Text analysis, via natural language processing (NLP) methods, aims at high-level information extraction. Applications vary from document classification to automatic answering machine. 

Our works aims at the development of “intelligent” human-machine interfaces using probabilistic models applied to text. We are currently collaborating with the company ELTEIDE/MAZER for the development of “Laila, a chatbot with a human touch”.

 

Word2Vec Networks

A Word2Vec network produces a mapping of large word dictionary of words into points in a multi-dimensional space. Word2Vect are used in a large number of applications, that range from document classification to question answering. The mapping is learned in an unsupervised way and leads to representations that reflect also a “closeness” in meaning among the words. Algebraic manipulations can answer analogy questions such as

King + Man – Queen = Woman

mobile int

 

We have developed one of the most efficient Word2Vec systems for the Italian language and analyzed its behavior when the hyperparameters change.

  • G. Di Gennaro, A. Buonanno, F.A.N. Palmieri, "Considerations About Learning Word2Vec," in The Journal of Supercomputing, Springer, pp. 1-16, 6 Apr. 2021. DOI: 10.1007/s11227-021-03743-2. 
  • G. Di Gennaro, A. Buonanno, A. Di Girolamo, A. Ospedale and F.A.N Palmieri, "An Analysis of Word2Vec for the Italian Language," in Progress in Artificial Intelligence and Neural Systems (Smart Innovation, Systems and Technologies 184), A. Esposito, M. Faundez-Zanuy, F.C. Morabito and E. Pasero, Eds., Springer, 2020, pp. 137–146. DOI: 10.1007/978-981-15-5093-5_13 - Paper presented also at the 29th Italian Workshop on Neural Networks (WIRN), June 12-14, 2019, Vietri sul Mare, SA, Italy. [Preliminary Version on arXiv:2001.09332] [Web page]

 

Intent Classification 

In a text dialog with a machine, questions have to be interpreted and classified. This is referred to as “intent classification”. We have proposed a neural network architecture based on LSTM (Log Short Term Memory).

 

  • G. Di Gennaro, A. Buonanno, A. Di Girolamo, A. Ospedale and F.A.N Palmieri, "Intent Classification in Question-Answering Using LSTM Architectures," in Progress in Artificial Intelligence and Neural Systems (Smart Innovation, Systems and Technologies 184), A. Esposito, M. Faundez-Zanuy, F.C. Morabito and E. Pasero, Eds., Springer, 2020, pp. 115–124. DOI: 10.1007/978-981-15-5093-5_11- Paper presented also at the 29th Italian Workshop on Neural Networks (WIRN), June 12-14, 2019, Vietri sul Mare, SA, Italy. [Preliminary Version on arXiv:2001.09330]