In the last two years, the term chatbot has become a buzzword and it has gained so much popularity to be often described as the `next big thing'. Many companies and people invested large sums of money in research and development of this technology and a considerable number of start-ups specialized in chatbots appeared. Nevertheless, many attempts to build chatbots gave rise to poor results because of too high expectations. In detail, many companies built chatbots as a customer engagement strategy without understanding their limits and benefits, failing to deliver what they promise. In light of this context, the aim of this work is investigating the problem of building an AI-powered transactional-based chatbot, i.e. a close-domain chatbot able to help the users to get services from a provider. We investigate the possibility of building such chatbot by favoring machine learning techniques over rule-based approaches. To do so, we analyze some existing commercial solutions, find their limitations and propose a meta-model able to solve - or at least lessen - these issues. Furthermore, we implement our meta-model by creating a chatbot in collaboration with an operator of the mass distribution industry. Usability testing was used to evaluate and improve the prototype. The results show that people are satisfied with the ability of the chatbot to complete the transactions, however, they are disappointed with its conversational capabilities, which are its main limitations. Further studies need to be conducted in this area. To sum up, the main novelty of our approach with respect to the scientific literature is twofold: first, we formalized the mathematical abstraction of conversational agent and its properties. Secondly, we formalized and described a chatbot architecture based on the concepts of intent, context and entity, which are present at an early stage in commercial solutions and are unexplored by the scientific community.
Negli ultimi due anni, il termine chatbot è diventato di moda e ha guadagnato così tanta popolarità da essere spesso definito come 'la prossima grande novità'. Molte persone e aziende hanno investito grosse somme di denaro nella ricerca e sviluppo di questa tecnologia e si è formato un considerevole numero di start-up specializzate in chatbots. Nonostante ciò, innumerevoli tentativi di sviluppo di chatbots sono sfociati in risultati insufficienti a causa di aspettative troppo alte. Nello specifico, molte aziende hanno costruito chatbots come strategia di customer engagement senza davvero capire né i loro limiti né i loro vantaggi, non riuscendo dunque ad offrire un servizio all'altezza di quanto promesso. Alla luce di ciò, l'obiettivo di questa tesi è investigare il problema della creazione di un chatbot transazionale, cioè un chatbot di dominio capace di assistere gli utenti nella fruizione di servizi offerti da un'organizzazione. L'utilizzo di tecniche di apprendimento automatico saranno privilegiate rispetto ad approcci basati sul riconoscimento di patterns. Per fare ciò, il nostro lavoro consiste nell'analizzare alcune soluzioni commerciali esistenti, trovare le loro limitazioni e proporre un meta-modello capace di risolvere - o almeno limitare - questi problemi. In aggiunta, il meta-modello proposto viene anche implementato creando un chatbot in collaborazione con un operatore della grande distribuzione. La tecnica chiamata usability testing è stata usata per valutare e migliorare l'applicazione stessa. I risultati mostrano come gli utenti siano soddisfatti di come il chatbot completi le transazioni, ma rimangano delusi dalle sue capacità conversazionali, le quali rappresentano il principale limite di questa tecnologia. Ulteriori studi in quest'area sono richiesti. I principali aspettivi innovativi del nostro lavoro sono i seguenti: da un lato, abbiamo formalizzato un'astrazione matematica dell'agente conversazionale e le sue proprietà; dall'altro, abbiamo sistematizzato e descritto un'architettura di chatbot basata sui concetti di intento, contesto ed entità, che sono presenti in maniera embrionale negli approcci commerciali ma scientificamente inesplorati.
An intent-based and context-dependent approach to the design of chatbots
MIGLIORINO, LORENZO
2016/2017
Abstract
In the last two years, the term chatbot has become a buzzword and it has gained so much popularity to be often described as the `next big thing'. Many companies and people invested large sums of money in research and development of this technology and a considerable number of start-ups specialized in chatbots appeared. Nevertheless, many attempts to build chatbots gave rise to poor results because of too high expectations. In detail, many companies built chatbots as a customer engagement strategy without understanding their limits and benefits, failing to deliver what they promise. In light of this context, the aim of this work is investigating the problem of building an AI-powered transactional-based chatbot, i.e. a close-domain chatbot able to help the users to get services from a provider. We investigate the possibility of building such chatbot by favoring machine learning techniques over rule-based approaches. To do so, we analyze some existing commercial solutions, find their limitations and propose a meta-model able to solve - or at least lessen - these issues. Furthermore, we implement our meta-model by creating a chatbot in collaboration with an operator of the mass distribution industry. Usability testing was used to evaluate and improve the prototype. The results show that people are satisfied with the ability of the chatbot to complete the transactions, however, they are disappointed with its conversational capabilities, which are its main limitations. Further studies need to be conducted in this area. To sum up, the main novelty of our approach with respect to the scientific literature is twofold: first, we formalized the mathematical abstraction of conversational agent and its properties. Secondly, we formalized and described a chatbot architecture based on the concepts of intent, context and entity, which are present at an early stage in commercial solutions and are unexplored by the scientific community.| File | Dimensione | Formato | |
|---|---|---|---|
|
2017_10_Migliorino.pdf
non accessibile
Descrizione: Testo della tesi
Dimensione
6.14 MB
Formato
Adobe PDF
|
6.14 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/136034