THE 3 rd INTERNATIONAL SCIENTIFIC CONFERENCES OF STUDENTS AND YOUNG RESEARCHERS dedicated to the 99
th
anniversary of the National Leader of Azerbaijan Heydar Aliyev
128
Taking the complex morphological structure into account, the character
tokenization of the sentences is preferred over word tokenization to consider
all possible word occurrences in the sentences. Then, padding is applied to
define maximum length for the input sentences.
Recurrent Neural Networks (RNN) is implemented with encoder and
decoder layers which provides an advantage of holding result from previous
step in its hidden state and using it predicting the next character in the word.
Encoder transforms input sentence into feature vectors that were used by
decoder as an initial hidden state with the combination of current word or
character and its own hidden state. Both encoder and decoder are built by
defining embedding and LSTM layers. Encoder-decoder-based model can
result in performance bottleneck when a long sentence is inputted which
makes it difficult for RNN to transmit all the required information from the
previous step. To prevent this problem, Bahdanau attention mechanism is
implemented between encoder and decoder to keep and pass all information
over full length of input sentence efficiently. Attention mechanism has 3 parts,
namely Feed forward neural network, softmax calculation and context vector
that contain information about attention should be allotted for each element
in the sequence.
Finally, the model is evaluated on test dataset that contains total of 1000
misspelled words and the accuracy of 94% is achieved as a result. Moreover,
it can be further improved by expanding the training dataset and continue the
study of typing behaviour of Azerbaijani people. The platform can be created
where the help can be acquired from the community by developing the game
for labelling the correct and wrong sentence pairs as true or false which can
provide us with the advantage of utilizing feedback mechanism in the
supervised learning to minimize the loss. As a future work, the software
extension as Grammarly can be developed to ease the usage of the model
for community.
References [1] Haldar, Rishin & Mukhopadhyay, Debajyot, “Levenshtein Distance Technique in Dictionary
Lookup Methods: An Improved” 2011.
[2] Minjoon Seol, Aniruddha Kembhavi, Ali Farhadi, Hananneh Hajishirzi, bi-directional
attention flow for machine comprehension, conference paper at ICLR, 2017
[3] Samir Mammadov, Neural Spelling Correction for Azerbaijani Language, IEEE Xplore,
2019