“ChatGPT is the start of a new NLP era – but there is still a long way to go”

Enlarge

Barbara Plank is Chair for AI and Computational Linguistics — MaiNLP lab, and co-director of LMU’s Center for Information and Language Processing.

“ChatGPT is an example of an NLP language model (LM). Natural language processing (NLP) is a very active and interdisciplinary research field (also known as computational linguistics), with the aim to teach computers to understand human language automatically by applying artificial intelligence and machine learning.

Everyone uses NLP already - but may not be aware of that yet (think at spam filtering, machine translation etc). ChatGPT is about to change that – it is the beginning of a new era in NLP: models that can produce longer, more fluid texts, that interact with the human, producing texts that look remarkably coherent.

ChatGPT was released by OpenAI on 30 November 2022 — the very day before we held a workshop on Next Generation AI about “Writing with Artificial Intelligence” at the CAS LMU Center for Advanced Studies.

The LM can do many things, for example it can answer questions, generate texts, generate stories, and even write software code. Generating texts from data, such as weather or sports reports, is easy. This kind of data-to-text generation is widespread — Associated Press has been using it for about ten years.

It’s the faithful generation of text that stretches it to its limits. Say, if we ask it, “How do I get from Munich to Venice without crossing the Alps?” Then ChatGPT (in its version of January 30, 2023) answers as follows. Coherent at first glance, but without understanding the essence of the question.

Example 1: A question for ChatGPT

Barbara Plank: How do I get from Munich to Venice without crossing the Alps?

ChatGPT: 1. Drive or take a train from Munich to Venice. 2. From Verona, you take a train to Venice.You can also fly from Munich to Venice, which avoids having to cross the Alps as well.

Hours of human labor invested

Is ChatGPT new? No. ChatGPT was released recently, but a similar model and predecessor has already been around for a year now (InstructGPT, January 2022). A big difference between the two models is the public access: OpenAI made ChatGPT available to the general public on its platform. Anyone can now test the model.

What few know is that a lot of painstaking work went into it before releasing it to the public: Over the past year, the technology has been improved and made more secure by creating filters and preferences based on human labeling work. It has taken not only computing power but also manpower to bring the models up to present-day standards.

Compared to older chatbot models, new models like ChatGPT are characterized by their increasing coherence and linguistic fluency. The newer, bigger models are fed with vast amounts of text files from the Internet, and ChatGPT is subsequently refined with a great deal of additional data to rank human-preferred outputs higher and create filters to avoid amplifying bad language and biases, which takes hours of human labor.

Faithfulness remains a major problem, as does avoiding bias, supporting smaller languages and dialects, and also hallucination. Hallucination describes the phenomenon whereby an AI model generates an answer that does not make sense, is illogical, or is not relevant.

Another problem is that it’s also increasingly difficult to distinguish AI-generated text from human-generated text. More and more classifiers for this are becoming available (OpenAI has just published a model); ideas for detecting it and research on it have been around for years, but it is becoming increasingly difficult.

Man and machine need to work together

ChatGPT and similar language models are here to stay and they will partly shape our future. That calls for research and a lot of resources. I believe one particular challenge is in learning language variants with little digital data available, which includes Germanic dialects. That’s one of the things we are currently working on in my recently started ERC project “Natural Language Understanding for non-standard languages and dialects.”

ChatGPT and other chatbots will improve in the future, they’ll provide more faithful answers, probably also offer source information — which is missing from ChatGPT today. But to improve the model, I think that there needs to be a strong link with knowledge databases and that man and machine will need to work together, because without human-centric NLP, the machine will only be able to understand language and solve tasks to a certain extent.

Example 2: What ChatGPT thinks about its own future

Les us ask ChatGPT itself:

Barbara Plank: Where will chatbots go from here?

Chat GPT: It is difficult to predict where chatbots will go from here, but there are some trending areas where they are likely to evolve:

1. Artificial intelligence and machine learning: Chatbots will continue to improve to better mimic human communication and demonstrate understanding.

2. Natural language processing: Chatbots will get better at understanding and processing natural language, including the use of context and emotional signals.

3. Personalization and contextualization: Chatbots will be able to tailor their responses to the needs and context of an individual or a situation.

4. Multi-channel integration: Chatbots will continue to integrate with other channels such as voice assistants and virtual customer service agents to provide a better user experience.

5. Transition from rule-based systems to learning systems: Chatbots will transition from fixed rules to systems that are able to learn from data and adapt.

However, it’s important to note that the development of chatbots will also be influenced by other factors such as technological advances, regulatory considerations, and social trends.”

Was ist von ChatGPT zu halten? Einschätzungen von LMU-Forschenden