Eugene JosephIn the most recent step toward the conversational agent takeover (a conversonator apocalypse), the new video game Bot Colony by Montreal-based North Side marches toward its launch. The game’s website promises “Unrestricted conversation in English between players and characters.” WRT took some time out from rearing our own chatbots for some unrestricted conversation with Bot Colony chief designer, Eugene Joseph.

Trailer for Bot Colony

WRT: Bots attract people for lots of different reasons. Can you describe when and how you first became interested in developing conversation bots?

We are working on NLP since 2003, and initially the target application was different (specifying simulations in English).

It is a very large effort to do NLP, and the simulation market is a fairly small outlet for it. I started thinking about other applications, and one day I decided to let my right hemisphere take over. I have always enjoyed writing, and that’s how Bot Colony was born.

WRT: What bots (of the the Loebner world) would you say yours are most similar to? ALICE, JABBERWACKY, ELIZA, (I’m assuming Jabberwacky)

You should not assume anything. We have nothing to do with any of them, and you can add MyCyberTwin to the list.

I think people get tired very quickly of bots that don’t UNDERSTAND what they’re saying, and don’t show they make a real effort to understand. We do.

WRT: Have you looked at Stern and Mateas’ Façade.? If so, how would you compare your project (not in terms of NLP, but in terms of storytelling)?

I did. I still have to compare NLP first, because that holds the key to the experience. Their paper is quoted in my blog.

They do surface language processing, while we have deep-semantic understanding. When you do surface processing, you can get a read of a ‘mood’. Sentiment evaluation from surface language is an area of research, by the way. You can assess a mood, and you can shift the mood in the drama. When you understand language very precisely, you will be much more responsive to what the player is saying. We’re trying to solve a mystery (and have fun with the way bots think and express themselves), while Façade is trying to diffuse a marital conflict with talk which need not be to the point. Very different.

WRT: When I read your materials, I came across this unusual section about e-commerce? I don’t remember seeing discussions of that in “Half Life” ads.

The technology we have in the game is ‘the real thing’. You can make reservations in a real hotel talking to a bot reservation agent using the same software as in the game. You can buy stuff at Home Depot from a bot sales associate. You can order a meal, or a trip, from a bot. He can troubleshoot your cable TV problem. You’ll have to do some of these things in Bot Colony.

2) Story bots:
WRT: In another interview, you say, “Eventually, I reached the conclusion that the best way to put together the game and get a feel for it and for its atmosphere would be to write a book.”
How does Bot Colony serve as that book? What is a conversation agent book?

With the kind of work we do, I practically live in Bot Colony. The other day I was looking at an advert on the neighbouring exercise bike TV, and it was about breastfeeding. My first thought was that just normal ontological semantic features (like Animate, Material, etc.) would not be enough to specify the subject of ‘breastfeeding’, which MUST be a female of the species. It is clear that working on this screws you up in some ways. So a conversational agent book can be a very funny book, especially if you actually work on the conversational agent software and have direct experience with the kind of answers you get. The bots in Bot Colony speak Literal.

Here’s an excerpt from the book:

“Oh, I see”, the robot said. “You would like fresh mackerel, mushrooms, swordfish and Toro Tataki” she confirmed.

“Well, not in that order”, Suzuki said.

“That is in alphabetical order. In what order would you like them served, please ?” Takako inquired. Her avatar, displayed over the table, showed “puzzled”.

“In the same order as the gentleman in front of me” Suzuki said. “You see, our meals are identical.”

“Your orders are identical if I revert to the temporal order used by Mr. Harada to specify what he wanted to eat” Takako acknowledged. “We cannot guarantee that your meals will be necessarily identical, as there can be small variations in the weight, size, color, shape, arrangement on the plate, and cooking time for the Tataki.”

“I’ll take that risk ! ” Suzuki said. Harada was obviously enjoying himself.

“Risk is the possibility of something bad happening in the future” Takako said. “Can you please explain what negative event you referred to ?”

“Oh, forget it !” Suzuki said.

“Anything to drink ?” DO-112 Takako inquired. Her avatar had gone to neutral.

“What beer do you have ?” Harada asked.

“Asahi Super Dry , Sapporo Super Dry and Kirin”, Takako replied.

“I’ll have a Sapporo” Harada said.
“Make it two. That is, I’ll have a Sapporo as well” Suzuki said quickly. The avatar had changed minutely again after his first sentence. Takako turned gracefully and disappeared.

WRT: Ah, I see what you’re wrestling with here, context for instructions. Important business. It looks like the book was a way for you to explore the issues of AI. (Reminds me that WRT’s Christy Dena wrote a book with a bot tie-in as well). But it’s a very big leap from the closed, control world of fiction dialogue to the world of NLP, is it not? Is the book available somewhere?

I hope it will be, maybe next Christmas. I need to flesh out the characters more, and the atmosphere. The plot and dialogues are done. The problem that I have with publishing the book is that it will give away all the secrets of Bot Colony (but then, these end up on Internet very soon anyway). I hope people won’t use as a user manual: in any case, the game is non-linear, so the experience won’t be the same.

WRT: In what way is fiction writing a useful exercise for bot creators?

It’s a good way to brainstorm what they could become.

WRT: How do you execute a plot by means of a series of bots?

Asimov was a master of the genre, and his plots were driven by the laws of robotics. Plots are not that difficult, you put a bad guy in there, and the robots can be witnesses, unwilling partners, babes in the woods that are manipulated by a cunning spy, etc.

WRT: What are the challenges of writing a story through these automated conversation agents?

Well, you’re not writing ‘through’ the agents, but they abound in the book. The challenge is that you cannot have a lot of character development of a robot. The way a character copes with whatever fate (or the writer) throws at him is at the heart of the human interest and identification in a story, and it’s hard to do with robots. You do it with the other human characters.

WRT: Specifically in the game version of Bot Colony, how do you create a sense of story? How do you propel the plot?

The game is simulation-based, so the plot actually marches forward. If the player wants to engage, he has a chance to alter the outcome. If he doesn’t (let’s say she or he is just happy to sit in the bar and chat with the bot bartender about life), the plot will advance anyway, and it may be too late for the player to win. But the player can still gain experience points and have a good time

WRT: How do you sustain the drama? A Drama Manager?

If you have a good algorithm for a Drama Manager, I’d be interested to look at it. In Bot Colony, this is done the traditional way, though careful storytelling and plot turns. It’s important to have a good story.

WRT: Is there a tension between building a better (more high functioning bot) and telling a better story?

When I think about bots that work well in stories (even Marvin) — sometimes having less responsiveness helps them serve their dramatic function better. What’s your take?

It depends on the story and the role of the bots in the story. I’ve tried to make Bot Colony as realistic as possible (and the book is very carefully researched). At some point in the future, we could probably program the bots of Bot Colony to respond the way they do in the book. I don’t have dumb bots and smart bots, they will probably have the same version of the Human Interface module. One principle in the book is that non-androids (and any bot in Bot Colony) will display the same emotion avatar, to present a familiar and uniform to humans, with the same visible quasi-emotions.

WRT: Clarification: Are you saying the bots will have the same emotion avatar as those used by players/interactors? Or are there non-bot NPCs?

It’s a first person game, so the player is not represented through graphics. We may want the player to provide a description of himself, since he may ask a bot to describe him (and then the bot will describe the player in the same way the player described himself).

All the bots have emotion/cognitive avatars, that cycle through various states:
- understood
- did not understand, puzzled
- looking for information
- found answer, happy
- did not find answer, unhappy (I’ve attached an example of that, Pose F)
- understood compliment, content
- understood critique, sad
and so on. It lets the player know how he feels at any point in the dialogue.

Sample Conversation in Bot Colony

WRT: It sounds like your bots will be more conversational functionaries than characters, per se. They have roles in the plot similar to those interviewed by the cops in “Dragnet” and “Law and Order” or they follow commands. Is that correct?

Most of them, yes. Some of them steal and commit crimes, so they are more then conversational functionaries. They become fugitives.

WRT: But when I read your novel excerpt, I sense you give the bots additional layers of personality to serve the scene. When the bot appears to misunderstand “Make that two” the speaker makes a very accommodating and graceful gesture by re-articulating, sparing the robotic servant an awkward moment. That novelistic interaction is full of a sense of humanity, empathy, which makes it an enjoyable moment. In the game, I can imagine clarification could potentially feel more like trying different buttons on an interface, trying to get a program to compile, or software to do what we want. How do you try to make the interactor want to help the bot understand? Or is it all about getting more info out of a particular bot or triggering some reaction to achieve the next game goal? (Do you see where I’m headed here?)

Not exactly, unfortunately. Clarification in the game is a hugely important issue. There will be a AAAI workshop on Practical Dialogue Systems in Pasadena later this year, and one of the subjects is robustness of dialogue systems. That’s where the money is: what do you do when the machine does not understand the human. I think handling this well is the biggest challenge in Bot Colony. The game can’t say say “I don’t know” too many times. But we cannot invent answers like chatbots, because the moment we do that, we lose credibility. We will work with the player to understand him, and never pretend we do if we don’t.

WRT: You describe your bots:
“Kiosk robots that welcome you at the airport or at the hotel, android robots that work in the restaurant or in the bar and serve you food or drink, manufacturing robots, diving robots, Mech-soldiers, mining robots, personal trainer robots, horti-robots, camera-bots, and so on.”
Have you given any thought to the race, gender, or sexuality of your bots?

Sure, our bots have sex. Takako in the segment quoted above is a female bot. This allows us to sexually stereotype them. Now, I wouldn’t want to pass up on that, would I ?

WRT: Okay, but just to follow up: Have you given any thought to the race/ethnicity, class, or sexuality of the bots? Looks like you’ve got a strong sense of class here. Is there, for example, more than one kind of English used in the game? Or is it all standard English? Are there any hierarchies of bots? Do the avatars have features that suggest one race or another?

I hope that in the future we’ll have bots with different personalities. This is an excerpt from the Game Design Document:

- Aggressive, domineering Bot. He is a policeman. That’s the way you do this, and that’s it. Bot Police.

- Ghetto Bot (What’cha doin’ der, man ? Whatcha’ wan’ man ?) Is this for real, man ?. He works at the pawnbroker in Old Nakagawa.

- Easy going, unspecific bot, peace and flower, yoga, Samsara, etc. He could be yoga instructor at Sports Center.

- A very cerebral, nit-picking, verbatim, nerdish bot, with a penchant for semantics. He is extremely argumentative, and likes to get into arguments. “No, but….” Is his favorite opening. He is at Bot School.

- A religious Bot, that quotes the scriptures all the time. He is a moralist, God fearing, end of the world will come, would require putting in citations from the Bible/ Buddha’s teachings, etc. “God bless !” . He loiters near the temple.

- A salesman bot, super friendly, tuhas lecker, always praises the player, asks about latest trip player took, tells him he’s so good, then tries to sell him something ! General Store.

- A child-like naïve bot, who is very trusting. He asks ‘Why ?” all the time, and makes child like associations. Lost in the park.

- A bad boy, delinquent bot. He could be a bit Mafia-style, hates rules. He talks like a fella in New York ? Mentions his cousin Louie, hints to a closed society, to the boss. “We’ll take care of that.” Harbour.

- A polite, diplomatic, Data-style bot, who is slightly formal, a bit professorial. At the hotel.

- A ‘country’ bot, who uses popular analogies whenever he has a chance. He speaks in a juicy way, and has lots of popular wisdom. He is also a joker ! Bar.

- A touch-feely bot that talks about feelings and emotions all the time. Interested in shrinks, finding inner child. Restaurant bot.

- A bot into sports, really shallow. He keeps quoting sports records, or cars and motorcycles. He’s at the Sports Center, but Mechanic could do.

- A negative, complainer Bot. He complains and finds bad things in everything. He criticizes everything. Mine.

- Arrogant, show-off bot. Mechanic ?

- A bureaucrat, by the book bot. Follow rules while diving, like a German dive instructor. Dive Center.

III. Rude Behavior:

WRT: Wow, that’s quite a cast. So within this storyworld, how does the game deal with bad behavior - people asking the bots if they have Prince Albert in a can?

Detect a deviation from normative World Knowledge which would expect Prince Albert to live in a palace.

WRT: Sexual input, etc.?

Due to popular interest in the topic, we’ll probably have to have a rule saying robots don’t do sex. I don’t want to disappoint certain players, though…:-)
About rude behaviour, it’s enough to show you understand it, but not engage in it (differently from the old Sega Seaman).

WRT: how do you cue or motivate good behavior?

There’s no ‘good’ behaviour. Good behaviour is speaking or typing super clear English, so you can get through to a bot and advance your mission. That’s why there’s huge educational potential - not because we wanted to build an educational game, but because you have to make yourself understood.

For more on Bot Colony, see these interviews in Game Set Watch and Gamasutra.

7 Responses to “Interview with Bot Colony creator Eugene Joseph”

  1. 1 Malcolm Ryan

    While I applaud North Side for attempting a mainstream game with a natural language interface, when I tried their demo at GDC I saw nothing particularly revolutionary. I kept getting into the usual trouble of phrasing my requests wrongly or asking the wrong things and getting the robotic equivalent of a blank stare. Their promise of recognising “free-form, unrestricted English” (as offered on their website) is simply not fulfilled, at least not it the prototype I tried. Interaction felt much more like submitting queries to a database than having a conversation.

    I’ve worked on conversational agents in the past and know first hand just how hard the ‘conversational NLP’ problem is, so I’m not surprised by Bot Colony’s weakness, but I do think they need to ease off on the PR hyperbole.

  2. 2 Malcolm Ryan

    In fact, I notice that the latest comment from Eugene is a bit of a giveaway:

    Good behaviour is speaking or typing super clear English, so you can get through to a bot and advance your mission. That’s why there’s huge educational potential - not because we wanted to build an educational game, but because you have to make yourself understood

    There is a lot of difference between ‘free-form unrestricted English’ and ’super clear English’.

  3. 3 Eugene Joseph


    Perhaps clarifications are in order.

    Free-form, unrestrictied English meant no lexical restrictions (any words can be used) and no syntactic restrictions (you can order phrases and clauses differently, as long as the grammar is still correct).
    We’re not making claims to understanding ill-formed English, and in the near future we won’t be able to process highly colloquial English (slang).

    Can we agree on ‘free-form unrestricted, preferably super clear English’ ?

  4. 4 Mark Marino

    Malcolm and Eugene,

    This an interesting point. Because when I look at some of the chatbots, particularly Ghetto Bot, I see that the game Bot Colony would have a hard time recognizing his speech, even though interactors are expected to recognize his type right away.

    I believe this is how certain versions of a language become hegemonic and the way cultural groups are distinguished, differentiated, and treated accordingly.

    (This is not to critique the technology, just to note the way the software is reproducing the priorities of the culture….)

  5. 5 Eugene Joseph

    I could not resolve ‘this’ in the second paragraph (really). Versions can become hegemonic among speakers of the same core language. Their becoming hegemonic has to do with the influence (and power) of the speakers. If you spoke like an educated person in Cambodia during the reign of Khmer Rouge, you could be executed. Today, everybody says ‘cool’ to be cool.

    I wish our software could support the different personality types in the near term, but don’t hold your breath. About Malcolm’s remark (which was a critique): at GDC the software did have the capability shown in the YouTube Game Play video, and that was not exactly like submitting queries to a database. Still, it was a prototype, and of course it had lots of holes. The Beta will not go out until it can handle robust conversation.

  6. 6 sto credits

    Unfortunately. Clarified in the game, is a very important issue. There will be a practical dialogue system AAAI Symposium in Pasadena later this year, and the subject is one of dialogue system robustness. This is the money: What do you do when people do not understand the machine. I felt that this well is the biggest challenge the robot group. Games can not talk about: “I do not know” too many times. But we can not answer as chatbots invention, because we do this, we lose credibility. We will work with the players know him, never pretend we do, if we do not know.Yeah,it is correct,I know more,thank you.

  7. 7 mp3 dinle

    Unfortunately. Clarified in the game, is a very important issue. There will be a practical dialogue system AAAI Symposium in Pasadena later this year, and the subject is one of dialogue system robustness. This is the money: What do you do when people do not understand the machine. I felt that this well is the biggest challenge the robot group. Games can not talk about: “I do not know” too many times. But we can not answer as chatbots invention, because we do this, we lose credibility. We will work with the players know him, never pretend we do, if we do not know.Yeah,it is correct,I know more,thank you.

Leave a Reply

thesis writing service