Made available in DSpace on 2014-06-12T15:54:30Z (GMT). No. of bitstreams: 2 arquivo7155_1.pdf: 666963 bytes, checksum: 56eeb6eb903215d0a8b285686ffe780a (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2005 / O trabalho de pesquisa aqui apresentado teve como objetivo principal melhorar o desempenho de chatterbots em diálogos livres com usuários. Chatterbots são sistemas computacionais que se propõem a conversar em linguagem natural como se fossem humanos. O primeiro desses sistemas foi ELIZA, desenvolvido em 1965 por Weizenbaum. Desde então, inúmeros sistemas foram produzidos com esse mesmo objetivo. Porém, uma série de problemas ainda continuam em aberto, dentre os quais, o tratamento de intenção, questão central na interpretação de diálogos naturais. Nesse sentido, desenvolvemos um mecanismo para tratamento de intenção para ser incorporado a chatterbots baseados em AIML. Adotamos como base conceitual para o trabalho a Teoria da Análise da Conversação (TAC), por considerar a intenção em pares adjacentes, e não apenas na sentença do falante, como a Teoria dos Atos de Fala. Com base na TAC e em experimentos realizados, selecionamos um conjunto de intenções, que foram utilizadas na criação de regras em AIML que utilizam informações de intencionalidade para interpretar e gerar sentenças em diálogos naturais. A solução final foi testada em uma série de experimentos, e demonstrou ser capaz de corrigir alguns problemas presentes em diálogos com chatterbots. Por exemplo, o sistema baseado em AIML padrão tratou 40% das sentenças dos usuários como sendo desconhecidas, enquanto o nosso sistema classificou apenas 3,5% das sentenças como totalmente desconhecidas. Além disso, o sistema foi capaz de manter a estrutura global dos diálogos, criticando turnos de abertura ou fechamento que foram ditos no desenvolvimento, ou turnos de desenvolvimento ditos na abertura ou fechamento. Por fim, implementamos três aplicações com chatterbots, o que demonstra que a solução adotada favorece o reuso de categorias em bases AIML, processo extremamente custoso do ponto de vista de engenharia de software com os sistemas atuais

Made available in DSpace on 2014-06-12T15:58:31Z (GMT). No. of bitstreams: 2 arquivo4642_1.pdf: 3290764 bytes, checksum: 0f87315cf72b415c90e599ead8d47c07 (MD5) license.txt: 1748 bytes, checksum: 8a4605be74aa9ea9d79846c1fba20a33 (MD5) Previous issue date: 2003 / Chatterbots são sistemas que se comunicam com usuários em linguagem natural, e podem ser utilizados como interfaces nas mais variadas aplicações (e.g., comércio eletrônico, ensino a distância, entre outras). Estudos recentes têm demonstrado a importância de explorar o comportamento social do usuário perante o computador, destacando-se o papel da personalidade. Além disso, a personalidade também é vista como um componente importante no processo de tomada de decisão. Dentre as ferramentas e linguagens para a criação de chatterbots, AIML (Artificial Inteligence Markup Language) é, atualmente, uma das mais bem sucedidas abordagens. Entretanto, AIML não oferece suporte para a modelagem e a implementação de personalidade em chatterbots. Esta dissertação apresenta a arquitetura Persona-AIML, uma extensão do AIML original, que possibilita o desenvolvimento de modelos de personalidade para chatterbots. Na arquitetura proposta, a personalidade é composta pelos seguintes elementos: traços, atitudes, humor, emoções e estados físicos. Neste trabalho, chatterbots são modelados como agentes racionais e utilizam uma base de regras para descrever sua personalidade (tendo como objetivo a modularidade, extensibilidade e reusabilidade). O protótipo desenvolvido utilizou um modelo de personalidade baseado no Modelo dos Cinco Grandes Fatores . Entretanto, a arquitetura proposta permite a utilização de outros modelos de personalidade. Essa arquitetura foi implementada em Java, e os testes revelaram resultados satisfatórios do protótipo

Uppsatsen behandlar artificiell intelligens ur ett service- och tjänsteperspektiv. Den huvudsakliga inriktningen är på området virtuell assistans, med fokus på virtuella agenter och chatterbots, vilket är en slags kommunikativ mjukvara. Studien syftar till att undersöka vilket mervärde virtuell assistans kan medföra, detta både ur ett företags- och kundperspektiv. Studien syftar även till att undersöka hur en lyckad implementering av tekniken fungerar. Utöver detta behandlas ett moraletiskt perspektiv, där utvecklarnas ansvar av artificiell intelligens tas upp. En induktiv forskningsansats har använts, medan det insamlade resultatet kommer från fyra intervjuer från olika företag som alla har arbetar med virtuell assistans. Som ett komplement till materialet från intervjuerna tillkommer lämplig teori på området artificiell intelligens, virtuell assistans och chatterbots. Studiens resultat har visat på att det stora mervärdet för intelligens programvara ligger i dess tillgänglighet och effektivitet, som kund, eller användare, tillhandahålls ett svar i realtid. Att ha en virtuell agent implementerad på en webbsida har visat på trygghet för kunder vid internetköp. Trots att uppsatsen behandlar virtuell assistans, har det varit återkommande att den mänskliga kontakten fortfarande är viktig. Det innebär att om kundens fråga är av svårare karaktär, är den en god idé om en människa tar över. Ur ett moraletiskt perspektiv måste utvecklare av artificiell programvara ha en medvetenhet kring vilken information den intelligenta tekniken utsätts för och hur röststyrda agenter kan bidra till att hjälpa en nödställd människa. / The purpose of this paper is to study artificial intelligence from a service point of view. The main focus is in the area is virtual assistance, with a focus on virtual agents and chatterbots, which is a kind of communication software. The study aims to examine the added value that virtual assistance can bring, both from a business and customer perspective. The study also aims to examine how a successful implementation of the technology works. In addition to the above, the paper deals with a moral and ethical perspective, and brings forward the developers responsibility. An inductive research approach has been used as a method, while the collected results come from four interviews from different companies who all have used virtual assistance in the service field. As a complement to the interviews, relevant theory in the field of artificial intelligence and virtual assistance has been used. The study results have shown that the added value of intelligent software lies in its accessibility and efficiency. To have a virtual agent implemented on a web page has demonstrated safety for the customer when shopping online. Although the paper deals with virtual assistance, the results have also shown that it is important that a human handle the customer, if the issue is of a complex nature. From a moral and ethical perspective, developers of artificial intelligence has to be aware of the information that the virtual assistance is exposed to. They also need to be aware how the voice control software can help a human in need.

This thesis explores the idea of how people 'do gender' in their online use of avatars, specifically avatar choice. A secondary question of whether or not a chatterbot can be used as a potential interviewer will also be examined as a tool acquiring large amounts of interview data. Gender is one of the ways in which we structure our society, and is completely omnipresent. We cannot opt out of participating in our gender, as we are constantly performing and reaffirming it. Because of this, gender performance and choice spills over into all domains. This includes entertainment such as massively multiplayer online games, both in how the designers make the game, and what the players bring to the game. Deconstructing how and why people engage in these gendered practices and choices becomes an interesting avenue of research, because it allows researchers to partially separate the mental aspects of gender from physical attributes, as the players' physical bodies are not actually in the game. Through the lens of the popular massively multiplayer online game, World of Warcraft, this thesis will utilize a qualitative user research study to understand how gender affects avatar choices. Prior research identified areas where players brought real world gender norms into the games they played. This research study will extend previous research by having players identify why they made the choices they made for their avatars, and how they feel about those choices. The methodology for this study will also involve using a chatterbot as a way of gathering interviews. In normal person-to-person interview studies, recruiting and organizing meetings for these interviews can often be a difficult task. This thesis brings in the idea of using a chatterbot as a mechanism to gather more interviews in a shorter time span to alleviate the problem of getting these one-on-one interviews in some types of studies.

This work deals with current chatterbot systems. It describes problems and possibilities of improvement with emphasis on natural language processing and emotion modeling during conversation. There is an implementation, based on the described knowledge, introduced in the second part of the thesis, also with experimental success rate evaluation.

Chatbots are becoming increasingly popular in various industries, and thereare many options available for businesses and organisations. Several studieshave investigated open-source chatbots and identified their core strengths,implementation, and integration capabilities however few have investigatedopen-source chatbot frameworks and libraries in a specific use case such asmedicine. The project's objective was to evaluate a selection of chatbots ormore specifically two frameworks: Botkit and Rasa, and two libraries:ChatterBot, and Natural which was utilised together with Botkit and alanguage model which is DialoGPT. The evaluation focuses specifically onaccuracy, consistency, and response time. Frequently asked questions fromthe World Health Organization and COVID-19 related Dialogue Datasetfrom GitHub were utilised to test the chatbots' abilities in handling differentqueries and accuracy was measured through metrics like Jaccard similarity,bilingual evaluation understudy (BLEU), and recall oriented gistingevaluation (ROUGE) scores, consistency through Jaccard similarity betweenthe generated responses and response time was taken to be the average timefor a response in seconds. The analysis revealed unique strengths andlimitations for each chatbot model. Rasa displayed robust performance inaccuracy, consistency, and customisation capabilities if the chatbot works ina particular topic with acceptable response times. DialoGPT demonstratedstrong conversational abilities and contextually relevant responses withtrade-offs in consistency. ChatterBot showed consistency, though sometimesstruggled with advanced queries, and Botkit with Natural stood out for itsquick response times, albeit with limitations in accuracy and scalability.Despite implementation challenges, these open-source frameworks, libraries,and models offer promising solutions for organisations intending to harnessconversational agents' technology. The study suggests encouraging furtherexploration and refinement in this rapidly evolving field.

This paper researches how different amounts of data affect different word vector models for classification of dialog system user input. A hypothesis is tested that there is a data threshold for dense vector models to reach the state-of-the-art performance that have been shown with recent research, and that character-level n-gram word-vector classifiers are especially suited for Swedish classifiers–because of compounding and the character-level n-gram model ability to vectorize out-of-vocabulary words. Also, a second hypothesis is put forward that models trained with single statements are more suitable for chat user input classification than models trained with full conversations. The results are not able to support neither of our hypotheses but show that sparse vector models perform very well on the binary classification tasks used. Further, the results show that 799,544 words of data is insufficient for training dense vector models but that training the models with full conversations is sufficient for single statement classification as the single-statement- trained models do not show any improvement in classifying single statements. / Detta arbete undersöker hur olika datamängder påverkar olika slags ordvektormodeller för klassificering av indata till dialogsystem. Hypotesen att det finns ett tröskelvärde för träningsdatamängden där täta ordvektormodeller när den högsta moderna utvecklingsnivån samt att n-gram-ordvektor-klassificerare med bokstavs-noggrannhet lämpar sig särskilt väl för svenska klassificerare söks bevisas med stöd i att sammansättningar är särskilt produktiva i svenskan och att bokstavs-noggrannhet i modellerna gör att tidigare osedda ord kan klassificeras. Dessutom utvärderas hypotesen att klassificerare som tränas med enkla påståenden är bättre lämpade att klassificera indata i chattkonversationer än klassificerare som tränats med hela chattkonversationer. Resultaten stödjer ingendera hypotes utan visar istället att glesa vektormodeller presterar väldigt väl i de genomförda klassificeringstesterna. Utöver detta visar resultaten att datamängden 799 544 ord inte räcker till för att träna täta ordvektormodeller väl men att konversationer räcker gott och väl för att träna modeller för klassificering av frågor och påståenden i chattkonversationer, detta eftersom de modeller som tränats med användarindata, påstående för påstående, snarare än hela chattkonversationer, inte resulterar i bättre klassificerare för chattpåståenden.