 Alright, thank you Michael, thank you Alec, thank you to Joyce and Laura, and thank you to each and every one of you for being here today. I'm really very grateful for your presence. I know we've been and could continue to be through a difficult time, so having you here in person today means a lot to me. And especially I'm very grateful to you, Janice, for being here today. I'm inspired every day by everything you have done and by all your amazing contributions in research and contributing toward making computing a more diverse field for all your accomplishments for being such a strong advocate for women. And I'm really happy to have you and your family here today, so thank you for that. I was asked to talk today about my research and what matters to me, and I did spend some time thinking about what I should be talking about. Obviously, national language processing is my field of research. It's been for the past 20 plus years and I absolutely love it. I love languages, I love computing, I love everything about national language processing. That's really the thing that makes me excited in the morning thinking about all these interesting problems and the impact it can have. On the other side, I'm also really passionate about people. I really see how they make a difference every day, including you here today, my students, my collaborators, all the wonderful staff that put together this event here, of course, my family and friends. So it's really, they are not two separate things, but national language processing and people think they really belong together. So what I want to talk about today is about this interplay between national language processing and people, how national language processing can actually have impact on society, how people need to be at the center of the research that we do in this field for it to be equitable, and how it really matters who's doing this research. And so that's really the main message that I want to send across today, that research, particularly talking about national language processing, but other research too and people cannot really be detached from one another. So specifically looking at this field of national language processing, we've been working on this for 60-plus years and the past few years have seen a lot of advances. So now we know a lot of things. If we take one individual word, say the word son, we know how to create representations. You might have heard about all these advances in neural networks, all these GPT-2, GPT-3, trained on billions and billions of words that created fancy representations and eventually allow us to compute with words. The same would go for phrases. We take a phrase like with paper, we know how to computationally identify the relation between the words and know what's a verb, what's a noun, what's the relation between them, and on and on going to larger units of text, entire sentences, how to represent them, entire paragraphs and documents. So this is really what we generally do in natural language processing. We really focus on words. Now who's behind these words? They are people. So people are those that eventually create language. People are those who benefit from the technologies that we are producing in natural language processing and beyond. And people are really those who are working on their research. There is a very close connection going both ways, from language to people and from people to natural language processing. One at a time and illustrate with some of my thoughts or things that I've been trying to contribute to work. So first off, thinking about natural language processing and what is that these technologies are trying to achieve. There is a lot of very exciting research and a lot of basic research and we are all for that. Laura Gordy, for instance, did 30 HD around word representation, which is really a fundamental problem. But we are also motivated as she said, we are motivated by what is the societal impact of the word. So in my lab, a lot of the work that we are doing has this as a goal. We have that front and center thinking about what is the kind of societal impact that we have. So these are just some examples of where that we do. For instance, we work on understanding what are different, the differences between cultures in terms of behaviors, beliefs, value to a computational lens. How can we identify misinformation, which is something that we are running against every day, including now thinking about misinformation about COVID or vaccine and so forth. How can we really understand learners, people, students more than just by looking at their grades? What are the things that really make them think and there is a lot that we can do in language? How can we understand how mental health is on set and what are the things that we could do to... I realize when mental health issues may occur and what are the things that we could do to help. And it turns out that there is a lot that we see in language. There are a series of projects in my lab that accompany natural language with other modalities, for instance, vision or physiological processing. And there are a series of projects where we are really interested in understanding more about behavior and human behavior when it comes to deception, the example of the joy scape, when it comes to discomfort, stress, emotion and so on. We also care about applications that would have to do with detecting activities or being able to answer questions in real settings. Imagine for instance something like Alexa or Siri that would be able to tell you things that might be important to you, such as where did I leave my keys or did I take my medicine and so forth. So having that ability to actually provide support in daily life. Conversational technologies. Here a lot of the work that we do is in collaboration with researchers from Michigan Medicine, from School of Public Health, so they are around healthcare. It's really core natural language processing in terms of conversations, understanding conversations and facilitating conversations, whether it's about collecting medical family history, about providing counseling or support in settings such as the pandemic. So I want to zoom in on a couple of projects, which are some of the recent projects that we've been working on and illustrate some of the impact that the kind of research that our lab has been focusing on can have. All the things that happen around us, we see often time public health issues. The stress and anxiety that people have experienced over the past three years is just one. There is also substance abuse, where more than 20 million Americans, only focusing on the US, need some kind of substance abuse treatment. There are other health issues. There are people who want to stop smoking, losing weight and so forth. And for those health issues, behavioral interventions are considered to be very helpful. They can actually encourage people to pursue those health treatments. It turns out that although millions of people are in need of such behavioral interventions, there is a shortage of counselors. One, and second, the counselors that we have need to get to periodical training. They need to do a lot of self-reflection to realize what is that really works and what works for us. And so in collaborations with School of Public Health and Michigan Medicine for several years, we've been working on national language processing technologies for counseling conversations, looking primarily at two aspects. One aspect is to go deep in the conversations and understand what are the behaviors that counselors have. Are they asking a question? Are they making a reflection? If they are making a reflection, what kind of reflection is that something that's triggering the patient to speak more, to maybe decide that they want to change, which is really what these behavioral interventions aim to achieve. And so rather than having a human being to go through those conversations, which is really currently what's happening and giving feedback that way, we are able, through our technology, to provide a feedback that could happen on a daily basis. So you have a counseling intervention and a system based on national language processing could provide you with that kind of feedback. And I will tell you, you ask this many questions, you make this many reflections. Successful counselors for this particular behavior would typically do this and that. So really providing counselors with the kind of training that will get them to be better counselors. So this is not to replace counselors. This is really just to bring the wisdom of thousands of counselors and the fingertips of every single counselor. So they themselves can get that. The other angle that we looked at is how we can again assist counselors in what would work best to say. So again not replacing counselors, but really just helping them with language that might be useful for a certain setting. So for instance, assuming somebody who is trying to lose weight and they receive counseling, they may say something along the lines of, I had a really hard time sticking to my diet this week. And so the question is, what should a good counselor say? And what we've been doing in recent work is looking at how can we bring in medical knowledge? How can we bring in semantically related information? And like I said, the wisdom from other counselors. So what other counselors have said in similar situations so that we can actually generate a reflection, generate something that's empathetic and would possibly help these patients to stick to their life. So it might come something like, you are wondering whether you'll be able to adhere to your diet, which what this will do will encourage them to say more. Why is that an issue and so? And what we found is that this is comparable to what people are doing in terms of quality of the output and in terms of how empathetic this output is perceived, which is really what counselors are trying to do. Now building upon this work, when the pandemic hit, we realized again in collaboration with psychologists that stress and anxiety was one big thing. It hit every single one of us, just wondering what tomorrow would bring. And so we said that we could perhaps help people think through these issues and building upon theories and practical examples from psychology and also really simplifying the work that we've been doing that I talked about earlier. We created this online, you could call it chatbot. It's not really a chatbot. It's nothing like Siri or Alexa. You cannot really have open conversations. It's more like an interview. It's a system that tries to really get you to think through the issues that you might have. So it's really the system is not really helping you in any ways, just helping in the sense of getting people to think about what is this causing this stress? What is this causing anxiety? And just it's people themselves who help themselves because there is a system that just keeps some prompts at the right time. So that's all that is not necessarily super sophisticated technology, but it's knowing when to ask and what kind of questions. Similarly, when we've been having the Black Lives Matter, all the challenges that we've been facing, it turns out that people are sometimes not really open to talk about the issues they would have, whether they're open or not to talk with others, whether they're open or not to meet with people of a different racial identity. So we created a similar intervention which builds upon the kind of dialogue system that we have, but again simplifies the sense of it's system-driven. It's like an interview and it's really to get people to think about what is that going on? What are they thinking about the other race? Would they like to meet with others and really getting people to internalize this? And what we found is that this system, although it's again it's a simplified version of all the advanced technology we have, it can make a difference for people in terms of removing stress, which is the metric that we look for when we build a system around relieving stress during the COVID pandemic. And one user, it's of course a cherry-picked quote, but one user did indicate that I just completed the interview and I might say I really enjoyed and loved the analysis afterwards. It actually made me feel better about myself, which is really the goal that we have, is give prompt at the right time to get people to feel better about themselves. Here's another project that I want to briefly talk about, which is how to understand people's values across cultures. A lot of the work, including work that we do in my lab, focuses on English and a lot of the work focuses on just a handful of cultures. So a question that we ask again in collaboration with psychologists is whether we can use language to understand the values that are driving people. And so we ask people just to tell us about their values in national language and they came up with things such as just to be at peace with myself and others or being honest and straightforward is the most important value in life. Here we wanted to distill what are the set of values. We didn't want to work with thousands of these value expressions, which are all wonderful to read, but we cannot really process them and distill them. So we use topic modeling, which is a process that we often use in LNLP, and with that we're able to come up with a few values that seem to be what's driving people. Among others, people talk about respect for others, about religion, about family, hardware, time and money, problem growth and so on. Now the next question is, are these values really the same across cultures? So are the people from the United States and India driven in the same way? So for that what we did, we once again collected these statements from people from these two countries. And then we wanted to see to what extent the values that we identified would distribute across the two countries. Now we ran into a challenge here, because aside from the culture itself, so we knew the location, this is all data collected through crowdsourcing. We knew the location, but then there was also gender, which maybe wasn't uniformly distributed among the people who contributed data and age. So we used topic modeling coupled with some disinfanglement to figure out what are the different aspects that actually matter in these values. And it turned out that those values that we found previously, for instance respect for others or religion or family, they are distributed differently across cultures or genders. And this all is done computationally. So we start with the language and then we build bottom up using these technologies. So in this example just to illustrate, it's harder to read because you have to look left and right. So for instance for the red columns, left would be US, right would be India. So we see that people in US would really value hard work. So that's a top value for them. Whereas people in India would really value problem-solving. And that's just one example from things that we can uncover by analyzing them. So that's maybe a little closer to hope, which is these faculty reviews. So there is my professor, of course, we also have internal reviews. Right now professor happens to have data that is public. So there are all these comments that students would submit in response to the classes they take. And while there is some rubric that they would fill in with numbers, what we really care about was language. And also what we care about was to understand what is the students like in a professor with the hypothesis that maybe not every single student likes the same thing. Like say for instance computer science students may appreciate a certain characteristic of a faculty, maybe somebody in arts, something else, maybe somebody in psychology, something else and so forth. So we applied the same idea that I described earlier, where we start with really the student comments, we build topics, we disentangle them because we want to know what is the discipline that these students are from. We also threw in red to see if red matters for the university and country. We had these comments from Canada in the US. And all in all we had close to one million comments from students in this natural language form. So from here we came up with several traits, which at least students in the room will probably recognize. I guess faculty in the room as well. So students really talk about approachability. So that's something that matters. Clarity, course logistics, enthusiasm, what are expectations, how helpful faculty are, whether they are humorous, interesting, what are reading discussion and what is the study material. So these are sort of the main themes that students will talk about when talking about their professors. Now the question is, are all students talking the same? And what we found is that it's really not the case. So again a little harder to read, but I'll help you through. So green columns represent country. To the left we have US. To the right we have Canada. Then red would be the light green. I won't pay so much attention to that. We didn't really find any distinctions with respect to red. And then the other colors are disciplines. So for instance, when it comes to course logistics, American students really care about that. So they will often talk about course logistics. They want to know what's expected from them, where they are coming from, where they are going. So that's one thing. That will same go for reading discussions. And just to go through with the country, so the green one, students in Canada would really appreciate approachability and humor. So when they comment about their professors, that's one of the answers that you'll see among Canadian students. Now when it comes to other disciplines, we see for instance in biological science, that will be the light blue. So biological sciences care about clarity, but do not really care too much about humor. They also care about approachability. We see for instance, physics also care about clarity. And we see some others, if we go back here, we see for instance, physics or art, fine arts will not care too much about study material. And there are other things that will matter for arts today. So this is just one way to dive in data, language data, which we see around in so many different parts, whether it's student comments, conversations that counselors have, misinformation is out there in the form of news, and so much more. So natural language processing is really giving a way to process the data, and then in turn also have made a difference in society through the kind of technology that we can. When we talk about natural language processing, we assume that we can build one technology and that will do it for that particular problem that we work on. As it turns out, that's really not the case. And there was a conversation that I had relatively recently with a data science group in Detroit, and they ran into a problem. With the pandemic, there was a lot of homeschooling. Families were given devices, as it happened with my children as well, and then they were supposed to just use them. And as it turns out, a lot of these families didn't know a lot of things, like how you get this particular application on this device, what to do if you start connecting to the internet, how to take a test, and so on and so forth. So this center in Detroit was taking all these questions from families about technology, and it turned out a lot of them were repeating. And they could see you in using a question-answering system, so something that would have maybe a list of questions and answers, a little bit of intelligence on top and would be able to handle that, like getting a question, understanding what is the goal of that question, looking to all the information they had and provide an answer. And in question-answering, we claim, at least in natural language processing, that we made a lot of advances, and we are feeling fine. It didn't work for them. It didn't work for this large community in Detroit. And why it didn't work? It's because this technology was built for English. And in Detroit, most families were speaking, they would refer to it as the Detroit Slam. So it was the language they used on a daily basis. It's just that this technology completely failed to understand them. So it was just giving failure out of failure to the point that it wasn't really usable. So really what we are saying in research that we do in my lab is that one size does not fit all. When we build these technologies, it applies to natural language processing, but with other areas as well. So I want to step back a little bit. We really started from the basics. One of the basic units in natural language processing is word association. So with word association, I give you a prompt and you have to tell me something in later. So we'll do just a little exercise. So if I say cat, what is that you would say? First thing that comes to mind? No overthinking. Dog. So I see a lot of dog, which is really typical. I always see cat and dog. Now, how about if I say sleep? What is the first thing that comes to mind? Night. Away. I don't agree. But for others, we do not. And this is exactly what two psychologists about a hundred years ago did. They took a hundred prompts and asked people exactly what I asked here. What is the first word that comes to mind? And it turns out there is variation of this word association with age and also gender. So for instance, for sleep, the younger group dominant answer was dream. Whereas for a less young group, the dominant answer was awake. If we take food as a prompt, we'll see dominant answer for the young group eating less young drinking. So what we've done, we said, well, we want to see to what extent we can replicate this. And also to what extent we can build natural language processing representation that will be closer to really the group of people that has these responses. And so we took 300 words, starting with those hundred that had class 200 others. We did crowdsourcing. We asked people from India and US to tell us what is the first word that comes to mind. And we had some spend checking, questions, and also demographic questions. All in all, we collected more than 200 responses and which we then balanced across gender and culture. So here is an example of what is the first word that comes to mind. Female respondents from US, the dominant response was watered. When we look at male respondents from India, dominant response was watered. When we look at female respondents from India, dominant response was snow. Female respondents from US, dominant response was bubble. And there are a few other such examples. For instance, for inspect, dominant response for male respondents. Female, baby. Or if we look at location, I mean in India, dominant response would be hospital, in US it would be. There are differences between how people think about the world, how they perceive the world. And we did some analysis of all these responses that we collected. We found that indeed, people would agree more with other people from their same group. So if we take one respondent out and compare their response with people from the same group, whether it's the same gender or the same culture, we see that they agree more than would be with people from the other group. So that's confirming this hypothesis which was really formulated many years back in psychology, that there are differences between groups of people. And then the other thing that we wanted to do is to see to what extent we can build models that would more closely emulate what people are really thinking about the world. Again, going against this, what's the general strategy that one size fits all and really say no, one size does not fit all. We need different sizes for different groups. So we use a lot of data, primarily from loss for millions and millions of words. And then we train a neural network. So neural networks are currently used these days to create word representation. And we did a very similar model to what was used before with one exception. We now said, we know who said this word. We know that there was somebody from India or from US or somebody from a male or a female. And so we actually attached that label saying we know who's behind this word. The same for the word that we are targeting as well for the word in their context. And when training this network, we end up with is a vector like a representation for that word. But now instead of having just one representation that presumably would fit every single person out there in terms of how they see that particular word or how they interpret it, we have different vectors or different word representations for different groups. So with the data that we have, we compare how the generic model do, which again, it's fairly sophisticated. It's based on this neural network but it turns out that you actually account like who's behind the language you can do better. And that we saw that for gender and for culture and in more recent work, we saw that if we increase the amount of data that we have, we can actually do even better. So it's a lot of improvement just coming from this fact that we account for who's behind the language. So the main message here really is that data degradation is critical for equitable models. We cannot assume that that question answering system out there would solve everyone's problem. It turns out people from Detroit cannot use that and it turns out we will need to think of how to create a model or a system that actually works for that particular community and the same goes across many other communities. Let's get to the last part of my talk. I try to illustrate on why I believe it's very important who is that we are target or addressing world, who is that we are impacting to our research. So actually working on research that has positive impacts on society and also putting people at the center of that research do not assuming, not assuming that we can build the same technology for everyone. It turns out if we do that, we will really benefit the majority and we'll not benefit the majority. Now the other final argument that I want to make here is that it really matters who's doing that research. So that's the vital part. And these are efforts that I've been having really the pleasure and joy of being a part of and were mentioned previously in the generous introduction that Michael made and Joyce and Laura made in their comments. With this belief that it really matters who's doing the research in my role of president of the association for computational linguistics, aside from starting a new position for diversity and equity and a new committee on ethics for NLP, I've also been working on this ACL year round mentorship together with people also in this room, Ashkan Kazemi, one of my students, Xi Jinping, again another student and also other collaborators from other institutions. We opened this mentorship which has received huge response. We have close to 1,000 respondents, students from around the world who sign up to receive this mentorship and this number is really telling us how needed this kind of mentorship is. Yes, we all observe here, being where we are, is that really in the United States we are very privileged. We have a lot of resources. We have a reward structure. See this very event and be worthy of what I've been doing so far. But that turns out not to be the case in actually most countries around the world. And so trying to pay it forward and inspired by the presidential award for scientists and engineers that I received years ago from President Obama, I started together with collaborators a similar initiative in Romania and that has been very rewarding. It's really about how you can inspire others to do research and this is a word for young researchers in science and engineer. In Romania as in many other countries they don't really receive much for the research they do, much appreciation. It's mostly sort of the salary they will get but that's about it. And so getting that kind of recognition for the work they do it really makes a difference. So those are, I would say little things that can make a difference in terms of building a diverse a group of researchers who work on these kind of big problems like national language processing or something else. Trying to contribute here at Michigan after I moved here in 2013 and these were mentioned already for instance the Discover Computer Science class that Laura Burdick and I have started with the goal of encouraging diversity. It's a class that's open to everyone but it's particularly emphasizing the importance of diversity plays in computing or explore computer science research which we've been doing for three years now it's the fourth year. The diversity in computing seminar series and also this new project we have for renewing CS, recruitment and retention of women in computer science along again with wonderful collaborators who are really working to make sure that we have a diverse group of people who now are educated in computer science and we will do the research we'll build those technologies of the future and without taking any sort of credit for that back in the day when Janice was a faculty member she was the first woman so there weren't really any faculty aside from herself there were very few women students when I joined in 2014 among students there were about 17% and now we are 22% which is not a big number in absolute but it's really there's a trend and again it's not for putting it next to things that I've been involved with sounds like I'm taking credit for it, absolutely not I think it's every single thing that's happening that actually makes a difference like everyone in this room making those kind of contributions and now at the point where I can I want to shift toward people who do natural language processing so it's really by people but also people who I'm particularly grateful for having them part of my life and my career so these are the people in my research group I think I have the most wonderful amazing research group several of them are here these are all super smart, super hard working very collaborative students working on projects that are making a difference in the world and themselves are paying it forward even if that's not necessarily something that's expected of them at this point in time many of them are mentoring other students many of them are involved in initiatives that are making the difference for others around them I'm also proud I have to say that I try to live by example I think I have a truly diverse research group so we have a wonderful environment and wonderful group meetings because of those diverse points of view that are already brought to the table I'm very thankful for the alumni who went on to do wonderful things each in their own way you've already heard from Laura she's here at the teaching faculty others went on to other places, other directions and I'm super proud of everything they've done and most importantly I'm very proud that they continue to be amazing human beings and I'll be plus there's one group that I'm particularly fond of and that's the AI lab I really appreciate every single individual faculty and staff and students alike in the computer science department but I feel the AI lab has been particularly I personally particularly as an academic home in terms of being very encouraging towards things that I was hoping to achieve and really very supportive since the very beginning so I'm thankful to everyone for their support of course there is all the collaborators, too many to list I try to do justice and I'm sure I miss many things the ones in here is that there is as one former student said I really try in interdisciplinary spaces I love working with people from obviously from within my discipline but from across disciplines and these are just some examples of other departments with whom I've been working with and here on this list I would also mention all the wonderful staff that we have some of whom are here today and have helped to together this event and in general just supporting the work we do on a daily basis thanks specifically to the raw models and in particular I want to highlight Professor Jenny Jenkins today I'm extremely happy to have had the opportunity to meet her today and like I said I'm inspired every day by everything she has done so here is Jenny's some years ago with one of some of the equipment that I learned today she was carrying it around she traveled with equipment around the world and she was as was mentioned by people time she was the first woman in X and interestingly that maybe also pulled some challenges it turned out that people were trying to help her but they figured out that really Jenny's really wanted to carry her own oscilloscope he would let other maybe open the door if it was bargain but otherwise they figured she really wanted that independence and that I think speaks a lot about what it means to be the first in a group and Jenny's was that person she was very dedicated to students she is here in one of the classes that she taught the digital design class surrounded by her students in a news article I read about Jenny's she's quoting saying life is good and the students are the best finally that has been wonderful and inspiring second career like having a bunch of kids again as you know Jenny's joined the computer science department after having raised five children so the comparison was quite appropriate having a bunch of kids again except they are brilliant students don't sass you always take your advice look up to you and best of all I don't have to pay their car insurance so that's the weirdness that many today have remember from Jenny's time here and Jenny's was very encouraging toward women she graduated several PhD students female PhD students she was awarded the Sarah Goddard power award for outstanding professional achievements and contributions to the education of women and she also received an NSF faculty award for women in science and engineering who are here today this is a recent picture with Sarah and Carlos on the left those are my family and then my close friends are here Alina and Isabelle and Carlos and Edo I'm very grateful to them for being there for me and just providing the kind of emotional support that everyone would need this as they go to and last but not least my foundation and that's my family back home they're not necessarily back home they're spread all over the world I have two siblings my brother and my sister to whom I'm really grateful for the times that we spend and continue to spend together and an extremely close relationship that we have and my mother and father whom have been contributing to everything that I am today those of you who know me well now that I grew up during one of the darkest times in Romania's history communism when there was no food to put on the table hours and hours without energy no books, no this, no that and yet my family they raise us in a way that made us appreciate education really be happy for who you are and just having each other so they really achieved a lot very difficult times to raise children they each went on their way and being happy the way I am here today my main message really there is a close connection between natural language processing and people in many different ways it really matters that we think about the impact our research has it really matters that we put people at the center of the research that we do and it really matters who's doing that research and who are the researchers what are the bi people and thank you for your attention