Murauer, J. (2023):

Analyzing word predictions by quantum natural language processing

Quantum natural language processing is an emerging field, that combines principles of quantum computing with natural language processing with the aim to enhance the capa- bilities to process and analyze human language. In classical natural language processing, word prediction is an ingredient of the pretraining task for large language models. This forms the motivation behind the exploration of a word prediction task in QNLP. This thesis deals with the implementation of a word prediction task in quantum nat- ural language processing using a mathematical framework, particularly the DisCoCat framework. Firstly, we reformulate word prediction as a binary classification task and train quantum machine learning models on that task. Secondly, we implement a word prediction task as a multiclass classification and also train models on that. In the course of this implementation, we give a comprehensive explanation of how to design a mul- ticlass classification in the DisCoCat framework. Afterwards, we show how masking can be conceptualized inside a quantum computing environment. Then, a strategy is presented, which reduces the number of qubits and lowers the task complexity. This strategy has the potential to be extended beyond word prediction and to be applied to various other problems as well. The evaluation of the models which have been trained with the binary classification task shows promising results. The masking approach and the strategy for reducing the number of qubits were both a success in the evaluation. However, the evaluation of the models trained on the multiclass classification approach is not of the same standard as the binary approach. The reasons for that outcome are discussed in depth. For both tasks, we document the applied hyperparameters. Further research topics include the continued development of multiclass classification in the DisCoCat framework and re- search on new ansätze in quantum machine learning.

• PDF Version (4,513.6 KB)