Multi-task Learning in Deep Neural Networks at EVALITA 2018

English. In this paper we describe the system used for the participation to the ABSITA, GxG, HaSpeeDe and IronITA shared tasks of the EVALITA 2018 conference. We developed a classifier that can be configured to use Bidirectional Long Short Term Memories and linear Support Vector Machines as learning algorithms. When using Bi-LSTMs we tested a multitask learning approach which learns the optimized parameters of the network exploiting simultaneously all the annotated dataset labels and a multiclassifier voting approach based on a k-fold technique. In addition, we developed generic and specific word embedding lexicons to further improve classification performances. When evaluated on the official test sets, our system ranked 1st in almost all subtasks for each shared task, showing the effectiveness of our approach. Italiano. In questo articolo descriviamo il sistema utilizzato per la partecipazione agli shared task ABSITA, GxG, HaSpeeDe ed IronITA della conferenza EVALITA 2018. Abbiamo sviluppato un sistema che utilizza come algoritmi di apprendimento sia reti di tipo Long Short Term Memory Bidirezionali (Bi-LSTM) che Support Vector Machines. Nell’utilizzo delle Bi-LSTM abbiamo testato un approccio di tipo multi task learning nel quale i parametri della rete vengono ottimizzati utilizzando contemporaneamente le annotazioni presenti nel dataset ed una strategia di classificazione a voti di tipo k-fold. Abbiamo creato word embeddings generici e specifici per ogni singolo task per migliorare ulteriormente le performance di classificazione. Il nostro sistema quando valutato sui test set ufficiali ha ottenuto il primo posto in quasi tutti i sotto task di ogni shared task affrontato, dimostrando la validità del

[1]  Chih-Jen Lin,et al.  LIBLINEAR: A Library for Large Linear Classification , 2008, J. Mach. Learn. Res..

[2]  Ting Liu,et al.  Document Modeling with Gated Recurrent Neural Network for Sentiment Classification , 2015, EMNLP.

[3]  Felice Dell'Orletta,et al.  Tandem LSTM-SVM Approach for Sentiment Analysis , 2016, CLiC-it/EVALITA.

[4]  Huizhi Liang,et al.  UNIMELB at SemEval-2016 Tasks 4A and 4B: An Ensemble of Neural Networks and a Word2Vec Based Model for Sentiment Classification , 2016, *SEMEVAL.

[5]  Jeffrey Dean,et al.  Efficient Estimation of Word Representations in Vector Space , 2013, ICLR.

[6]  Janyce Wiebe,et al.  Recognizing Contextual Polarity in Phrase-Level Sentiment Analysis , 2005, HLT.

[7]  Danilo Croce,et al.  Overview of the EVALITA 2018 Aspect-based Sentiment Analysis Task (ABSITA) , 2018, EVALITA@CLiC-it.

[8]  Felice Dell'Orletta,et al.  Multi-Task Learning in Deep Neural Network for Sentiment Polarity and Irony classification , 2018, NL4AI@AI*IA.

[9]  Zoubin Ghahramani,et al.  A Theoretically Grounded Application of Dropout in Recurrent Neural Networks , 2015, NIPS.

[10]  Malvina Nissim,et al.  Overview of the Evalita 2016 SENTIment POLarity Classification Task , 2014, CLiC-it/EVALITA.

[11]  Felice Dell'Orletta,et al.  Overview of the EVALITA 2018 Hate Speech Detection Task , 2018, EVALITA@CLiC-it.

[12]  Jürgen Schmidhuber,et al.  Long Short-Term Memory , 1997, Neural Computation.

[13]  Kuldip K. Paliwal,et al.  Bidirectional recurrent neural networks , 1997, IEEE Trans. Signal Process..

[14]  Veselin Stoyanov,et al.  Evaluation Measures for the SemEval-2016 Task 4 “Sentiment Analysis in Twitter” (Draft: Version 1.13) , 2016 .

[15]  Felice Dell'Orletta,et al.  Building the state-of-the-art in POS tagging of Italian Tweets , 2016, CLiC-it/EVALITA.

[16]  Anders Søgaard,et al.  Deep multi-task learning with low level tasks supervised at lower layers , 2016, ACL.

[17]  Paolo Rosso,et al.  Overview of the EVALITA 2018 Task on Irony Detection in Italian Tweets (IronITA) , 2018, EVALITA@CLiC-it.

[18]  Malvina Nissim,et al.  Overview of the EVALITA 2018 Cross-Genre Gender Prediction (GxG) Task , 2018, EVALITA@CLiC-it.