Deep Semantic Learning for Conversational Agents

I did my Master’s Thesis under the supervision of Giuseppe Rizzo (from Istituto Superiore Mario Boella) and Maurizio Morisio (professor in DAUIN). The thesis, openly available, is centred on the topic of conversational agents and the task of Natural Language Understanding.

The main contribution of this work is a conversational model, that learns how to perform on a specific and narrow task, by taking into account the sentences from the users. The outputs of this work are:

• A seminar done with HKNPolito to introduce Deep Learning for Natural Language introductory event to Machine Learning. You can find the slides(italian) used for the presentation.
• Two papers accepted to workshops at The Web Conference 2018:
  • Multi-turn QA: A RNN Contextual Approach to Intent Classification for Goal-oriented Systems: focused on the modification of the model to create context-aware responses
  • The Rise of Emotion-aware Conversational Agents: Threats in Digital Emotions: describing the usage of emotions in conversational agents
• The release of the source code for the Natural Language Understanding model

The defense of the thesis was held on 12 April 2018 (slides, video) leading to my graduation with full marks.

I’ve made available a set of repositories that can be used to reproduce experiments and know all the implementation details:

• Bot Core and NLU: this component has the implementation of the neural network, the dataset, together with the implementation of the logic of the bot prototype
• Bot Server: this component is the public endpoint and acts as a proxy between the different messaging platforms and the Bot Core (communication through a web socket)
• Thesis source code: this repository can be useful for people that want to write a thesis with LaTeX and want to have an example
• Italian GloVe embeddings: the release of the italian word embeddings that have been trained for the italian version of the bot