بسم الله الرحمن الرحيم
06 February 2018

I’ve always found it frustrating that chatbot developers seem to be satisfied with frameworks that don’t even attempt to mimic anything close to real intelligence. There seem to be two basic approaches:

  • Intent-based approaches, in which a query is mapped to a template, perhaps with some slots to be filled, for example, a query like “I need a train to Norwich” would prompt the chatbot to question the user with the goal of filling in slots relating to the departure location and the desired arrival time.
  • Tree-based approaches, where there is a tree of possibilities the user can explore, kind of like the “make your own adventure” books I used to read as a kid. This is useful in informational settings like customer support, where giving a useful responses depends on exploring a tree of possibilities to determine the user’s problem.

Sometimes these are augmented with features like the ability to remember past values for slots, that improve the perception that the chatbot knows what’s going on. But as soon as the user steps off the beaten track, the chatbot will be confused and the user experience will suffer. And what about if you want to combine the tree-based approach with an intent-based approach? So far there’s no clean way of doing this.

I believe there is a better way, and that’s why I’ve started working on my own chatbot framework. These are the design goals of the framework:

  1. The chatbot should automatically choose the next best action out of all possible actions
  2. The chatbot should learn which responses are most likely, and optimise its behaviour accordingly
  3. The chatbot behaviour should be specified by independent modules that can be combined freely

As an example, I’ll describe how these goals could work out in practice for a bot to allow users to make purchases on an e-commerce website. Imagine a user is buying their weekly supermarket shop from BigMart. The conversation might go something like this:

“I’d like some apples.”

“How about 6 Russet apples for £1.20 or 12 Golden Delicious for £1.70?”

“I’ve just remembered I need milk”

“1 litre of whole milk like last time?”

“Yes”

“Do you still want apples?”

“Yes, the Golden Delicious.”

“Ok.”

In this conversation, the bot has remembered that the user wanted apples, even after the distraction of buying milk. The point is that this behaviour shouldn’t need to be explicitly planned by the bot designer: the bot should automatically know that the user has a goal of buying apples that needs to be fulfilled. Also, note that the bot has learnt the type of milk that the user likes to buy, which saves the user time. Again, this behaviour should be built into the platform rather than needing to be programmed by the bot designer.

If the bot designer does not need to program these behaviours, what would bot development look like? We envisage three types of bot “modules” that can be developed:

  • Modules that specify the style of conversation between the bot and the user. This allows the bot designer to specify the preferred expressions to be used by the bot when interacting with the user. For example, you could write a module for bots to talk like pirates, perhaps restricted to a specific domain.
  • Modules that describe world knowledge. For example, you might try and write a bot that helps people choose the correct visa for a journey to the UK (this is actually something I’ve done before, and it’s a non-trivial problem). Such a bot would need to know about the different types of visa available, the conditions for each one, their cost and so on.
  • Modules that endow the bot with new abilities. For example, a module may allow a bot to interact with a specific API. Different e-commerce bots could then choose the correct API module for their e-commerce platform, while re-using the same style and world knowledge modules as other bots.

The three goals combined should make it very easy for a bot designer to create a bot: in the most common case, their job would be to simply choose the best modules for their application, customizing each one according to their needs.

Is it possible?

I can hear you thinking, “It’s all very well having such lofty goals, but is it achievable?” I believe it is, and in this section I will outline my proposed solution.

Chatbot platform architecture

The above diagram is a very rough idea of what the new platform might consist of. My goal is just to show how I think the proposed goals can be achieved using existing technology. I’ll try and flesh out in future posts what each component might look like, but for now, here’s a high level summary, following the diagram anti-clockwise from the user:

  • A natural language query or response from the user is received. This is parsed by a semantic parser. I’ve not found a good concise description of semantic parsing on the interwebs, which is strange, but it’s not the same as parsing (although similar) and it’s not the same as (traditional) semantics. A semantic parser takes a natural language expression and translates it to some “logical form” where the logical form is anything that a computer would naturally understand, such as a SQL query, an expression in first order logic, a JSON string or an “intent”. The typical application is to use natural language to perform database queries. Anyway, this is a well studied sub-field of natural language processing (despite its lack of a Wikipedia page). An example of an almost-state-of-the-art system is the SEMPRE system from Stanford.
  • A planning system then chooses the next best action to take given its knowledge of the current state of the world and the latest user input. This problem is also a well studied one. A very general way of describing planning problems is something called a Partially Observable Markov Decision Process, or POMDP (pronounced “pom dee pee”) for short. In fact, POMDPs have been used to plan dialogue, as described in this overview by Steve Young at Cambridge. My idea is to use Monte-Carlo tree search to solve our planning problem, an aproach described in this paper from NIPS 2010. I’m really excited about the potential for Monte-Carlo tree search to do something other than playing games really well (in case you didn’t know it’s a large component of AlphaGo). The planning system makes use of the Knowledge Modules provided by the bot designer to inform the decisions it makes.
  • Once an action has been decided upon, an action interpreter makes use the the ability modules provided by the bot designer to perform actions on external APIs, or passes on a logical form to the next system to send a response to the user.
  • A natural language generator interprets the logical forms and sends the response back to the user. The generator can make use of the style modules to determine the best expression for each logical form.

Hopefully this is enough to convince you that the plan is not entirely crazy. Each component is well studied (at least in a research setting), so it is not too far-fetched to assume that they can be put together into something useful. The biggest uncertainty in my mind is around the planning system, and exactly how this will work effectively. I plan to flesh that out in a future blog post.

Some readers may be disappointed that I’m not proposing some new-fangled deep learning technique to solve this humongous problem. In fact, I’m pretty much proposing the same good old fashioned AI techniques that were popular in the 70s and 80s. Actually I think systems built in that time period got a lot of things right, but the individual components were not developed enough to make the system as a whole a success, at least when applied to a general setting. In fact, in some cases, the improvements in the individual components are because of algorithmic developments like deep learning, along with the abundance of data and computing power. There is definitely potential for making use of deep learning to improve the three major components of the system:

  • Here’s a paper on using deep learning for semantic parsing
  • Deep learning was a large part of AlphaGo’s success so it can definitely be used to improve planning. Here’s a paper on using deep learning to solve POMDPs which happen to be Atari games (what is it with the games?).
  • And here’s a paper from NIPS 2014 on natural language generation using deep learning. Also language modeling is often an important component in natural language generation, and neural networks have been very successful at this task.

It’s almost inevitable that deep learning will take over most components of my proposed system at some point. But they are not essential, at least initially.

But still, I should probably try and answer the question of why not build a single big deep net to rule them all? One answer is that we don’t know how to do this yet. But even if it were possible, I do not think I would want to try and do this. The answer is engineering. When I know how each component is supposed to work, I can fix it. When a deep net doesn’t work, all I can do is add more data and tweak the algorithm, which may or may not solve the problem (and may introduce new ones).

The argument I’m trying to make here is that natural language interfaces should be a solved problem, given that we have such sophisticated components around now, and all it requires is putting them together in the right way and engineering the thing correctly. Of course, that’s still a huge challenge, but one I’m quite excited about undertaking. I like big challenges.

Practical considerations

Now I can hear you thinking “It’s all very well taking on such a grand challenge, but who’s going to pay for it?” One option would be to try and build this thing in academia, after all, I’ve done the academic thing, so it should be possible to follow that route. The problem is, speaking with my metaphorical mortarboard on, in my experience, and speaking for myself, us academics tend not to build useful things things. You see, our motivation is naturally skewed towards publishing papers, which is what academics do, rather than building something that people actually want. And if we can squeeze a few percentage points of improvement out of a problem, we can publish a paper.

So if not academia, then what? I happen to find myself in the lucky situation of having some spare time at the moment. My current contract requires me to work only 15 hours a week, so that leaves plenty of spare time. I could try and build this just for fun, as a side project. However, at some point, my spare time will run out and I suspect this is going to take a lot longer than the three months I have left on my contract. The next obvious option is to try and build a company from it. This would either be a traditional startup, with funding and all the craziness that goes along with it, or a bootstrapped company. Actually a startup would not be a bad vehicle for something as ambitious as this. However, there are at least two reasons I don’t want to go down the traditional startup route:

  • I’m not convinced that this can be a billion dollar business (yet). The thing is, people want chatbots, but they don’t know they want what I’m building (although they may well need it). At some point in the future that may change.
  • I don’t personally enjoy the pressure to grow quickly that comes with a startup. I would rather build a successful and sustainable business slowly. That’s particularly true because I think it’s going to take a long time to build this properly. Also, I don’t hold with “stealth mode” - I’d rather do this out in the open.

So the current plan is for a simple bootstrapped company selling chatbots as a service. Yes, I know, there are a lot already, but I think there is space for another. The market is predicted to grow quickly in the next few years - we’ll see whether this turns out to be true or not. Of course I won’t be building my full crazy idea above straight away, I will only build each component properly as it is needed, and instead focus on building something that people want, preferable in focused niche.

One niche that I think is likely to be profitable is chatbots for marketing, specifically, Facebook Messenger bots as a landing point for Facebook ads. So, if you’re interested in this idea, please get in touch! I think it has a lot of potential for increasing the return on investment of Facebook ads.

If you liked this, please sign up to my blog Machine Learning in Practice - and thanks for reading!