Morphological segmentation using ConvLSTM networks

OData support
Supervisor:
Ács Judit
Department of Automation and Applied Informatics

This thesis consists of experiments on morphological segmentation using convLSTM

and LSTM networks. We define multiple new neural network architectures to reach

state-of-the-art performance on the morphologically rich Hungarian language. This

work provides a comparison of static and dynamic neural networks by describing the

current deep learning frameworks and benchmarking their speed and convergence

capabilities. We give a brief description of the main machine learning architectures,

such as convolutional networks and recurrent neural networks and how we can con-

nect them together to reach better performance. Our best performing models utilize

most of the recent improvements from the deep learning field. The thesis introduces

new architectures which can be further improved due to their completely data-driven

behaviour. We also perform a wide hyperparameter search on all kinds of models

to drive further current neural network approaches for character-level segmentation.

All the introduced models are completely language-independent and they can serve

not only morphological but other sentence-tagging tasks due to their very natural

and highly recurrent architecture. We compare the best performing models trained

with sentence and word inputs. The experimental results are automatically evalu-

ated during the training in order to help the development of the networks and to

prevent large models from overfitting the train dataset. An inference framework

is provided for testing the machine learning models on unseen words, in order to

evaluate the results of the segmenting algorithms. All the notebooks and code can

be found in the open-source repository along with the model descriptions and data.

Downloads

Please sign in to download the files of this thesis.