Inaugural Lecture by Roman Klinger in Bamberg: Natural Language Processing Research in the Era of AI

So, sehr geehrte Damen und Herren, sehr geehrter Herr Präsident, lieber Kai, als Dekan der Fakultät Wirtschaftsinformatik und Angewandte Informatik freue ich mich ganz herzlich, Sie heute zu dieser festlichen Veranstaltung zu begrüßen. Wir sind ja als Fakultät in den letzten Jahren stark gewachsen, haben uns in wenigen Jahren verdoppelt. Ich war der 16., jetzt sind wir bei über 30, und das hat die freudige Konsequenz, dass wir bei Antrittsvorlesungen einen gewissen Stau haben. Das ist für Sie und uns heute Abend insofern angenehm, dass wir gleich 2 Antrittsvorlesungen mit einem Termin abwickeln können und Sie gleich 2 Vorträge von neuen Kollegen sehen können. Den Anfang wird Roman Klinger machen, den wir von der Universität Stuttgart zu uns holen konnten, wo er sehr lange am Institut für maschinelle Sprachverarbeitung tätig war. Guten Abend. Und Roman arbeitet interdisziplinär, bringt auch noch Startup-Gründererfahrung mit, das ist also genau das, was die Fakultät hier dringend brauchen kann. Und als ich ihn gefragt habe, ob er denn wohl auch an künstlich intelligenten Systemen arbeitet, hat er gesagt, so, nee, würde er eigentlich nicht, sondern er arbeitet nur an Systemen, die Sprache verstehen. Das fand ich total nett und bin sicher, er wird uns das heute noch ein bisschen genauer erklären, weswegen ich sagen würde, Roman, die Bühne gehört dir. Wir freuen uns auf deinen Vortrag. Vielen Dank.

Welcome everybody to this inaugural lecture. It's my pleasure to be here. Thank you for being here. Mr. President, Dean, thank you very much for coming to this lecture. When I heard about this event 1.5 years ago that I need to do this, I was a little bit intimidated. I found this a challenging event, so I looked up online what inaugural lectures are about. And I tried to find information on impressive, important inaugural lectures that have been given in the past, and I came across this person here. So this is Friedrich Schiller. He gave an inaugural lecture in Jena a while ago, 1789, and this was on the topic: Was heißt und zu welchem Ende studiert man Universalgeschichte? Okay, what is and to what end do we study universal history? Not less intimidating. I don't think I can really keep up with such a big topic, but I still found this, yeah, inspiring to think about how to make a good inaugural lecture based on such example. So what Friedrich Schiller did in his lecture is that he divided students into 2 groups. One group were bread scholars, those who essentially study to have practical advantage and advance their career, and others who are philosophical minds who seek knowledge and human betterment. And I found this an interesting hypothesis that there are these 2 groups of students. I'm not really sure if I can agree with that. But I find it interesting to ask ourselves, for whom do we teach and for whom do we do research? I'll come back to this. So now, this inspiration essentially also brought me into what I want to talk about today, and that is essentially what is and why do we study natural language processing? And we do this in the context of AI, which seems to change our field. These days. So, to be able to discuss this, we first need to know— when we want to understand what natural language understanding is, we need to discuss what language understanding is. So, let's say we are in this situation that there is one person and another person, and this one person A says something to person B, perhaps something about the world, perhaps they share some news item, Perhaps they share a social media post, and then the other person wants to understand that and does do something with that. And essentially, our research is about replacing the human, at least building helpful systems which are able to do the same thing. We want to have a machine that is also able to understand language, perhaps as humans do. So we want to study how machines can understand human language. And in our research, we mostly focus on written text, not so much on audio signals, but that's just to focus on something. Still, we did not really discuss yet what it means to understand. So let's look up what understanding means in a lexicon. They should— this should tell us something. So the Cambridge Dictionary says, Understanding means to know the meaning of something that someone says. I can't disagree, but it's also not too helpful for a computer scientist to build on top of this and build systems now. So what we would like to do is we would like to build computers which understand. We want to measure if we are successful, and we want to make use of the meaning that we understood with these computers. So, the goal is essentially that we build a machine that understands language as humans do, but that's still not really concrete. We still don't really know how to start this, how to practically do this. And this is just a very big task, which might confuse this tiny robot a little bit too much. So, we somehow need to operationalize this task. We can't really study language in its entirety. It's just too big of a task. We can't really aim at studying universal language understanding abilities. So what we do instead is we typically define more narrow particular phenomena that we then can actually study. So we define concrete tasks that we want to solve. And that's essentially a pragmatic approach to natural language understanding. So it's always good to look at examples, and since Donald Trump is president again, we have tons of examples that are interesting to look at on the media, and this is about Donald Trump who mocks Bill Gates after billionaire's humiliating backflip on climate change. So, if you believe that Sky News from Australia has something to say that you care about, you might want to understand what is inside of this headline. Donald Trump mocks Bill Gates after billionaire's humiliating backflip on climate change. So what tasks, language understanding tasks, are in there? What's worth understanding? So, one comparably straightforward task, but still not necessarily simple to solve, is that we want to find entities in there, names. So, for instance, Donald Trump and Bill Gates, these are both mentioned there. And if we can solve this automatically, we could do this with large amounts of data and analyze which entities are often mentioned together. In which relation, et cetera. We could recognize the sentiment. This text is very clearly negative. So, we could also link this to the entities and recognize which entities in large amounts of data are perhaps more frequently mentioned in a positive or in a negative connotation. There's a topic in there, perhaps climate change, and there's also something about stances in there. This is a description that Bill Gates changed his opinion on climate change. There's also a relation. Donald Trump is described here to have a negative opinion about Bill Gates. That's in the word mocks. So if we do this and if we aggregate such information, large amounts of data, this enables a lot of use cases. For instance, we could find texts in which this stance is differently described, and we could recommend news to people which have the one description and the other to diversify recommendations. Recommendation systems. We could also do social network analysis, opinion mining, and such downstream tasks. So, how can we build such systems? Typically, we start with a researcher, this guy, with formalization of the task. So, the person writes down the guidelines, what the task actually looks like. We have guidelines. The next step is typically that typically another person or students annotate text which follow these guidelines. So they write, hey, there's an entity, here's the negative sentiment on top of documents in a computerized way. Once we have these annotated data, we typically have the same person or another person who makes use of these annotated texts to optimize automatic systems which can then, based on these data, automatically replicate the task and take over the task before humans needed to do. So then, if we have this machine, the machine could just take non-annotated text and do automatic annotations for these new texts. So now, this is something that some people would call an AI, but it's this very specialized AI. Now we have more general systems that are referred to as AI. And the question is if this more novel development perhaps helps us with these things. So, can an AI, another AI, that's why it's green, can an AI help us with the formalization of the task? Well, perhaps, but do we want this? Probably not. The human should specify what they— what the task is that they want to solve. For text annotation, maybe, maybe an AI can take the guidelines and just generate annotations on text. And perhaps an AI can even help with building computational models, sometimes called AutoML or something like this. But let's focus on this part. Can an AI just take guidelines and annotate texts? So, does this work? Well, if it works, if an AI could just solve a task by being instructed, for instance, with guidelines, we wouldn't need manually annotated data, and we would not need to spend a lot of effort in model development. So if this is the case, perhaps we don't need a researcher in natural language processing. Yeah, this guy is sad now. So is this the case? Well, this is a reasonable question. We should reflect on our role. We should see if we do something meaningful, if we do something helpful. So, a lot of research goes into formalizing concepts, collecting and annotating corpora, advancing machine learning methods, evaluating methods. If an AI can essentially just do this, our work is done. We should ask ourselves if this is the case and how we should change our research field. So, to be able to answer this, we need to ask ourselves, what is AI? Now we have an idea what understanding language means. Now, what is AI? So, that's also a difficult question. And so this term typically encompasses many different topics. Many researchers and topics agree that one important aspect is that a system behaves in a way that appears to be intelligent. Pragmatic decision. If it looks intelligent, it's intelligent. Methods include a lot: logical reasoning, machine learning, Learning semantic language representations, search or planning, integration of world knowledge. And these are just some. So what I want to point out is that artificial intelligence is not just one method or one technology. It's a set of methods and technologies. And my personal opinion is that this is why we should be specific with terminology. So in the moment when somebody talks about AI, there's quite a high chance that people actually mean language models. This is perhaps the predominant meaning. Of AI these days. And language models essentially represent probabilities of combinations of words, of language. How likely is that a particular word is actually seen in the real world? So they can essentially help us to solve this task. Which word fits here? I came here by bike. Food, car, train, helicopter. So a lot of different things which essentially are all means of transportation. You could also say something else that doesn't fit into this category of transportation, but a lot of these things share semantic meaning. We could also fill this gap here. The sentence, it's my birthday party, expresses Joy. Many people like their birthday parties and enjoy this. So we could also ask the model to analyze a text that is embedded in some other text. What goes here? You need to go through the— that's unclear. We need more information to reach Untere Sandstraße from the island at Bamberg. That could be City Hall, Rathaus. So these are different tasks which essentially are all about filling a gap with a language model. So language models like GPT and all these things, they represent probabilities of words in context, and that's all they do. And language models can be used for different types of tasks. So the first one here is perhaps just sentence completion. This one is text classification. We ask the model to categorize the text, which is in quotes. This one is perhaps even knowledge retrieval. We even get a piece of knowledge, a piece of information that we needed to solve a task like navigation, for instance. So how do these models do that? How does inference work in modern machine learning language models? So let's ask the question to this language model. Is an inaugural lecture interesting? And the language model should now Generate words which are likely to follow based on this question. It's a question, so the model could say yes. Unfortunately, it could also say no. Now, if it says yes, it could follow— could continue with if or because. Same for no. So you can see where this is going, right? So it's expanding different possibilities, and these are already a lot. So perhaps we should limit these different paths to only those which are most likely. So we asked this language model, which of these paths are most likely? And perhaps we just keep 2, which is a parameter. But then we continue. Yes, if the— no, if the— yes, if the speaker— no, if— yes, if the audience— no, if the prof— no, if the topic. So these are all possibilities for an answer. Again, perhaps too many. Let's drop those which are not very likely. And then perhaps we come up with yes, if the speaker expires, no, if the prof fails. And obviously, there's only one possible answer here, which is yes, which is hopefully the most— more likely output following what the language model estimates. So this decision— this is not a trivial process. The yes can perhaps just be output based on this question mark. The, the perhaps just follows because of the if, but for speaker and inspires, it needs to relate these words to Inigo or lecture or interesting. So there can be quite long distant relations between these words. So it's a complex task. So, how can this model do that? It needs to learn how to factorize probabilities or how to estimate probabilities. So these language models typically have billions of parameters. I didn't want to make billion dots, but I mean, these dots should mimic the parameters in there. And they see huge amounts of text to find values for them. And then typically you optimize these language models in multiple steps. The 1st step is a language modeling objective. So you find these parameters that I came by bike or plane or train is more likely than Potato. Not sure what that means. It's not likely that somebody would say that. Yeah. So the model should optimize the parameters such that real answers that are seen in text are more likely. The 2nd step is typically that you make the model be more likely to be able to answer in a way of answering instructions, following instructions. So, for instance, I say, list all German chancellors, and then we want to have an answer like a list. So we use dedicated data which follows this instruction-answer structure to find good parameters. And then what's typically done is a so-called alignment task where we ask the model to generate the most likely answer and the 2nd most likely answer, and we ask humans Which of the 2 do you like more? And then we adapt these models a bit more to what humans actually prefer. And perhaps people prefer that the answer starts with, let's delve into that. And that's essentially an effect that we've seen with GPT, that it prefers this otherwise comparably uncommon word delve more often than we typically do. So, and finding these best outputs, this requires logical reasoning, semantic understanding, efficient search, and world knowledge, which comes from the data. So you could say, you could claim that these language models essentially integrate many other approaches from AI that we've seen already before. So now, what is our research about? Well, what can these models do and what can't they do? They are so complex, it's difficult to say. What can they actually do? So we want to find out what they can do. We need data for that to evaluate them. We want to challenge these models, and we want to challenge them with tasks that we assume are difficult to solve. So perhaps tasks would be also challenging for humans. And now I'll look into 3 cases where— which don't really work like expected. We do a lot of emotion detection, so emotions are also difficult for humans to understand and to interpret. So, this is an interesting task to work on. So, we could instruct an AI, you are an expert in emotion annotations. Which emotion does the author of the following text feel? Select one from sadness, joy, anger, fear, surprise, disgust. And the output for a text like, I needed to cry, is perhaps sadness. There are other emotions possible, but this might be the most, most probable output. And if you do this, you get a performance of an evaluation score between 0 and 1 of 60 to 70%, which is not much worse than what humans can do when they want to replicate the emotion of another person. So that's okay. I wouldn't say it's solved, but it works quite okay. Yeah. Then we moved to arguments. We wanted to understand how people develop emotions based on arguments in a debate. So we asked the language model, you are an expert in emotion annotation and arguments. Which emotion does the reader of the following argument likely feel? We should ban plastic bottles because they harm the environment. Difficult to decide, right? I mean, personally, I think this is a great argument. I agree with this, makes me happy if somebody brings it up. The plastic bottle industry worker will perhaps be less happy about it. So without context, we can't really make a decision here. That's quite difficult. Interestingly, when you ask large language models to solve this task, they have a tendency to say this is about fear and anger. Difficult to say why, probably because arguments often just cause that. If they don't understand the argument itself, they just say this is about fear or anger. So, how can we do this? How can we conduct contextualized emotion detection and what we also do, argument convincingness assessment? So, we somehow need to know And this is a typical example for what research in our field looks like in our group, not generally in our group looks like. We need to know for whom we make predictions, and this is difficult. This is a challenge because we need annotated data that we can show the model to learn from, which is— which also includes information about the person who reads this argument, who sees this argument. But typically you don't have this information, right? If I bring up an argument to you and I don't know anything about you, I can't really make this assessment. So we somehow need to get such data. So what we did, and we means mostly Lynn and Savina, is that we asked people to role-play a debate and annotate arguments. So we told people, hey, you are in a town hall meeting, and on the stage is a person who brings up an argument. This person looks like this and brings it up in this and this way. And then we asked them about how they perceived this argument. So this person says we should ban plastic bottles. And then a person from the audience in this role play says, this argument is in line with my goals. I'm familiar with it. It makes me happy. I find it convincing. They don't really generate text. We asked them to fill a questionnaire to be able to use this data in a structured form. In addition, we get information about the person demographics. Opinions about the topic, et cetera. So this is already a challenging task, and perhaps it's something that you didn't expect from the field of computer science. We need to get data to understand what the computer should do at the end. It's perhaps a bit closer to a psychological study. We need to acquire information about people who are in interaction in a debate. And we are currently working on integrating these methods with computational methods such that we can personalize such estimates for people about whom we then hopefully know something. But sometimes we don't know, right? I mean, in most cases, we don't know about who's the receiver of a text. And in this case, the question is for whom we make predictions with the model. So, this is work where I only listed Johannes here as a first author, the lead author, but essentially, this was an effort in the whole group. We wanted to understand for whom these models make the best predictions. If we don't know who to expect to use our system, for whom do these models work? We did this in the context of offensive language offensiveness detection. So we wanted to decide if a text is perceived as being offensive or not. And we want to understand for whom the predictions of a model who doesn't know who the audience is, for whom does this model work best? So let's say we have text like, an American building would collapse before fire spread to more than a few floors. Now, how you perceive this really depends on who you are. If you are an American worker, you might say, well, this is offensive. Our buildings are just fine. Yeah, this person might say this is an offense. If you have, let's say, a British scholar, they might say, well, this is a nice joke. I know it's not true, but it's funny. So, so it really depends on who you are, how you interpret the text. What we did is we asked people to disclose their annotation, their offensiveness annotation. We retrieved This is actually from the original data, right? No, did we do this? Is it from the original data? Okay, so we took the annotations from the original data. This is a corpus which already existed where we got this information, who made this particular annotation, but then we asked language models for their prediction, and we compared with which demographics these predictions work best. And not too surprising, the models work best if you're white, comparably young, and male. And if you are in a different group of the society, the model might not do what you expect them to do. That's not too surprising because this demographic group is probably dominating the data that you find online. And therefore, the model is somehow adapting well to such data. So we need to understand that. That's the first thing. And secondly, we need to make the models work well for everybody. However, models anyway vary in the output. They do because there's randomness involved, that the output is not always, always the same. But that's also what humans do. So perhaps there's something good in there. Let's look at an example. Let's again prompt a model for emotion annotation. You are an expert in emotion annotation. The label set is— we'll come to this— the instance to classify is the dog ran towards me. So now I put this into a large language model as a prompt, and this LS I replace by anger, fear, joy, disgust, sadness, and surprise. But I could make an experiment And instead put this label set there. Same labels, different order. I could use the original order and introduce typos. When I do this 1,000 times for each of the 3 variants, I get this distribution for this first set. Yeah, there is some randomness involved, but there is clearly dominating surprise and a bit of disgust. This is reasonable or not, could be questionable, but this is what the language model says. Now, I change the order of the labels and I get this output. This is a known problem. Yeah, it's not something that we discovered, but it is a problem. When I told this in a dry run to my wife, Liebke, she said, yeah, this totally makes sense. Humans are also susceptible to these changes. Okay. But these typos, people can typically resolve, but they are not. I mean, they are, humans still are, but the language model is not. Again, a different output. Now, do we want that or not? So I asked my wife again because she had an opinion if this is expected or not. And she said, no, we don't want this. This is not a human. I want this system to be helpful. I want it to be reliable. So this is not really clear, right? I mean, we know that humans are differently susceptible to such prompt changes than machines. In some cases, the differences are bigger to humans than in others. And now we could ask ourselves, do we want models to output— the output to vary as humans do? If yes, how do we achieve this? We had Barbara Flang for an invited speaker. She cares particularly about this. If no, the question is, what should we do? What do we expect from models? Just this one answer for the dominating group in our society? Perhaps not. Yeah. So there are many different questions where the models do not behave as expected. Yeah, it looks like they do something meaningful, but they do not always do that. I heard we have another lecture today, so I'll need to wrap up. So what is and why do we study natural language processing, and how does this field change? Well, natural language processing is a research area that aims at understanding how to capture language meaning and structure computationally. Automatically understanding language helps to distill information from large amounts of text. More than anybody would be able to read. But we've seen that models do not always behave as expected, particularly not for everybody for whom these models should be useful. And to understand what machines should do, we need to understand human language in context. And that's something that is just starting. This research area is just starting. We've also seen now that LLMs do not really just solve tasks, not all tasks off the shelf. You couldn't just not, just not take this model and run it and expect it to work. But they show promising results for many tasks. However, in the moment, still more traditional methods from AI often work better than large language models. Still, it's worth following this route and making sure that it overall becomes better. So now back to this person. So are we bread scholars now? We just care about career and so on. Well, perhaps yes. I mean, we do teach how to practically build systems that do understand language. We help people, we educate people such that they can find a job in industry and make use of this knowledge. And build helpful systems and earn money and all these things. But at the same time, to advance the field, we also need to have a deeper understanding of what humans expect, how they behave, their own properties, demographics that influence language interpretation and production. So perhaps no. So can we be philosophical minds? And I would like to claim that We need to be, otherwise it might be a bit risky in the future to use these models if we don't really understand whose output they are generating, for whom they are helpful, what biases they have, if they are reliable, if they are robust. So these are all questions that we need to follow further, that we need to investigate further. And for that, we need to work in an interdisciplinary way. And many questions are not core computer science questions, but they go beyond. That's also why this is what our group looks like. We build a lot on machine learning, computer science research, but we also care about linguistics and language and understanding these properties. We care about social sciences, psychology, and digital humanities, and we try to integrate these different fields. With we, of course, I don't mean me. I mostly mean the group with whom I'm working. I'm really blessed with this nice group of people who are diverse in their backgrounds and have different opinions and bring up these opinions in group meetings. And we have nice debates that, that I believe advance really the research field and in the research questions that we work on. I heard that it's not expected to ask you now for questions. So this seems not to be a thing in this event. But of course, I'm really happy that you are here, and I'm more than happy to have further discussions with you during the session, whatever this is called, after the next lecture. Thank you very much.