In this paper we describe our approach to EVALITA 2016 POS tagging for Italian Social Media Texts (PoSTWITA). We developed a two-branch bidirectional Long Short Term Memory recurrent neural network, where the first bi-LSTM uses a typical vector representation for the input words, while the second one uses a newly introduced word-vector representation able to encode information about the characters in the words avoiding the increasing of computational costs due to the hierarchical LSTM introduced by the character-based LSTM architectures. The vector representations calculated by the two LSTM are then merged by the sum operation. Even if participants were allowed to use other annotated resources in their systems, we used only the distributed data set to train our system. When evaluated on the official test set, our system outperformed all the other systems achieving the highest accuracy score in EVALITA 2016 PoSTWITA, with a tagging accuracy of 93.19%. Further experiments carried out after the official evaluation period allowed us to develop a system able to achieve a higher accuracy. These experiments showed the central role played by the handcrafted features even when machine learning algorithms based on neural networks are used.
Building the state-of-the-art in POS tagging of Italian Tweets
Cimino A;Dell'orletta F
2016
Abstract
In this paper we describe our approach to EVALITA 2016 POS tagging for Italian Social Media Texts (PoSTWITA). We developed a two-branch bidirectional Long Short Term Memory recurrent neural network, where the first bi-LSTM uses a typical vector representation for the input words, while the second one uses a newly introduced word-vector representation able to encode information about the characters in the words avoiding the increasing of computational costs due to the hierarchical LSTM introduced by the character-based LSTM architectures. The vector representations calculated by the two LSTM are then merged by the sum operation. Even if participants were allowed to use other annotated resources in their systems, we used only the distributed data set to train our system. When evaluated on the official test set, our system outperformed all the other systems achieving the highest accuracy score in EVALITA 2016 PoSTWITA, with a tagging accuracy of 93.19%. Further experiments carried out after the official evaluation period allowed us to develop a system able to achieve a higher accuracy. These experiments showed the central role played by the handcrafted features even when machine learning algorithms based on neural networks are used.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.