Chatbot Mechanisms

The chatbot contests introduced in the previous section determine one or more relatively more human-like chatbots based on subjective votes or a quantitative scale evaluation without specifying the reasons behind the scores. Kirakowski et al. (2009) in a content analysis study inquired their fourteen participants the reasons why the ELIZA-style chatbot they talked to was natural or non-natural. Six unconvincing chatbot characteristics were documented from their stimulus recall interviews: (1) failure to maintain a n initiated theme of conversation; (2) formal or unusual treatment of language; (3) failure to respond to specific questions; (4) failure to respond to general questions or cues requiring elaborations; (5) time delay or promptness in responding; (6) inappropriate phrases or lack of reference to previous dialogues.

On the other hand, seven convincing aspects were also identified from the transcripts: (1) human-seeming greetings; (2) success in maintaining a theme; (3) admitting the breakdown of conversation and redirect the talk to a more fruitful direction; (4) reacting appropriately to a cue; (5) offering a cue for further discussion; (6) using conversational or colloquial English;

(7) manifestation of a personality.

To rectify the problems with chatbots’ unnaturalness in human-computer interactions, different chatbot mechanisms have been developed to improve the performance of chatbots.

Bruce Wilcox (2011a) categorized chatbots into two general kinds, including the data- mining chatbots by Jabberwacky (2011) such as Cleverbot and rule-based chatbots like A.L.I.C.E.

and Wilcox further identified four kinds of engines included in the rule-based chatbot category: AIML, ChatScript, Façade, and the Personality Forge. Wilcox (2011a) also discussed the underlying mechanisms of ALICE bots, ChatScript bots and a distinct AI technology used by a one-act interactive drama called Façade, but that technology seems to be limited to the virtual characters in the game, Façade, not the other chatbots designed to maintain open-ended conversations. Jabberwacky.com and Icogno.com are affiliated AI businesses that developed a chatbot mechanism collecting what humans say to it as its own

15

output responses, thus belonging to the category of the data- mining chatbots. This section reviews the above mechanisms in more detail, including AIML, Jabberwacky, Façade, the Personality Forge, and ChatScript.

2.1.5.1 AIML

ALICE bots, including the three-time Loebner Prize winner, A.L.I.C.E., the Artificial Linguistic Internet Computer Entity, are constructed using the Artificial Intelligence Mark-up Language or AIML, a chatbot architecture language that Dr. Richard S. Wallace and his colleagues developed during 1995 to 2000 (Abu Shawar & Atwell, 2007a; A.L.I.C.E. AI Foundation, n.d.; Atwell, 2005; Wilcox, 2011a). Pandorabots.com is a chatbot platform that hosts thousands of chatbots constructed using AIML (Abu Shawar & Atwell, 2007a;

A.L.I.C.E. AI Fundation, 2007; Atwell, 2005). Abu Shawar and Atwell (2007a) and Wilcox (2008, 2011a, 2011b) pointed out that AIML language is basically divided into two units called topics and categories. Topics are the top- level label representing a set of categories.

Categories are the knowledge unit in AIML and are scripted in three kinds, including atomic categories, default categories, and recursive categories.

Atomic categories are those matching rules that involve an exact match of the input for an exact output to be triggered. This kind of categories does not have a wild card symbol and thus are less flexible. The second kind is the default categories, the pattern rules with one or more wildcard symbols in the form of a star symbol “*” or an underline “_” to accommodate partial matching of the input. The AIML system first screens through the atomic categories and when that fails, default categories are screened to find the one that best matches the user input.

The last one is the recursive categories, those rules that intend to simplify, divide, or transform the input for better matching. Before all the above matching occurs, AIML uses a normalization process to remove all punctuations, divide sentences, and change all inp ut into

16

uppercases. Therefore, AIML disregards mechanics and is case- invariant. The design is straightforward with exact inputs matched with one or more pre- fabricated or random responses when matching is not made.

According to Wilcox (2011a, 2011b), there are several shortcomings in AIML. To begin with, although wildcard symbols are available in AIML, at least one word should be put in the wildcard position for an input to be matched and to activate the paired output. Therefore, for sentences without additional modifiers or articles in a wildcard symbol position, an input will not be matched with a pre- fabricated output, since no exact matching is achieved. In that situation, a random response will be given (Jia, 2004a). In order to match all possible variations of one statement, for example, “I love you,” at least four rules have to be made to have most conditions matched, including “I love you”, “ * I love you”, “I love you *”, and “*

I love you *.”Such a restraint makes AIML inflexible and more time-consuming when scripting a chatbot to deal with one question in a specific language (Abu Shawar & Atwell, 2007a; Wilcox, 2011a, 2011b) since several predetermined patterns need to be manually coded for one of them to be properly matched.

Secondly, AIML can transform complex sentences with multiple clauses into simple sentences with similar meanings to be properly matched. For instance, the input “Tell me what * is” can be transformed into “What is *” before the system processes the output matching. Nevertheless, with more and more rules of this kind, the botmaster cannot easily spot the outcome of any new rules he creates, since some previously- made self- modification rules might get in the way and some unexpected output might be produced. This lack of transparency also makes debugging the program difficult.

As an open chatbot language, AIML has been the most popular chatbot language for some time since A.L.I.C.E. has been ranked the best program in the 2000, 2001, and 2004 Loebner Prize competition and has won other minor awards in numerous similar contests.

Ongoing attempts are also made to refine the ALICE bots and the underlying system by

17

training ALICE bots, for example, using existing conversational corpora (Abu Shawar &

Atwell, 2005; Atwell, 2005) or other domain knowledge data (Heller et al., 2005; Schumaker et al., 2006).

2.1.5.2 Jabberwacky

Jabberwacky chatbots store everything everyone has ever said and then select the most appropriate responses using contextual pattern matching techniques (Jabberwacky, n.d.; Kerly et al., 2008). They are created to entertain, to gather conversational data and to pass the Turing Test (Icogno, n.d.).

In 2005 and 2006, George and Joan respectively won the bronze medal award in the Loebner Prize competition. Both are Jabberwacky bots that can capitalize on the inputs generated by the human interlocutors, making it more complicated as more interactions take place (Child, 2006). Cleverbot has been introduced as the latest work by Jabberwacky.com and has won the grand prize of the 2010 BCS SGAI Machine Intelligence Contest and came in second and third in the 2009 and 2010 Loebner Prize competition. Aron (2011) reported a Turing Test held in Guwahati, India, where 1,334 attendants voted for the human- likeness of Cleverbot based on the transcripts obtained from thirty volunteers, Cleverbot was voted to be 59.3% human while the counterpart human entity was only 63.3% human. The Cleverbot in the contest was a more powerful one as indicated by Cleverbot’s creator, Rollo Carpenter that the online version of Cleverbot only searches its database for three times before deciding on a best response while the complete version on this occasion did 43 searches before responding.

Since Jabberwacky bots gather data from all the random chats with humans (Jabberwacky, n.d.), a deficiency with the system is its lack of a defined personality (Wilcox, 2011a). Jabberwacky bots like George, Joan, and Cleverbot all respond using data- mining techniques and might not be consistent throughout the chat because their responses are based on different human beings they have previously conversed with.

18

2.1.5.3 Personality Forge

The Personality Forge is an online chatbot platform that has its own chatbot mechanism (Personality Forge, n. d.). Wilcox (2011a, 2011b) considered the technology employed by the Personality Forge to be less complicated to construct than AIML. Instead of matching the exact words, the Personality Forge can match the examples of the same concept with a built- in dictionary, making it more flexible than AIML. The introduction page of the Personality Forge describes its technology using the following statement:

The Personality Forge's AI Engine integrates memories, emotions, knowledge of hundreds of thousands of words, sentence structures, unmatched pattern- matching capabilities, and a scripting language called AIScript (Personality Forge, n.d.).

The most famous example of the Personality Forge is Bildgesmythe, created by Patti Roberts.

Bildgesmythe has won first place in the Chatterbox Challenge for three times in 2007, 2008, and 2011. On June 30th, 2012 at State of Mind: A Consciousness Expo in Brighton, England, 101 visitors chatted with Bildgesmythe and a human; afterward, the visitors tried to identify the machine and the human. Of the 101 visitors, thirty-four visitors selected Bildgesmythe to be the human entity (Adams, 2012), exceeding the 30% threshold established by Turing (1950). However, the platform that host this kind of chatbots do not offer complete transcripts of conversations.

2.1.5.4 Façade

Façade is a one-act interactive drama with AI-based, real-time rendered 3D virtual characters, Grace and Trip (Procedural Arts, n. d.). Instead of matching the exact input to a reaction output, Façade is designed to match one of the fifty discourse acts such as agreeing and apologizing (Wilcox, 2011a). Façade’s technology however, seems to be limited to the

19

virtual characters in the game for the researcher in the current study did not find any other chatbot entities or literature claiming to adopt the same technology by Façade. Wilcox (2011a) concluded that overall Façade is more efficient to construct than AIML bots, but matching discourse acts does not necessarily entail accurate interpretation of the input. Based on the review in Wilcox (2011a, 2011b), Bruce Wilcox designed a new chatbot mechanism, ChatScript to developing his own chatbots (Anonymous, 2011) by drawing on the advantages of existing chatbot mechanisms.

2.1.5.5 ChatScript

In 2010, 2011, and 2012, Suzette, Rosette, and Angela, created by Bruce Wilcox in Telltale (Anonymous, 2011), won the Loebner Prize competition for three consecutive years.

According to Wilcox (2011a, 2011b), there are several advantages of using ChatScript, the new chatbot language he created after reviewing the existing mechanisms. First of all, the wildcard symbols in ChatScript are more flexible as they can match zero, one, or more words, thus making the scripting process a lot easier and efficient. Secondly, the same themes are structured under the same domain, making ChatScript bots more focused on the same topics.

Thirdly, for every scripted response, the same words or phrases that can be generalized to fit the incoming patterns can be pulled together, so ChatScript bots are more likely to respond correctly even when a substitution in the same domain is used. Finally, an embedded syntactic parser also allows ChatScript bots to match only the keywords that appear in the predetermined part of speech position, leading to responses that are more accurate.

In addition to designing chatbots for human- likeness contests, Wilcox (2011b) has also used ChatScript to design chatbots for SpeakGlobal, Ltd., an online English learning site in Japan. Among them, Meg is a free trial version with voice synthesis and basic voice recognition (http://www.speakglobal.co.jp/member/bots/meg_trial.php ). Rather than being open in topics, Meg also specifies the topics she can converse about.

20

The above five chatbot mechanisms are mainstream chatbots available online. But chatbot mechanisms might not be limited to the types reviewed here because there are variations of the categories above and new chatbots like Goh and Fung’s (2008) Artificial Intelligence Natural Language Identity (AINI) might adopt more mechanisms than just keyword and pattern matching. It is apparent that chatbots have progressed after the introduction of ELIZA, and with the more complicated systems they are structured now, more successful human-computer interactions might be possible.