Welcome to the final post in our dedicated Natural Language Processing Blog Series. In our first post, we asked Xcede’s Data team about the current state of the hiring market for NLP specialists. We then spoke to University of Edinburgh Speech Technology PhD student Jacob Webber to learn more about the uses of NLP within academia, current research in the field, and his connected field of study.
Here we speak to Srini Janarthanam, Lead Conversational AI Scientist at Builder.ai about responsible AI, AI bias, and how revolutions in ML and Deep Learning are advancing the NLP space.
Read his thoughts below.
Srini, can you describe what your role is at Builder.ai and the wider role of ML and NLP within the organisation?
My role at Builder.ai is Lead Conversational AI Scientist. I lead the NLP team working towards building a conversational AI agent that will assist customers and colleagues in engaging with the Builder platform and products in a seamless manner. Our team is part of a wider AI team called the Intelligent Systems team that is responsible for all the AI models that power the Builder platform like models that estimate time, cost and feature sets for customer projects.
What does ‘responsible AI’ mean to you?
Machine Learning models have been very successful in solving many NLP tasks. Researchers are coming up with more complex tasks to challenge them. Although the power of ML is growing exponentially, there are some problem areas too. ML models can be biased towards or against certain sections of the population. This is because they were trained on biased data. When using such models, we need to exercise caution. ML models based on certain kinds of architecture like neural nets can become opaque as decisions made by them could not be easily explained. How could you then explain to a customer why their application was rejected by the model?
Another important problem is privacy of data. Are we training our models on data that we shouldn’t be having access to? Models can learn to generalise but sometimes they memorise. Big transformer models have been shown to have this problem - memorising personal data that they were fed with during training time. How do we avoid these issues? How do we consume these models to build AI services? With great powers comes great responsibility. Responsible AI is about understanding the underlying problems, the quality of the datasets, limitations of our models and taking responsible action.
How, in your words, would you describe NLP, and has this changed in any major way over the past few years?
Natural Language Processing is a subfield of AI focused on getting computers to understand and generate natural human languages like English, French, Hindi, etc. Researchers in the field have built tools and techniques to analyse language at several levels - morphology (words), syntax (phrases), semantics (meaning) and pragmatics (context). These tools are used to do several NLP tasks like recognising and generating speech from text, translating text from one language to another, answering questions from a large body of text documents, etc.
How would you describe the evolution of NLP?
The earliest form of NLP involved building rule-based tools in which linguistic rules were hand-coded manually by language experts. Then came statistical models that mined the rules and patterns from language data. These were very robust and more flexible than rule-based approaches. Recently, in the past ten years, deep machine learning methods have overtaken the past approaches. We now have machine learning algorithms that can ingest large quantities of text and build huge state of the art models for addressing NLP tasks.
ML models like BERT and GPT-3 which were introduced recently are huge models with millions of trainable parameters that have learned to represent any sentence in an N dimensional semantic space. This means these models can take any sentence and, in a sense, ‘understand’ the meaning of it. So, they can compare two sentences and tell if they are similar to each other. GPT-3 can understand a prelude and generate a whole story that could follow the prelude meaningfully. Of course the results may need to be fine-tuned by humans but still I think the quality of their output has improved a lot.
What is the most exciting thing to have come up in the conversational technology space over the past year?
The most exciting thing in Conversational AI in recent years is the emergence of transformer-based architectures. Several transformer-based conversational AI systems were introduced by organisations like OpenAI, Google, Facebook and Amazon.
In early 2020, Google introduced Meena, a conversational AI open domain chatbot that can talk about anything. It was based on transformer architecture, much like BERT. It was trained on hundreds of gigabytes of social media conversation to come up with sensible and specific responses to the user. In 2021, they released Lamda, which is an upgrade to Meena. It focused on making conversations not just sensible and specific but also interesting and factual too.
Amazon have been running the Alexa prize challenge for a few years now. It is a global competition for university teams to create the most sociable chatbot that can interact with users on Alexa. This competition has generated many new architectures for open domain social bots. This year’s competition also includes a task bot challenge where the teams are required to build a chatbot that will assist users with DIY tasks like cooking a recipe.
Facebook came up with Blenderbot after Meena in which they focused on giving an open domain socialbot a consistent personality and empathy skills. Their study showed that 67% of human evaluators thought it sounded humanlike.
How much is sentiment and emotions a part of your work and where do you think this (area of the space) can go?
Sentiment and emotions are essential components of human conversations. Making conversational AI empathic is one of the goals of the field. There are two sides to this problem - recognising emotions and generating them. We want chatbots to understand the emotional state that the user is in so that it can adapt its behaviour. We also want chatbots to generate emotions and display them during conversation to make the conversation engaging. Some open domain chatbot researchers have started to focus on this problem.
In task based and goal-oriented chats, emotion recognition can be used to direct the conversation in optimal directions. If the customer seems to be angry or disappointed, chatbots could consider handing off to the human team quicker than usual. If they are confused with jargon, chatbots could generate a simpler version using descriptions for concepts.
I think we still have a long way to go to fully understand how to responsibly use emotions in conversational AI settings.
What’s next for the sector?
Conversational AI is a very exciting field. We are tackling one of the hardest challenges AI is facing - that is to build AI systems that can have engaging and useful conversations with us in natural language.
The research in the area has grown multifold in the last decade. New architectures are emerging every day. New datasets, models and code are shared under open-source licensing which I think is a great way to accelerate the field.
While open-domain chatbots like Meena, Lamda and BlenderBot have made big news, task based chatbots haven’t. Scaling up task based chatbots like the ones you would use in a sales, troubleshooting or customer support scenarios is still a challenge. Transformer based architectures for taskbots are emerging however they are not mainstream yet. Even the most basic conversation rules have to be hand-coded in current approaches.
The next big thing would be to build task-based chatbots powered by transformers in a meaningful way so that organisations need only provide content of conversations to chatbots (what to talk) and not worry about the rules of conversations (i.e. how to talk) because those are already learned.