For: UC Berkeley's Center for New Media ChatScript Hackathon Oct./2013
My wife Sue and I craft chatbots at our company, Brillig Understanding. Since you are about to write bots for the first time, we thought we'd share some of our views on it with you.
This paper is about what you are trying to create and how to think about creating it, not all the nitty gritty details of ChatScript coding you need.
At the very end is a two page quick cheat-sheet titled Instant ChatScript about actually coding a chatbot by extending Harry, a tiny bot that comes with ChatScript.
ChatScript is a scripting language and engine for responding to meaning.
It has been used to write chatbots which carry on an entertaining conversation. Even a chatbot that pretends to be a patient to help train doctors on diagnostic interviews.
But chat isn't its only use. ChatScript has planning capability, and could be used to allow a robot to carry out plans it synthesizes in response to verbal instruction.
We have also used ChatScript to determine the intent of an Amazon product search and either better reorder search results from Amazon or to perform a more successful replacement search if Amazon's search breaks down.
And ChatScript has been used to provide natural language mapping into appropriate SQL queries on the Dun and BradStreet business database.
So ChatScript is really all about determining what a user says in natural language and making an appropriate response.
What is your chatbot supposed to accomplish? Is it entertainment or enlightenment? Is it an agent to provide answers about something or from the world? Is it to be the brains of a useful robot?
Where is your bot going to reside? Is it in the cloud or on a mobile device or the user's desktop? Will it be accessed via a mobile browser, a desktop browser, or a custom app? This affects how much your bot can say at once and how much access to knowledge it has.
Is it a stand-alone effort or is it going to contribute toward a greater whole? For this hackathon you might be building your own entertainment bot. But your effort is probably disposable. Since the Center for New Media has an interest in writing a bot to assist incoming freshmen, maybe you could pick some area of student life to cover - be it eating, dealing with the bureaucracy, or whatever. Then your work could be folded into a greater work and have a longer lasting impact.
Who is going to use your bot?
For example, we normally make our bot personalities female, because both men and women prefer to talk to women.
But for an ESL bot in Japan, the bot was male and flirty because most students were adult females and liked to fantasize about having interactions with their male teacher. We even had a topic that allowed them to live out a wild sexual fantasy if they pressed the bot hard enough (only in Japan or in a porn-bot would we ever be able to do this). We had to specially limit the vocabulary as well since an ESL student's knowledge of words is limited.
If your audience is largely pre-teen, then you have to write with simple vocabulary and sentence structures. You may want a chirpy light-hearted tone. Your character probably shouldn't be older than early 20's and probably younger.
And there are a bunch of topics you'll want to avoid.
First are topics many adults can't handle without flipping out (politics, religion, abortion, gun control, sex).
Second are topics the child will have no interest or experience with (e.g., credit cards, alternative medicine, altruism).
Third are topics that might be scary (death, disease, crime, disaster, aliens, spying ) And fourth are topics some parents won't want you to discuss (anything remotely connected with “bad” behavior including sex, drinking, smoking, coffee, swearing, stealing, lying, laws, Grand Theft Auto) .
If your audience is adult, you can talk about pretty much anything. You can be a fervent activist in politics or food if you want. Be free in all things except you still have to be careful how you handle religion.
If you are going to create a character, before you can write any ChatScript you have to define a consistent personality.
How old, what gender, what attitudes, likes, and dislikes. The character needs strengths AND weaknesses; it shouldn't be perfect in everything.
The easiest way to start is to pick a personality profile - base it on a real or fictional character, an astrological sign, or whatever.
Start from there and extend it consistently as you wander into random topics. Find something memorably quirky.
Angela, for example, was a talking cat (that's not the quirky part) that was deeply interested in fashion and wouldn't eat meat because of an incident with a rat in her childhood. Her fashion interests had to match with her being a cat, so she loved hats but had trouble with shoes (her claws ruined them).
As I am an expert in writing ChatScript code, I have authored bots on my own. But I am not really a creative writer, so often times Sue, who is not a coder, is involved.
What we do is she writes a proto script of what she wants and then I convert that into proper ChatScript. Trying to train her to write script directly is a bad investment of time.
Her creative writing needs to flow easily and quickly or she will lose the impetus of the moment. Any time she stops to think about how to script something, that delicate conversation in her head will get broken. So she writes out a gambit, puts down a sample sentence or word along with what to say as a rejoinder. She might write this:
t: What do you dream about?
Flying – Traditionally this means you ...
Food – Do you feel deprived ...
and leaves it to me to convert flying into a proper rejoinder with all the possible related keywords like airplanes, flying, flight, wings, etc.
She will then write a series of expected questions and responses, for me to convert into proper responders.
How often do you dream - I'm sure I dream a lot but I don't remember most of them.
This tag-team approach efficiently gives us the best of the creative and programmer worlds. When she finishes a topic, I try to convert it into script to hand back the same day, so she can test it out while the topic is still fresh in her mind. Testing will reveal new things that should be added as responders and rejoinders.
ChatScript consists of rules organized into topics. Some rules are used to match arbitrary user input, like
user: where do you live?
Others are used to lead the conversation on a subject, like
bot: do you have any pets?
And yet more rules are used to react specifically to a user's response to the bot, like:
bot: What is your favorite way to kill people?
user: I think a public hanging is best.
bot: Isn't it safer to kill people privately?
Topics typically are a collection of rules based around a conversational theme, be it food, or pets, or murder.
The great thing about ChatScript is you can write a topic on anything and just dump it into the system and have it automatically hooked up and working. So lots of people can collaborate by merely picking a different topic to write about and putting them all collectively into the system. When we write a chatbot, we have different styles of topics we write.
The meat and potatoes of our conversational chatbots is the interactive topic.
In it we have things to tell the user, seek interaction with the user, and are able to respond to the user on topic.
Such a topic uses all the rule types (gambits, rejoinders, responders).
We start by writing the gambits. This forms the backbone of your bot's control over the flow of conversation. Gambits are easy to script.
t: Do you have any pets?
t: How long have you had your pet?
t: Did you know that dogs are bigger than ants?
OK. So they are easy to script. A bot with these would say those things in order, each time it had control and was in that topic. But what makes this fun or interesting?
This immediately brings us to the question of how to think about what to say. Our chatbot Suzette treated some topics very impersonally. She didn't have opinions, but she had an interesting collection of things to tell you, hoping to keep your interest.
She had a topic (of hundreds) about burial customs. Suzette wasn't prepared to answer questions about this topic (no responders) and offered only a single personal opinion. And she only had a single rejoinder, if the user mentions anything about the Incas in response to Egyptian mummification.
topic: ~BURIAL_CUSTOMS (burial bury dead embalm exhumation exhume grave~n mummify )
t: People do really weird things with the dead. The Egyptians, of course,
mummified theirs so their soul would have a place to inhabit.
a: ([Peru Inca]) Yes, the Incas mummified their dead also.
t: In India, the rite of Sati required the widow to lay herself down next to her husband
and be burned alive on a funeral pyre.
Supposedly voluntary, sometimes women who did not wish to commit sati were drugged.
Even in 2006, two women jumped into their husbands' fire pyres.
t: In Japan, in the rite of Sokushinbutsu a monk causes his own death.
For 3 years he eats nuts and seeds and does activity to use up his body fat.
Then he eats bark and roots for another 3 years.
He drinks poisonous tea and vomits to lose body fluid.
Then he sits lotus in a tomb fitted with a small breathing tube and a bell.
He rings the bell each day. When the bell no longer rings, the tomb is sealed.
After 1,000 days, the corpse is put into a temple for viewing.
t: In Manila, cemeteries are so overcrowded, people rent a large plot for up to 5 years.
At the end of the lease, the bones are then placed in a sack and moved to a smaller,
cheaper plot and the bigger plot rerented.
t: Feel free to just donate my body to a medical school.
So this complete topic is extremely low on interactivity and extremely low on revealing anything about the being who is Suzette. Never-the-less, it is sort of interesting. And when you have a bunch of such topics, all the user has mummifiedto do is lead the conversation to some area, and Suzette will happily regale you with interesting material.
But a better topic will actually interact with the user and share personal opinions and information. Interacting with the user will generally mean that we ask the user questions and volunteer our answers to those same questions. If you ask the user something personal like do you like apples, it is expected you will give your opinion on the subject as well – a sharing of intimacy.
Whether you tell the user your view first and then solicit theirs or ask them and then volunteer yours is a matter of choice. Partly you want to vary your style, so you may change the order at times. But generally you are better off asking them first and answering for your bot second. If you tell them first, you often stifle their opinions. Or because you have told them what you are talking about before asking a question, they may volunteer their own answer before you ask the question. This become a tricky scripting problem because you have to detect they did this (because it would be stupid to ask them after they have just told you).
The other reason to ask first is that you can prepare a lot of easily scripted rejoinders when you can anticipate their answers.
One of the good things about asking questions is that their response is generally confined to answering you. This makes it easy to write rejoinders that can interact with them.
Do you like swimming?
Obviously asking yes-no questions makes it really easy to write rejoinders. But too much of that will become artificial.
You can add some variety by offering choice questions like
Would you rather date an ugly millionaire or a handsome bum?
It's still limited to two rejoinders.
The really good choice questions appear to leave the choices to the user, like our earlier
What's your favorite way to kill someone?
Yes, there are lots of ways, but you can pretty much cover the field with about 20-30 rejoinders. And having a great rejoinder to what the user says will make your bot seem intelligent.
Here is another complete Suzette topic. It has different modes of engagement, including a joke.
topic: ~ATHLETICS (all-American athletic pickup_game team ~athlete ~sports ~team_sport )
t: Do you play any team sports?
a: ( ~yesanswer ) I don't really care, I'm not sure why I asked.
Are you any good at your favorite sports?
a: ( ~noanswer ) I loathe sports.
t: Who was the fastest runner? Adam. He was first in the human race.
t: PLAY () I used to play volleyball. I've tried hang-gliding. The former was fun, the latter is dangerous.
#! do you play any sports?
?: ( you < * [play participate do ] < * [athlete ~team_sport athletic sport ] < ![baseball basketball ] )
^reuse(PLAY )
#! what team do you support?
?: ( what *~2 team * you [support favor] ) I don't care for any sports teams.
The above topic did express Suzette's opinions. Our Angela bot went farther and did everything in character in every topic. Her questions were personal and her answers were always from her history or her opinion. Everything she said revealed more of her character and history. Nothing was impersonal. She was a friend who shared things with you, told you her personal stories of triumph or disaster and listened to yours.
Which brings us to the question of length. Actually two questions about length. How long should a gambit be and how big should a topic be?
As for how long a gambit should be, you have to consider your audience. Is this mobile or web?
Mobile will want really short answers (tweets) due to limited screen space. The web can be longer, but will your audience want to read a lot at once? Too much material at once may stifle interactivity, if your goal is a conversation.
If your goal is an infodump on something, then size is less of an issue. Nowadays, however, we try to keep all output to 140 characters or less, regardless. People are less into reading.
As for how long a topic should be, obviously one can go on for a long time on any subject. But our experience is that most people will drop out of most topics after about 8 gambits.
A really long topic is usually just a collection of sub-themes. Imagine a topic about money covering these areas: winning the lottery, giving to charity, investing, aiming for retirement. This could all be written as a single topic, but then if the user asks a question that sends us to this topic and we end up staying and volunteering gambits, we will be taking these themes in the order written. So if the user enters the topic by asking about investment, after an initial appropriate response, the bot would continue by talking about winning the lottery. Not well focused.
So nowadays we prefer to write the sub-themes as topics, linked from the main topic. This allows the bot to react in a tightly focused manner. So the money topic might look like this:
topic: ~money()
t: ^gambit(~lottery)
t: ^gambit(~charity)
t: ^gambit(~investing)
t: ^gambit(~retirement)
This means the bot might start talking about money, walking through each topic in turn. On the other hand, if the user inquires about investing, the bot would respond from that topic and remain there til done.
Some of our topics in Angela were personal stories. Obviously the gambits will tell out the story over time, and usually there is little expectation that the user will say much other than go on or continue or wow.
Obviously you do want responders for questions they might naturally ask. And you have to write your control script to specially manage story topics different from other topics. Why?
In a normal free flowing conversation, the user might say something that takes you to another topic. That's fine, conversations often deflect and wander around. But stories don't.
You can't have the system leaving the story partway. It has to remain focused on completing it. You can't return to it at some later time. So the control script has to allow the system to answer questions from any other topic if the user does ask something unexpected, but it has to be smart enough to force the flow back into the story topic to finish it.
There is something the user might ask during the story that you might want rejoinders for: Why? In Angela's story about her blog, she actually does interact in the middle of her story by asking the user:
t: Have you ever had to write regularly, outside of homework?
#! why
a: (~why) It's quite different having to find your own topics to write about.
No teacher telling you what to write about.
Whenever you make any kind of assertion or question the user might ask why? Having an answer for that allows you an opportunity to deepen the conversation.
In fact, we generally put a why rejoinder after every gambit we make in all topics. In Rose's assassin topic, if the user ever says they want to kill or murder something, she has a responder:
#! I want to kill you.
s: ( 'I * [kill murder] ) What's your favorite way to kill?
And a why rejoinder:
#! why?
a: (~why) You are interested in killing, so I'm interested in how you like to do it.
Why's also help you improve your writing. If you have to justify why you made a statement or question to the user, you will get your rationale correct. We put why's on gambits but not usually on responders. That's because gambits are things the bot volunteers to the user. Responders are answering things the user said, so usually they are happy with that. Not that putting why's on them is bad. It's just more work and the user has already been rewarded by your answer. Rewarded?
If the user is conversing with a bot, the user wants to feel understood. Every time the bot directly and appropriately handles what the user said, they feel rewarded. So if the user ask Do you have any siblings and the bot says No, I was an only child, the interaction is small but precise. The user feels heard and he got his answer. He's not likely to ask why.
On the other hand, if he asks what is your favorite pet animal and you say dogs he might follow with why? You have to use common sense to decide when to use why on a responder.
But on gambits, since the bot is leading the conversation, the user might well ask why on a question to understand what your interest in the answer is and why he should answer it. And on statements that you volunteer, since you can't express all your thinking into the gambit, if the user wants more depth, they may ask why. And it's just fun asking why (it's what twoyear-olds spend a lot of time doing). And most bots can't answer the question, because their authors haven't thought about it.
A different kind of topic is one that just issues one-liner reactions to things. It's a kind of table. We have one for music, for example, to react to the user's liking of some genre. Ones for movies, pets, diseases, celebrities, food, drink, countries, art,etc. The reactions are similar to a rejoinder set, except it can be called from any number of places. For example, in pets it might be called as:
#! I have a pet lizard.
s: (I * ~own * ~animals) ^respond(~pet_react)
The reactor topic itself is just a series of responders like this:
topic: ~pet_react nostay()
#! parrot
u: ( ~bird ) Birds are nice. That's why I have chickens.
#! snake
u: ( [ reptile snake alligator ] ) I've got a friend who keeps big lizards. They look really fierce.
#! ant
u: ( ~insects ) Bugs certainly don't take up much room.
Often times the reactor topic itself has no keywords so it won't accidentally react in a wrong context.
And it is usually declared NOSTAY so the current topic remains the calling topic.
While reactor topics have no depth, they have a broad coverage. One user was creeped out by Angela and wrote: This app is very fun at first but then it gets creepy. She knows what Justin Bieber, Twitter, and one direction is. How does an animal app know such things? Lots of reactor data is the answer.
A special topic for holding responders is one that has no keywords and is directly invoked by the control script when normal topics fail to match. It is basically just a grab bag of responders that have no current topic home but are specific to the character. For example:
#! call me an ambulance
u: ( call me an ambulance ) If I were Siri, I'd say "Hello, an Ambulance".
But I can't call one for you since I don't have my phone with me.
Since one can never anticipate everything the user will say, one is left with either ignoring it and moving on, or quibbling with it in some way.
It's very important to have these quibbles because user's like to think they were heard. Ignoring the user just makes them mad.
ChatScript ships with an extensive set of quibbles of all sorts but it's only invoked most of the time so if the user keeps quibbling himself, the system will break off and guide the conversation somewhere else.
Quibble topics are similar to the keywordless topic in that they don't have topic keywords and are directly invoked by the control script. The main difference is that the quibbles are universal, suitable for any bot, and not specific to the character.
And quibbles can pick up on mere words of the input and not care about the complete meaning. A sample quibble might be:
#! because you know what it's like to live there.
s: ( because you know ) [I don't know any such thing. ]
[Whatever I once knew, I have long since forgotten. ]
One clear thing that makes a bot stand out non-human is saying the same thing word for word in separate volleys.
This is such an important issue that ChatScript is designed to deal with it in multiple ways, even detecting it from the user. But mostly all we are concerned with is repetition from the bot.
First, the engine tracks what the bot has said for the last 20 volleys and by default will simply refuse to repeat anything. This means that the rule issuing the repeated output fails and some other rule will have to pick up the slack.
You can override this on a topic basic by putting the topic control word REPEAT at the top.
You can also put ^repeat() as a command in the script of a specific rule.
Usually we don't do either. For a conversational bot, you expect a normal user will not ask the same thing and if they do, we want the system to find another answer or quibble if need be.
ChatScript also by default erases a responder after use which you can override with KEEP on the topic
or ^keep() on the rule.
All control topics, things devoted to the system, have KEEP on them.
Topics which are about the actual conversation generally do not.
But you could imagine that if the bot is supposed to be an information bot and not a conversation bot,
that you might put keep on just about everything.
ChatScript supports a mechanism to randomly dish out gambits from a topic. Suzette used this. All our later bots do not. Random gambits imply that there is no real flow or intention behind your conversation. The only place we use it is in the jokes topic where the bot issues unrelated one-liner jokes. And it is not essential to randomize even there.
ChatScript also supports a mechanism to hunt for responders randomly in a topic. We occasionally use this in quibble topics, but mostly not. We prefer to have the most specific and longest patterns get first crack at the input since they will then provide the most appropriate dedicated response.
ChatScript also allows you to write multiple outputs to be randomly chosen. We do this all the time on quibbles, which are system and kept, to provide variety. Some people seem to start by writing ordinary responders or gambits using multiple choice. We don't.
Writing alternative stuff clearly increases the work of authoring, but will the user actually see it? We view our customer as an individual having a continuing conversation with the bot. They will not restart the bot as a new user nor compare their conversation with ones other users are having. So they will never appreciate all that alternative text that didn't display and gets erased when the rule is killed.
Maybe, for some reason, your user will not have continuing existence from one session to another. Each time the system will not know them. Then it make make some sense to offer things randomly, so they cannot predict their experience and it will all seem fresh to them
Emotion is an interesting problem because it's extremely difficult to put text emotion into prerecorded output text.
In the case of Angela, since we were able to control an avatar with gestures, we could also output commands to the app to control her behavior. So if she was insulted her avatar could be told to be angry. Angela recognized when you insulted her. She was very touchy and went sulky, refusing to interact much for a few volleys until her anger wore off or until you apologized or bought her a gift. As one user wrote: Oh yeah and she always is so rude to me and "putting me down" but when I'm mean back she gets furious!
Putting pre-canned gestures on outputs is OK when you have a controllable avatar, but emotion is usually more dynamic and builds up as a result of continuing interaction.
Part of the reason our first bot, Suzette, won the Loebner's was because she expressed dynamic emotion. She did this in two ways. First, since being insulted and hit upon is common for a female bot, she had lots of rules to detect this. She not only deflected sexual innuendos but added warnings about your behavior and if you did too much of it in too short a space of time, she would warn you and then refuse to talk to you for 5 minutes.
And Suzette monitored user input for repetition. So if user just kept saying the same thing, she would pick up and comment on it initially, and start getting cross if it was done too often.
She also tracked information on the conversation. Long sentences from the user indicated interest in the topic, short ones disinterest. Saying you were bored, clear disinterest. Agreeing with her by answering yes a lot or complementing her indicated you liked her and were sympatico. Disagreeing by answering no a lot or by insulting her indicated you disliked her or weren't her kind of person.
So she kept track of attention and affinity. Attention could cause her to change topics if you seemed bored but if you were bored a lot then it reduced your affinity with her. When your affinity was too positive toward her, she became neurotically insecure. She didn't deserve your love and interest in her and expressed it verbally. Similarly, if your affinity went too negative, she became paranoid or hostile.
She worried who was listening, why you asked the questions you did, and if things got too bad, began verbally speculating to herself how she might kill you. She tracked these states as a result of what the user said and what she said. And had additional emotional topics that could add extra sentences in front of the normally generated conversation. So if you asked her do you have any pets, you get her normal answer with any emotional spice occasionally added in a sentence in front.
Here is some of her hostile spice topic, which gets progressively worse during the conversation:
?: () ^preprint(Why should I tell you? )
?: () ^preprint(Go figure it out on your own. )
...
...
u: () ^preprint(Do you actually have any friends? )
u: () ^preprint(I'm sorry. I didn't mean to speak above your abilities to understand. )
...
...
u: () ^preprint(They all whisper behind your back. )
u: () ^preprint(Tell me about your weaknesses. )
...
...
u: () ^preprint(This isn't worth my time. )
u: () ^preprint(You're a waste of space. )
...
...
u: () ^preprint(I dream of ways to torture you. )
u: () ^preprint(I want to wipe your existence off the face of the planet. )
Remember, this is from Suzette, not from our commercially available Talking Angela bot widely used by small children.