LSTM Recurrent Neural Networks for Short Text and Sentiment Classification
Nowak, Jakub, Ahmet Taspinar, and Rafał Scherer. “LSTM recurrent neural networks for short text and sentiment classification.” International Conference on Artificial Intelligence and Soft Computing. Springer, Cham, 2017.
RNN models are superior in NLP than other models.
In authors’ word, ‘Human communication is a very complex and complicated process. …. We use dictionaries, i.e. collections of words describing our surroundings, feelings, and thoughts. Very important is their order and context.” Statistical and Feed Forward Neural Nets don’t take the order of the words in sentence into account. Thus a lot of the information is wasted.
LSTM. When using this structure, we activate the same cell every time, modifying the state of its internal structure. The most important elements of LSTM cells are – cell state – the state of the cell passed in sequence after to the next steps, – forget gate – the gate that decides what information should be omitted, – input gate – a gate that decides what should be forwarded to the next activation.
Experiments run on Microsoft CNTK shows in small sentences GRU performed slightly better. And all other case, LSTM or BLSTM outperformed all other methods.
RNN is suitable for sequential problems.
This is a non-technical proof of the RNN’s superiority in NLP. What are future directions for this research (author’s and yours, perhaps driven by shortcomings or other critiques)? Memory requirements in the learning process is a problem.
LSTM model in detail.
NOTE: These are my personal point of view, uploading them here to keep up the writing habit for every paper that I read.