ChatGPT has become a very hot tool to use since it debuted late last year. While it’s not the first chatbot out there, there’s certainly been quite a bit of buzz around it. We’ve interacted with AI before when using Google Translate to communicate or try to get an issue resolved with Amazon before we get a real person on the line. There are a lot of chatbots out there, but what makes this one different? What could this mean for the future of Artificial Intelligence (AI)?
One of the first trends that has been improving quite a bit over the last decade is the quantity and quality of multilingual programs using Natural Language Processing (NLP). There is quite a bit of history behind that, which I have covered here. In short, machine translation goes back to rule-based machine translation programs in the 1960s and has improved over time as newer and different techniques have been introduced. Machine translation today currently makes use of neural networks, such as LSTM-RNNs (Long Short-Term Memory networks-Recurring Neural Network – also covered in my previous article). These neural networks have been key in making automated translation sound more human-like. While these translations can still sound or read a bit odd, it is still light years ahead of where we used to be. Take Google Translate for example, as when it debuted it was cool, but made so many translation errors. Its current state is much better than where it was during its debut, even if it may not be the best automated translator out there.
Machine learning. It is a hot topic and buzzword. We know that machines can learn, but what exactly under this broad umbrella term is trending? Reinforcement Learning. Reinforcement Learning (RL) in a nutshell is a branch of machine learning where an agent, the model, learns from its “experiences” in a dynamic environment. This method is different in that the model will receive “rewards” when it takes the correct action in response to a given input. You’ll have to define the reward as that will be a way to measure the performance of the model. The requirement for this though is that it requires many observations in order to learn what the correct response is before consistently giving the desired action (Mathworks). One of the benefits of reinforcement learning is the increase in response time from the model based on user input. However, depending on the training data, this can mean that the model can give a wrong answer fairly quickly. This has recently been seen with Google’s new AI chatbot BARD, where it gave an incorrect response about astronomy during its debut last month. Typically, models are trained on supervised algorithms and then fine-tuned with RL (MonkeyLearn 2020).
The last trend we’ll discuss is the increased use of automated chatbots. More and more services and companies are using chatbots to help direct consumers to the right channel for customer service. You have more than likely interacted with a chatbot when contacting Amazon for a return or called the bank to inquire about a transaction. Chatbots help lower the amount of support tickets while increasing the timeliness and effectiveness of resolved issues a consumer may have. We should see an improvement in timeliness and the complexity in tasks a chatbot can take care of before needing to speak to a human in the coming years.
CASE STUDY: ChatGPT
ChatGPT has become a popular chatbot for many people to interact with. Unlike the chatbots mentioned previously, the popularity of ChatGPT has been due to its ability to have full, mostly sensical conversations with the user. It is using OpenAI’s InstructGPT, which itself is based on GPT-3. The difference between InstructGPT and GPT-3 is that InstructGPT has been adapted to have the ability to follow user-inputted instructions. It is what allows ChatGPT to respond in the way you expect it to. ChatGPT used what OpenAI calls “Reinforcement Learning from Human Feedback”, or RLHF. It still retains the characteristics of RL, but the addition of human feedback and an increase the difficulty of tasks the system would learn to handle (OpenAI 2023). The below image describes in more detail the process they used to train ChatGPT.
There are different reasons that contribute to errors in responses from ChatGPT, mostly due to its training. It cannot search online, like a search engine, so it is limited to the data it was given and trained on, which is only as current as 2021. Any information given will only be as current as its data (OpenAI 2023). Despite this, this also means that responses may seem a bit weird to a non-English speaker as ChatGPT was trained on primarily English-based data, so any response templates will be biased towards English and (American) culture. A good example mentioned would be the standard five paragraph essay format that is used in US English (https://jilltxt.net/right-now-chatgpt-is-multilingual-but-monocultural-but-its-learning-your-values/).
Each day we become closer to what the future of AI may look like as new technologies are developed and improved upon. New developments in NLP have led to changes in how companies can improve their services and applications. Those that have been around for years have seen significant improvements with new techniques, such as neural networks. ChatGPT has shown a combined number of NLP improvements into one application. Many of these improvements have been seen in language translation, and as the world is so interconnected, it has become even more important that applications are able to handle tasks in multiple languages. Despite any errors it may give out, it is constantly improving itself as users interact with it, like other similar applications. Reinforcement Learning has helped applications improve their responses, so that rather than understanding the likelihood of certain words to appear next to each other, it understands what a typical response should look like. AI chatbots have improved customer service for many companies and decreased how much time we spend as consumers trying to receive assistance with businesses. While chatbots have typically been seen in the customer service industry, ChatGPT is showing how chatbots can be applied to other areas.
SOURCES
“9 Natural Language Processing (NLP) Trends in 2022.” MonkeyLearn Blog, 23 Dec. 2020, https://monkeylearn.com/blog/nlp-trends/.
OpenAI. “CHATGPT: Optimizing Language Models for Dialogue.” OpenAI, OpenAI, 2 Feb. 2023, https://openai.com/blog/chatgpt/.
Walker Rettberg, Jill. “CHATGPT Is Multilingual but Monocultural, and It's Learning Your Values.” Jill/Txt, 6 Dec. 2022, https://jilltxt.net/right-now-chatgpt-is-multilingual-but-monocultural-but-its-learning-your-values/.
“What Is Reinforcement Learning?” What Is Reinforcement Learning? - MATLAB & Simulink, https://www.mathworks.com/discovery/reinforcement-learning.html.
“Why Doesn't Chatgpt Know about X?” OpenAI Help Center, https://help.openai.com/en/articles/6827058-why-doesn-t-chatgpt-know-about-x.
Comments