157 - Diyi Yang: Socially Aware Large Language Models

0:00Su:Welcome back to the Stanford Psychology Podcast. I'm Sudha Karaja, a pre-doctoral fellow at Stanford University. I'm excited to be joining the show as one of your co-hosts. For my first episode today, I'm so happy to share my conversation with Professor Diyi Yang, whose research I've long admired and who has been an incredible role model since I took her human-centric large language models class. She is an assistant professor in the computer science department at Stanford University, affiliated with the Stanford Natural Language Processing Group, Stanford Human Computer Interaction Group, Stanford AI Lab, and Stanford Human-Centered Artificial Intelligence Center.

0:37Su:She is also leading the Social and Language Technologies Lab, where they study socially-aware natural language processing. Her research goal is to better understand human communication in social context and build socially-aware language technologies via methods of NLP, deep learning, and machine learning, as well as theories and social sciences and linguistics to support human-human and human-computer interaction. In today's episode, we discuss her interdisciplinary approach to research, along with her recent paper, Social Skill Training with Large Language Models, which introduces a new framework that supports making social skills trainings more available, accessible, and inviting.

1:18Su:So, without further ado, here's our conversation.

1:43Su:Thank you so much for joining me. I'm so excited to be hosting you today. Let's start off by getting to know your research a little better. You are in the Social and Language Technologies Lab, and your research revolves around human-centric and socially aware natural language processing. In your own words, how would you define socially aware natural language processing, and why does studying it matter?

2:09Diyi Yang:So, great question, because I also think about this question all the time. We call the research we are working on socially aware language technology. Basically, it's about the study and the development of language technologies from a social perspective. The goal here is that we want to enable today's AI systems to better understand and respond to social signals expressed in language and also the broader physical and social environment. So you can imagine that the socially aware systems can recognize social aspects such as social factors, cultural, emotion, perspectives, etc. And more importantly, they can help us produce or process implications and meanings behind the language in the same way like humans do.

3:01Diyi Yang:There are three dimensions I often use when I think about socially aware language technologies. So the first one is what we call social factors. If you think about the language communication between people for a specific message, it's not only just about the content. It's also about who is the speaker, who is the receiver, what's their social relation, in what context, guided by what type of social norms, culture, and ideology, and for what type of community goals. So taking all of it together, we will have a much better understanding of a sentence or a message. And when it comes to social interactions, to me, it's more about the activities and the interactions humans have with NLP systems.

3:48Diyi Yang:So this could include some of the broader organizational or cultural norms that have governed the interpersonal communication. The last dimension is what I call implication or social implication. It refers to the broader impact of the NLP system on society, including understanding both the positive and negative effects.

4:10Su:How did you get started with this? Because there is some CS, there is some social sciences. How did you become interested in what you research now? And what would you say were some of the major influences and experiences along the way?

4:27Diyi Yang:Yeah, great question. I think I first got attracted to this social dimension as I moved to the U.S. to start my graduate school. Before that, I did a lot of training in computer science and mission learning. When I moved to the U.S., I got very fascinated by different types of languages people speak around me. There are people from different countries. We all gather together to do graduate school together. You can see that there is a very strong difference when it comes to the language people use. The same expression might be perceived differently and a lot of cultural factors there. That's the first time I really get excited by all these kind of factors.

5:10Diyi Yang:And I realized that language or language technology is not only just about words, grammar and the sentences. It's also about people and what people mean when we use language to accomplish different goals, such as reading news or stay connected with friends and families.

5:30Su:In one of your lectures in the human-centric large language models class, something that stuck with me was that you kept emphasizing the data comes from humans, the model is produced by humans, the choices to build architecture are made by humans, and even the end product is for humans. Your work and you emphasize the human element at every stage of AI development or just any technological development. I'm just curious, what could go wrong if we don't pay attention to these human biases infused into these systems? And what can we do about it? Or just rephrasing it, How does your research play into it?

6:16Su:Because we talked about the social aspect, and I can't imagine a social aspect without human.

6:22Diyi Yang:Yeah, I think this is also why I wanted to create the course on human-centered large English models at Stanford. If we take a step back, a lot of the topics here is not only about the technical solutions. when I shared with you all data matters and data is produced by people and in the product and the system is still made for humans I think one side we really want to make sure the problems we are going to work on they matter in diverse real world domains when we think about finance education or even in health care I think taking real world impact into the first place to think about what the problems to work on is quite important.

7:06Diyi Yang:And going back to the data dimension, we briefly mentioned culture, language, etc. But if you think about that, not only just here are the tokens, words, here are the images, audios, it's also about human preference, our cultures, biases, a lot of signals that are embedded in data. So this is when the human factor or human-centered flavor or perspectives will help us understand our data better. And this is not only about data. If you think about training models, we have many different types of model architectures, et cetera. I think the researchers and the practitioners play a very big role to decide how models trend, what type of data we use, and how could we make trade-offs between accuracy, efficiency, et cetera.

7:57Diyi Yang:And more importantly, around risk and all sorts of safeguard we need to think about. I think that's a very important space for many of us to think about. And if we do not pay attention to those human-centric aspects, I think the consequences could be huge and really consequential in many ways. For example, our systems may not be aligned with human behavior. And in addition to the system themselves, When it comes to people, you can imagine a lot of issues around the trust or psychological impact of how AI systems may influence our behavior. All of those are very important today. And we really need to think about them at different layers so that we can avoid and prevent failures in communication, hallucination and other aspects.

8:48Su:I think it's a really good time to go into your paper because as someone working at the intersection of computer science and human behavior myself, I found your paper, Social Skill Training with Large Language Models, very interesting, where you introduced a large language model driven framework called AI Partner, AI Mentor that makes social skills training more accessible. What made you recognize this as a critical gap that AI could help address?

9:23Diyi Yang:Yes, this is actually one of my favorite work. I got excited about this topic for several reasons. First is that I, myself, I'm not good at many communication skills. Part of the reason could be English is not my native language. In the U.S., trying to navigate the context here was quite challenging when I first got here. So I always want to learn all sorts of social skills. And when I talk to my friends and students, I'll realize that learning these kinds of social skills is often out of reach for most of us. It's very time-consuming. Let's say if you say, oh, I want to learn coffee resolution.

10:02Diyi Yang:It's very time-consuming to do it. It's also very expensive if you are going to get a coach. Also, such coach and their availabilities might be limited. More importantly, I think many of those social skills, practicing them is actually psychologically unsafe. So if you think about, oh, I want to negotiate or do this kind of resolution about this topic with my roommate, with my boss, I think a lot of times people don't feel like they can easily open up. So I think a part of the reason is that I personally really want to learn that. The second thing is that I see a lot of the ways of how we build AI and evaluate AI today, or large English model today, is more about can those models do well on math, do well on coding?

10:49Diyi Yang:And I really want to see can they really also do well with social skills. When large English models get popular and the performance gets very impressive, I realized if we use them well, it's a great way to help enable very interactive training. Because one of the big advantages of large-time model is that you can have conversations with these AI systems, and you can do a lot of chit-chat. We can even role-play different characters there with large-time models. So I think it really provides an interactive and personalized training paradigm. that. To me, I think that this would help transform learning in many, many unthinkable ways.

11:30Diyi Yang:So this is like how this lab research got started. And I think you mentioned our framework. It's called AI Partner and AI Mentor. I use it as more like a conceptual framework to think about how we use large-time models to help people learn social skills. So think about you want to learn conflict resolution. Then you can actually talk to an AI partner. This AI partner might be a role play of your roommate. And then you practice different types of topics with this AI partner. And then the entire conversation is going to be coached by this AI mentor. So here, both AI partner and AI mentor are what we call large language model agents.

12:13Diyi Yang:One is that you can practice different type of conversations. An AI mentor is tailored in a way that we will bring in or build domain expertise into specific models so that those AI mentors can actually give you very realistic feedback. So this is like how I would describe this framework. And although it only has one AI partner, one AI mentor in the framework, you can actually use it in different ways. So think about if you want to practice how to talk to a group of people, then you can have multiple AI partners there. You can even have different AI mentors. Sometimes you may prefer, oh, what if my mentor has this type of personality?

12:55Diyi Yang:Or what if my mentor has this domain expertise? You can actually use it in very different ways.

13:02Su:So when a user wants to develop a new social skill, like talking to a roommate, it could be, or just having this interview, the AI partner, that agent guides them through a realistic scenario with a simulated conversation. And then we have the AI mentor, which offers insightful knowledge-based feedback at key moments during the simulation. So we have this conversation partner and we have this instructor that is like overlooking the whole conversation. As you mentioned before, the communication isn't just about syntax or meaning or just like the language. It's also about intentions, emotions, and social cues.

13:44Su:And you said that apart from all the other technical skills of large language models, you're excited to see these things. How does the AI partner, AI mentor, APAM system, take these deeper layers of communication into account? How do models understand these subtle social contexts that humans navigate almost instinctively?

14:07Diyi Yang:Yeah, that's a great question. So what I can share is that there are different layers when we think about the development of AI partner and AI mentor in practice. So, so far we have used it for conflict resolution where we build a system called rehearsal to help people practice how to have a difficult conversation. And we also have a domain where we use AI partner and AI mentor to help nervous counselors to learn how to be a good listener. Throughout those two different domains, first challenge, even before we get to this nuances or social awareness, is actually to make sure that technology would work well in the first place.

14:52Diyi Yang:Despite the fact that large language models are very powerful in doing role play, etc., etc. Their simulation, like if you are going to use LMS to simulate your roommate or simulate a typical learning partner, that process may not be very realistic because models tend to create a and they tend to amplify a lot of the attributes. If you tell them, okay, here is the specific person. So it's actually quite difficult to get the simulation to be realistic. And then if you bring in or think about the earlier topic we had about more human-centric perspectives, we actually first work with domain users.

15:37Diyi Yang:So we talk to senior supervisors, we talk to novice counselors, we try to work with domain users to create some prototype, because it's really difficult to simulate a specific individual. Instead of doing the simulation of a specific individual, we try to create a typical prototype for interacting on that specific skill. So here, we are not talking about a specific individual. We are talking about a type of individuals who you can actually practice with. By working with domain users, we actually bring in a lot of the nuances. Like sometimes people may not be easy to open up. So that's a feature we will build into the simulation.

16:22Diyi Yang:We'll get many, many of those. And then when it comes to the implementation stage, we actually develop some techniques such as self-critique and self-improvement. So the key idea here is when the simulated AI partner is in the process or is trying to output a sentence, We will try to let the model themselves self-critique themselves. Like, is this a reasonable response to use here? Does it align with all the domain knowledge? Is it appropriate, et cetera, et cetera? So many of the contextual assessments will be achieved at this stage. So this is how we actually make sure that the interactions could be realistic in the first place.

17:06I think the second dimension I would emphasize is this is not only just a process of building AI systems.

17:15Diyi Yang:We have greater theories when it comes to social skills from communication field, from psychology field, and also from psychotherapy field. So we actually use a lot of theories from those different research fields, try to see whether we could let them in the development process so that a lot of the feedback are actually very grounded to users. So I think that that's the two challenges and approach we are working on these days. In terms of a lot of the cultural dynamics, a lot of the social awareness, we haven't figured out a great way to integrate them into the space. I think that itself is an open question.

17:56And I wish that our AI partner, AI mentor framework could include more of those cultural related practice in the future. I'm curious about the after-production stage.

18:09Su:You mentioned that there's still work going on and it's not perfect, but it's well known that evaluating models for accuracy and consistency is key. And there are plenty of benchmarks that focus on more aesthetic, measurable aspects of model performance. But in the system you are talking about, these models are said to have, for example, different personalities, depending on the mode they're in and even give feedback on how empathetic or engaged the person practicing is. For example, in the case of a therapist you mentioned. These are skills that are pretty straightforward for humans to do. So how do you plan to evaluate these human qualities like having a personality, understanding emotions or showing empathy and expertise?

19:00Su:Are they measured using the same standards we would use for a person? And if so, how do you even quantify something like that? I'm just curious because you mentioned using social fields to build these models, but are we going to hold these models to human standards, to the standards of these social fields so that we are actually evaluating them as if they're an actual agent we are practicing with or that they're mentoring us?

19:29Diyi Yang:Yeah, very thought-provoking question. I don't think there is a single standard or like one face-all answer. I think it really depends on case by case. So maybe let me just use a few examples to go through this. If you think about learning to be a good supporter, so you want to provide a social support, and this is a skill that many of us want to learn, how to be more empathetic, etc. And in the practice stage, maybe you want to interact with different types of patients or clients with different personalities. So in that situation, you need to have the personality really embedded into the simulation.

20:12Diyi Yang:So then the evaluation would be as similar as what we do when it comes to evaluating human's personality. So this is like scenarios where you want to blend them in. And similarly, I think if you want to bring in a lot of other aspects, such as, oh, maybe this is the person who grew up in Brazil. Here is a person who has more knowledge about Canada, etc. I think a lot of the domain expertise could also be introduced in similar ways. Evaluation is really like it's not an easy task. I think for personality, because we have greater theories and frameworks, you can easily quantify it. But a lot of the social aspects are actually very fuzzy and not well defined.

21:03Diyi Yang:We don't even know how to operationalize them. So the evaluation will become something that's quite tricky and open-ended. And when we go to a little bit of broader social construct, it's also a bigger spectrum. because if you think about how to evaluate whether the simulation is a really good simulation of a person from Brazil, this is a really hard question because there are all sorts of different individuals, etc. There may not be a ground truth answer that you can use compared to other situations. So for many of the evaluation here, we actually sometimes work with human experts. We ask, is this a simulation that looks or sounds like believable to you?

21:54Diyi Yang:Is this a good reflection of the client you had in your interaction? So we actually developed a lot of these more human studies or evaluations to evaluate such simulation. And I think a similar challenge exists when you want to simulate a culture. for example in the context of psychotherapy if you want to learn how to talk to patients from different cultures there isn't a single formula saying that okay this culture has this kind of prototype so that's what makes sense very very difficult today and also i want to just take a step back i i like your question i think one thing i haven't mentioned is that social skill training The way how we think about it today is this is a first stage training for people.

22:46Diyi Yang:It's not like, OK, as long as you finish this, you will be expert in the world. It's not like that. We imagine that especially for novels, for beginners, for people who don't know the field very well, they can actually use this kind of AI partner and AI mentor in the starting stage so that they get a lot of practice, a lot of experience. And later, a lot of the social factors that we briefly mentioned earlier, I think you really need human connections and have this kind of real sessions to get to know it. If you can imagine being a good supporter does not only mean you produce the best message there or the language there, it's also about how you behave, your emotion, your pose, like how you do eye contact, a lot of those.

23:35Diyi Yang:So I think it will be very, very cool to think about how to make social skill training in a physical world or a 3D space so that we can actually bring in a lot of those to help people learn in the earlier stage.

23:49Su:If you're also okay with it, I want to dive into something you just mentioned because I noticed that in the paper too, you mentioned multiple times that you don't expect the system to just replace all these trainers or replace all these social skills training you want this system to be a collaborator where you just introduce as a helper to the systems like the whole social skills training system can you please elaborate on that