Publications

Prediction of words in Turkish sentences by LSTM-based language modeling

For my master thesis, I focus on language modeling. I have built alternative Neural Networks to model the Turkish language. I designed both LSTM and GRU networks as well as stacked LSTMs and GRUs using Tensorflow. I have done series of experiments such as generating sentences, word prediction, sentence ending prediction. However, the main motivation of my thesis was to explore the correlation between my language model predictions and human predictions on word level. I compared my model’s word predictions with human predictions gathered from an independent cloze test experiment.

Team ReadMe at CMCL 2021 Shared Task: Predicting Human Reading Patterns by Traditional Oculomotor Control Models and Machine Learning

We have participated in the CMCL 2021 Shared Task. The goal was to predict human eye movements while reading. We have built two different models. One of them is the linear regression model, and the other is based on the LSTM network. As input, we use word characteristics that have the most effect on human reading patterns. For example; word length, word frequency, word predictability. Word predictability is produced from another LSTM network based on my master’s thesis.