by representing words as complex numbers rather than standard real-valued vectors. By encoding a word's contextual meaning in the amplitude and its sentiment polarity in the phase, the proposed Lexicon-extended Quantum Probability Driven Network (LQPDN) successfully decouples semantic similarity from emotional valence. Experimental results demonstrate that this representation leads to state-of-the-art performance on standard sentiment benchmarks, particularly in fine-grained classification tasks.
Introduction
In traditional Natural Language Processing (NLP), word embeddings like Word2vec or GloVe map words into a high-dimensional real-valued space based on the distributional hypothesis: words appearing in similar contexts should have similar vectors. However, this creates a significant hurdle for sentiment analysis. Antonyms like "good" and "bad" or "optimistic" and "pessimistic" often appear in nearly identical linguistic contexts (e.g., "The outlook was ____"). Consequently, their vectors end up very close in vector space, making it difficult for sentiment classifiers to distinguish between polar opposites.
Methodology
The authors propose the Lexicon-extended Quantum Probability Driven Network (LQPDN), which moves word representation from the real plane to the complex plane, drawing inspiration from the mathematical framework of quantum mechanics.
Complex-valued Representations (CR)
Each word is represented as a vector of complex numbers. A complex number is defined by an amplitude ($r$) and a phase ($\theta$):
- Amplitude ($r$): Captures traditional distributional and contextual information.
- Phase ($\theta$): Specifically reserved to encode the sentiment polarity of the word.
Quantum Fusion and Lexicon Integration
The model utilizes a "density matrix" ($\rho$) to fuse word information within a sentence, a concept from quantum mechanics used to describe the state of a system. To refine the sentiment encoding, the authors integrate a sentiment lexicon, assigning higher weights to sentiment-carrying words during training. Additionally, a sub-network (SWPN) is employed as an auxiliary task to predict individual word polarity based solely on phase embeddings, forcing the phase component to learn sentiment features effectively.
Experimental Results
The LQPDN model was evaluated on three widely recognized sentiment analysis benchmarks: the Movie Review (MR) dataset, and the Stanford Sentiment Treebank in both binary (SST-2) and fine-grained 5-way (SST-1) formats.
As shown in the results, LQPDN outperformed several strong baselines, including standard CNNs and specialized sentiment models like CNN+SentiNet. The improvement is most pronounced on the SST-1 dataset, which is the most difficult task due to its fine-grained labels, suggesting that the complex-valued representation is superior at capturing subtle nuances in sentiment.
Discussion and Key Insights
The primary innovation of this work is the decoupling of semantic and sentiment information through separate dimensions. By using the phase of a complex number, the authors create a dedicated channel for sentiment that does not interfere with the word's semantic meaning.
This approach effectively resolves the "Good/Bad" paradox. In standard real-valued embeddings, "optimistic" and "pessimistic" are identified as highly similar. However, in the complex-valued embedding, their similarity is significantly reduced because their phases point in different directions, even while their amplitudes (representing context) remain similar. This demonstrates that quantum probability frameworks can be powerful tools for modeling the non-linear combinations and ambiguities inherent in human language.
Conclusion and Future Work
While the LQPDN shows significant promise, future research could address its scalability and computational overhead compared to real-valued models. Furthermore, integrating this complex-valued approach into Transformer architectures like BERT or GPT could provide a powerful next step for sentiment-aware language modeling. Finally, reducing the dependency on pre-defined lexicons would improve the model's robustness to slang and domain-specific language.
Summary updated