[0:06]Service Now is a platform company, so we provide automation process as a product. So if your product can do a little bit of emotional interaction with you, having empathy towards your problem, that only helps. Welcome to another episode of Tech Bites and I'm your host for this episode Bobby Brill. But more importantly, that is one of our guests on this episode, Mi Dam Kim, senior linguist for generative AI here at Service Now. And along with our other guests today, Roberto Bacci and Ryan Dukes, we are going to explore the importance of linguistics in making AI successful and the goal in making large language models more humanlike and emotionally intelligent for customer interaction. But to start us off and to get to know everyone as we are lucky enough to have three guests today, Roberto, why don't you go first and tell us a bit about your background? Yeah, I'm Roberto Bacci. I'm the manager of content development in ATG at service now. I manage a team of sixteen linguists at the moment. My background is in linguistics and foreign languages in literature with a BA from University of Bologna in Italy and a PhD from Brown University. So you are truly an expert in language. I hope so. Kim, uh introduce yourself and tell us a little bit about your background here. So hello, my name is Mi Dam Kim. I'm from originally from South Korea and I did my BA there at Seoul National University and also Masters both in linguistics. And then I went to Northwestern University for my linguistics PhD and my focus was always speech sounds and cognitive science. And I also taught in the School of Business University of Kansas for eight years, uh focusing on the communication between speech sounds and language and conversation and business. Because in my thoughts everything's foundation is language and communication. And that's how I was led to service now and now try to make bridges between different worlds and try to enhance customer's experience with LLM. Okay, just how we sound, how we present ourselves and how we talk really is that whole key to making us communicate better, correct? Yeah, how we sound, what not just what we what we say, but how we sound is very, very important. Because it has lots of information that's not in the text, that's not in the content, but uh that kind of speech tones and your voice and how you raise your voice, how you amplify your voice, how even your gesture, everything goes into your communication. Ryan, uhm, last but not least, introduce yourselves and and tell us how you came about to service now. My name is Ryan Dux. I've been at Service Now about two years. Prior to that, I was a German professor actually, so I have my PhD in German focusing on linguistics as well. I went to grad school at UT Austin, but also lived in Germany a few times and now want to do the transition from academic world service now. My main task is working on the question answering the Q&A project to improve the search results and with the development of LLMs, we are able to try to refine the results that people get from their when they when they search a query, make something more advanced, more user-friendly, more clear to the point to address what the what the user wants to know about. So we are going to try to cover a lot in a short amount of time, but Roberto, can you give us an overview of what the world of linguistics looks like in tech? Uhm, I would say in a tech company, linguists can range from computational linguists who are basically a figure that is a mix of an engineer and a linguist to other linguists who are more similar to storytellers or play rights who write dialogues and are being like creative. I would say like to boil it down, there are basically three main things that we do. One is creating synthetic data. So creating synthetic data is important because it is new data that does not contain private information, is not customer data. It's creation, is really creation of basically from scratch. The second thing that we do is labeling and leading labeling teams, that is also evaluating, assessing the data that that we have created. And the third part we have not mentioned it yet, but is also localization. This is maybe kind of task that happens less often, but we also localize and translate and adapt English data to other nations. For the creative aspect we create a variety of synthetic data. It can be dialogues, can be isolated utterances, can be knowledge-based articles, can be utterance maps, can be fake incidents or fake tickets. So this this creative aspect is is is a very important aspect of our job. Together with another point which is our role as a bridges or intermediary between on one hand engineers and data scientists and on the other hand, for example, labeling teams.
[5:46]We are in between these two elements of the company, so we very often also collaborate in creating guidelines, in explaining complex concepts in a way that is organized and and simple to absorb. So just how important is linguistics to making AI successful? It's called large language model, right? So large language model or this language technology is a big part of the current AI boom.
[6:18]And I think that is coming from our human desire to communicate more effectively and efficiently with larger data and larger audiences from far away from where you are, uh in a very convenient way. So it is all from our desire to communicate better. So in that way, linguistics, linguists and language uh are all very, very important. You know, when we're talking about linguistics, we can say like there's like the four Cs that we have. So we're we're critical, we're creative, we're culturally aware, and we're communicative in the sense where we can communicate between different teams, between like the data and the product, between, you know, the labelers and the engineers. These seem so powerful that we need to, you know, use these more humanistic skills of communication, cultural awareness in order to harness the power of LLMs and ensure that they meet service now goals of being trustworthy, accurate, user-friendly and making work more effective. Yeah, language seems to be where some of the pain points happen, is that correct? From a linguistic point of view, yes. Oh, okay. from a linguistics point of view. Well, well, explain that, explain the difference between that, Roberto, linguistics and language and how is that separated or what do we really need to focus on in that aspect? Yeah, we need to focus on on both, I would say, so language is a human complex communication system that can happen through speech, through text or through science. And linguistic is the scientific study of language and also the structure of the language. There are many subfields in linguistics. Linguistic can include the study of sounds, for example, phonetics and phonology. It can study word and sentence structure, so morphology and the syntax. It can study meaning, so semantics. It can study language use, so pragmatics, and the historical and social aspects of language, so historical linguistics or philology, how it's still common to call it today, especially in Europe. They constantly evolve as technology evolves, so this is a point of particular interest for us. We are in a period of extremely fast technological changes that of course have an impact on linguistics and vice versa. And also linguistics changes and evolves also with interaction with other disciplines. For example, Midam, you have studied also cognitive linguistics that can be a subfield of linguistics. Yeah, so in cognitive linguistics, cognitive linguistics is a study that explores the relationship between language and cognition, and we researchers have been investigating how language reflects and shapes thought processes, studying concepts such as metaphor, cognitive blending, and the cognitive aspects of grammar, for example. So basically understanding language as a sub part of cognition, as we do in cognitive linguistics, is enabling us to study language from a variety of different perspectives, including language acquisition, language development, language deterioration, um and and language communication, sensory perception production, neurological mechanisms, and multimodal processing and and everything. So this really enables uh language studies to uh so that it's it can blend into other studies in cognitive science. When you're talking about cognitive science, is this the idea of how people think or why people think and how does language plug into that? All of that, all of that. So this is not this is not going to be an easy conversation. This is a lot of stuff. This is okay. Yeah, cognitive science is a huge, uh it's a it's a vast amount of study and cognitive linguistics is part of it. So this is how we learn, that be the right way of looking at it? How we how our how our brain functions and develops and deteriorates, um and use and interacts uh with exterior environments. So that makes a lot of sense if we're if we're constantly reading and listening to something and our entire day is based upon reading and writing if you're working on the internet or working on anything tech related or at this point we're talking with service now and talking a lot about, you know, help desk things and and really interacting with people via copy via text.
[11:07]This plays directly into that concept of does the other person I'm interacting with and also the person who's doing the interacting, do their minds meet at the same point? Is that correct? It's about interplay of your mental model within yourself and with other people. Okay. Does it make sense? That makes sense, but that I heard you that makes so much sense, but it's like, wow, that that seems very almost science fiction. Oh, I think that's what we're doing right now. That's what I and you are doing right now. Okay, so it's more science, science realism then. So how how do you apply these linguistics and Ryan, you you have some insight into this. How how are we applying linguistics into into what we're doing? There's a lot of ways that uh linguistics can be applied to their in various topics within the academic field of linguistics. They make a distinction between theoretical linguistics and applied linguistics. Uhm, and applied linguistics can cover all of kind of the main theoretical aspects of language structure, you know, the um the sounds, the putting words together, sentences together, pragmatics, but rather than focusing only on how language works, understanding how language works, uh in its own right. Uh, applied linguistics tries to see how we can apply what we know about how language works to real world problems, um real world situations, making making the world better. Traditionally, applied linguistics is relegated to thinking about like, oh, it's just language teaching, language teacher training and, you know, how to get students to learn language better, but there's really a lot of ways that language can linguistics is applied, um. So the, you know, there's um translation, forensic linguistics, you know, trying to solve mysteries with it. Uh, a lot of sociolinguistic issues, trying to determine how language is used to mark somebody's social standing, uh interactional, how they're interacting with different people. And when it comes to large language models and generative AI, you know, they say applied linguistics is kind of like language teaching. Really, what we're trying to do is get large language models to learn how to speak a language that is more natural, more humanlike than just kind of regurgitating data. So how do we get that answer into a language that somebody can understand better over a chatbot? Or is that the correct way of thinking about it? You know, LLMs are initially created as just by a lot of just feeding it data, usually text data that comes from the internet or comes from just random texts and texts themselves are not human, you know, really the things that Mi Dom was talking about how interacting the tone is really, um, really plays a lot into how what you're trying to communicate. Because LLMs are just the way that they work is basically they're given a huge bunch of text and rather than actually understanding how to create language, they take the text that they've been given and draw on the probability, the statistical probability of what the next word is supposed to be of what the next proper response to a given question would be most likely. And because of that, they're very good at kind of the really structural parts of language, being able to put together words, put together sentences, knowing the grammar rules, but the other fields of language, specifically semantics, the study of meeting and pragmatics, kind of the study of more interactional aspects of language. Those can't be gained just from statistical probability and those are the essential parts of human communication. So linguists are important because with kind of the base language models, they are because they're only based on statistical probabilities, they can sit out information that's either like untrue or it's not culturally sensitive and kind of generic. So we need to use linguistics to try to train them to better understand people and to communicate in a way that's more acceptable for normal human users. And and Roberto, you you you've got some insight into that because I wanna ask you how do we get humanlike communication with all of this data? So from a linguistic perspective, humanizing, for example, an AI powered bot involves not only mimicking the language patterns of human communication, but also understanding and responding to the emotional and cultural nuances that are inherent to human communication. We don't only communicate words, but we also communicate emotions and also our culture. So there are several ways an AI powered bot can be humanized. One of them, for example, is natural language generation or NLG. So the NLG techniques allow the AI bot to generate responses that mimic natural human language. In many ways like AI models can be compared to imitators, impersonators, mimics. They are very good at mimicking the language of humans. So NLG involves also incorporating variations, colloquialisms and conversational tone to make interactions feel more human and less robotic. Another way is the recognition, emotional recognition and expression, how to express and recognize emotions. This enables AI to understand and to respond appropriately to the emotional cues of the user. So the model needs to be able to catch to understand the the subtle emotional signal that the user is sending. This can involve using emotive language, language expressing feelings, adapting the responses of AI, basing them on a detected sentiment. And after identifying making the interaction resonant, emotionally resonant with that sentiment. What are some some things that you're seeing or or give us some insight into that process of finding emotion within within the AI. That can be, for example, uh when we create synthetic data for for AI, we can take into consideration different personalities. If let's say we are recreating dialogues between a user and AI, we can have we can imagine different types of users that can have like different attitudes toward the bots. Another aspect related to this is that of politeness and empathy. Basically, we need to develop algorithms that simulate empathy and politeness in order that the conversation is on both sides respectful and considered. The side of the AI, but also on the side of the user. For example, I encountered in my career in in in different companies, let's say when a human interacts with a bot, for example, sometimes the interaction can be very, very rude from the human side. Like people, if you think of of assistance even outside of Service Now, like Alexa or Siri or Google Assistant, people may talk to these assistant like servants, you know, do this for me, you know.
[19:03]Right, we've all been an aggressive user with this type of technology at some point. This is not the case of service now, but in other companies many of the devices are also used by children, for example. What if a child ask a question in a very rude way, using bad words like should we actually should a bot actually answer or should the bot have a different behavior? Similar to to this is uh you know, should a user should a polite user be rewarded for asking politely to the bot compared to a rude user.
[19:40]Yeah, that that's almost even more humanlike that if you're a nice person, then I'm going to help you even better, like a good a good waiter is a better waiter because his customers is a nicer person. Exactly. Is being nice and making people happy what we always want, Mi Dom, you you've got I know you've got an opinion on this. I guess always, because we always want nice people to talk to, right? Uhm, there can be different types of use cases in different types of industry or different types of customers who might want different degrees of this kind of friendliness or warmth. I did a presentation on my research on that to the um internal engineers at service now and also externally in a conference. And their response was that I actually don't need that. I don't need that much of friendliness or niceness because I I focus more on the accuracy of the responses of the models. So different people might need different levels of friendliness, but I guess overall, as a conversational partner, you have to be nice, otherwise people the the other body will not listen to you. Well, well, that's an interesting thing you brought up where a a polite interaction may in some cases may slow down the process. Is that the right way of looking at it? Yeah, that is possible. That is very possible. One example I got from my interaction with an engineer was actually that so you have an emergency, you like cut your hair, cut your hands or whatever. I mean, it's it's a very, very, uh emergency situation. You just need the an answer. Then you ask the question to the LLM and then it says, nice to meet you, how can I help you? Then then you get so irritated. Just give me an answer, right? So that is possible too. This idea of empathy and emotional recognition and expression where we're actually helping our customers, is the empathy and emotion um, something we're finding that is really important? I think so. I think so, because Service Now is a platform company, so we provide automation process as a product. Many of the customer facing points are chatbots. It's around uh language and communication all the time. That's how customers are experiencing our product. When would you use uh a LLM or any of our product? When I I've got a problem. When I I've got a problem. Solve my problem. You are stressed, you are nervous, you are mad, right? So you have this negative feeling and you are full of feeling. Then if the chatbot is so dry and not really addressing my emotion, then that is just at least not helpful. So if your product can do a little bit of emotional interaction with you, having empathy towards your problem, that only helps. So if I can add something to this, we are talking of humanizing AI bots and one thing that service now cares a lot about is trust. Okay, so how is trust built and measured? I think having bots and virtual assistants that are not offensive are of course helpful, but also they are sensitive to customers, for example, in different regions. They don't provide harmful answers to the questions of the of the users, that's a very important point. Trust is measured in how helpful is is my is the solution that I'm giving to you, but is also measured the same as a person. Is a person polite and helping you or is the person offensive and harmful to you? So what we are saying here about language also has to do a lot about with trust. So there you have it, a quick primer on the world of linguistics and generative AI. And if you like what you're hearing, please hit subscribe on whatever platform you're listening on so you never miss an episode. And for more information, you can check out our newly designed docs website, docs.servicenow.com. I'm Bobby Brill, thanks for listening.
