With human trainers, bots may finally get smart enough to hold complex conversations
ParlAI, a Python-based tool for developing and evaluating chatbot dialog models, tries to teach bots by enlisting humans, via Amazon Mechanical Turk, to train their code-based replacements.
Facebook’s desire to turn mobile chat into an e-commerce command line hasn’t quite worked as well as planned. At its recent F8 developer conference, Messenger veep David Marcus acknowledged that Messenger bots work better when interaction comes in the form of menu selections instead of as written or spoken commands.
In an effort to downplay dialog-based interaction, he also said that Facebook never called conversational apps chatbots. Just bots.
FAIR boffins Jason Weston, Alexander Miller, and Will Feng say much the same thing, though they evidently missed the “don’t call them chatbots” memo.
“Existing chatbots can sometimes complete specific independent tasks, but have trouble understanding more than a single sentence or chaining subtasks together to complete a bigger task,” the three researchers explain in a blog post. “More complex dialog, such as booking a restaurant or chatting about sports or news, requires the ability to understand multiple sentences and then reason about those sentences to supply the next part of the conversation.”
ParlAI complements Facebook’s other text-oriented research projects, including FastText – a text-classification system – and CommAI, a framework for developing general-purpose software agents that rely on linguistic input.
The software is designed to allow researchers to submit multiple tasks and training algorithms into a shared repository, with an eye toward unifying the results into a coherent set of data that will allow bots to conduct complex conversations.
Tasks in this context refer to five types of dialog training:
- Question answering
- Sentence completion
- Goal-oriented dialog
- Chit-chat (social banter without a specific purpose)
- Dialog associated with objects (initially image-oriented)
ParlAI supports over 20 tasks, such as bAbI tasks, SimpleQuestions, and SQuAD, which are used for approximating the process of reasoning. For example, given the statements, “Daniel entered the kitchen,” “Mary took the milk from there,” and Mary went to the office,” a bot trained with bAbI might be able to conclude that the milk has been moved to the office.
In theory, chaining together various straightforward AI tasks can lead to complex AI, the sort of personal assistant technology depicted in science fiction.
Asked whether there are any examples of bots capable of this sort of complex conversation, Weston in an email to The Register said, “ParlAI has only just been released, so there aren’t too many trained models yet. We do include a dataset of this kind of task, which has nice results showing end-to-end neural networks outperforming traditional rule-based systems: Learning End-to-End Goal-Oriented Dialog.”
To refine bots further, ParlAI has been designed to collect feedback using Mechanical Turk, Amazon Web Services’ turnkey cheap labor service. Turkers, as Turk workers are known, may get presented with facts, about which they’re expected to pose and answer questions. This input gets used to build better bots.
“In the end, dialog with humans is necessary to build chatbots that can talk to humans,” the researchers observe.
Weston allows that chatbots leave something to be desired as conversational partners, but remains optimistic that they can be improved beyond menu-based interaction.
“Menus or apps are strictly limited to the options they include,” he said. “The full technology of chatbots isn’t there yet, but hopefully one day we’ll be able to serve all kinds of requests, ones that we couldn’t have predicted. We do believe we’ll get there eventually. ParlAI is one step towards that.” ®