This code implements a quantum model for binary sentence classification. It is built on top of lambeq, a Python software library developed for the design and training of Quantum Natural Language Processing (QNLP) models.
A JAX-based backend is implemented to provide just-in-time (JIT) compilation and automatic differentiation capabilities. qujax, a JAX-compatible library for quantum circuit construction and evaluation, is used as a part of this backend. A class interfacing with the Optax optimization library was developed, making most modern optimization algorithms available to use for training.
To use the code included in this repository, you must first install a custom version of lambeq which has minor compatibility changes:
pip install git+https://github.com/gamatos/lambeq.git@linear_model_changes
Other dependencies include qujax, optax, pandas, matplotlib, jax and jaxlib, all of which can also be obtained via pip.
You can see a full usage example in training_example.ipynb.
A tokenised version of the poem sentiment dataset is provided in the poem_sentiment folder for example purposes only. For simplicity, all but the positive (here labeled with a 1) and negative (here labeled with a 0) examples have been removed. Note that this processed version of the dataset is distributed under the same CC-BY 4.0 license as the original, and this license differs from the rest of this repository.
In the language of string diagrams, used by lambeq, this model treats the sentences as if they had been first converted to diagrams containing only information about the order of the words using lambeq's LinearReader class. Since the only relevant information is the word order, in practice this model works directly with tokenised sentences, and the diagram construction machinery is not needed. This, in conjunction with JAX, is used to dramatically speed up the evaluation of the model.