International Conference on Connected Intelligence for Industrial Applications (CI2A)
Hybrid ML-NLP Framework with Quantum Embeddings for Semantic Analysis and Multilingual Emotion Understanding
Muskan, Kulbir Kaur, Ankita Avthankar, Gyana Ranjana Panigrahi, Harpal Singh
Abstract
Emotions and semantics across languages are important in digital communication to be used in social media analysis applications, customer feedback systems, and humancomputer interfaces. Conventional machine learning methods tend to have issues with the specifics of multilingual data, resulting in less than optimal results in emotion recognition and semantic understanding. To cope with these issues, the paper presents a hybrid ML-NLP model based on quantum embeddings. The framework takes advantage of transformerbased models to extract the initial features and uses quantum machine learning to achieve high classification accuracy. The dataset utilized in the present study is a mix of PEMO dataset of Punjabi text emotions based on the work of [1] pemo and GoEmotions dataset of English with an annotation of more than 58,000 samples in 27 emotion categories. We supplemented this with multilingual samples of SemEval-2018 Task 1, which left us with an evenly distributed corpus of 100,000 entries with support to the English, Punjabi and Hindi languages. This heterogeneous dataset is robust in the way of dealing with code-switched and diversified language forms. Driven by shortcomings of the current model, including lack of dealing with ambiguity in semantics of low-resource languages, inability of classical embeddings to learn quantumlike superposition in emotional states, our model is an integration between BERT embeddings and quantum variational circuits. The approach is pre-training on semantic tasks, fine-tuning on quantum classifiers, and cross-lingual benchmark. According to the experimental results, an overall accuracy of$\mathbf{9 2 \%}$, precision and recall of over$\mathbf{9 0 \%}$across emotions have been proven. Quantum embeddings showed a 15 percent increase in semantic understanding compared to the classical approaches especially in the multilingual set-ups. This model exceeds such baselines as SVM and standard BERT by a wide margin, making it a promising solution in the realm of such applications in emotion-aware systems.