Latest / The Adults In The AI Room / Linda Andersson on AI Powered Semantic Search - Episode #5

Transcript

Linda Andersson on AI Powered Semantic Search - Episode #5 · The Adults In The AI Room

• 0:22
  SPEAKER_00: Thanks so much for listening.
• 0:23
  SPEAKER_00: If this is your first time, we invite you to subscribe wherever you listen to the podcast.
• 0:28
  SPEAKER_00: And now you're your host, Seth Early at Chris Featherstone.
• 0:38
  SPEAKER_01: Good morning, good afternoon, good evening, depending upon your time zone.
• 0:42
  SPEAKER_01: Welcome to today's podcast.
• 0:44
  SPEAKER_01: I'm Seth Early.
• 0:45
  SPEAKER_03: And as always, I'm Chris Featherstone.
• 0:48
  SPEAKER_01: You're still Chris Featherstone.
• 0:51
  SPEAKER_01: So I'm saying before we jump in, I just want to thank our sponsor very quickly, uh Early Information Science.
• 0:58
  SPEAKER_01: That's my company.
• 1:00
  SPEAKER_01: CMF Wire, a great uh website for uh up-to-date uh thought leadership and information about uh all of the technologies that we're we're using today and the Marketing AI Institute.
• 1:13
  SPEAKER_01: Uh they have some awesome courses.
• 1:15
  SPEAKER_01: Uh you should definitely check them out, and they have a terrific conference as well.
• 1:18
  SPEAKER_01: I actually used them uh for some case studies when I wrote my book uh a couple of years ago.
• 1:24
  SPEAKER_01: So uh really, really great uh organization.
• 1:26
  SPEAKER_03: Awesome.
• 1:27
  SPEAKER_03: Hey, you know, I'll tell you what, as still being Chris Featherstone, Seth, uh, you know, I cannot believe that we're already in 2022.
• 1:34
  SPEAKER_03: It's crazy.
• 1:36
  SPEAKER_03: I know, I know.
• 1:37
  SPEAKER_03: It's like, what?
• 1:38
  SPEAKER_03: What happened?
• 1:39
  SPEAKER_03: And it's it's it's been, you know, I I honestly I've been a little bit bummed because we haven't had any uh guests since, you know, since what I want to say the beginning of December, right?
• 1:50
  SPEAKER_03: So it's it seems like it's it's kind of too long for me.
• 1:54
  SPEAKER_03: Yeah, you know, yeah.
• 1:55
  SPEAKER_01: I I definitely want to increase our cadence here if we can, because we have so many guests lined up and so many great topics.
• 2:02
  SPEAKER_01: So it's really gonna be exciting to get to stuff.
• 2:05
  SPEAKER_03: I'm looking forward to 2022 as well.
• 2:08
  SPEAKER_03: Um, you know, besides you know, everybody's staying healthy for sure.
• 2:12
  SPEAKER_03: That's kind of number one.
• 2:13
  SPEAKER_03: But two, like you said, just the amount of of uh people that we have lined up and the diversity of people that we have lined up.
• 2:21
  SPEAKER_03: I mean, we've got folks from big companies, from consulting companies, from you know, from uh product companies, from hardware companies, and everything in the company.
• 2:29
  SPEAKER_01: Academia, researchers, absolutely some of the greatest minds in the space.
• 2:34
  SPEAKER_01: So it's very exciting and agree.
• 2:36
  SPEAKER_01: Yeah.
• 2:37
  SPEAKER_03: So with that, who's holiday, by the way?
• 2:39
  SPEAKER_03: It was how was your holiday?
• 2:41
  SPEAKER_03: It was uh uneventful, which is I think the the the good thing, you know.
• 2:47
  SPEAKER_03: How about you?
• 2:48
  SPEAKER_01: I actually went to a yoga and meditation retreat for six days, uh place called Kropalu, which is in western Massachusetts.
• 2:56
  SPEAKER_01: And all I have to say is I wasn't looking forward to it.
• 2:58
  SPEAKER_01: My wife arranged it and I was there and I was like, okay, I'll deal with this.
• 3:02
  SPEAKER_01: But you can't do that stuff for like four, six, eight hours a day and not have it impact you.
• 3:06
  SPEAKER_01: It was very, very meaningful, really valuable.
• 3:10
  SPEAKER_03: I don't know if this is a yoga thing, but you know what they say, you know, those that are flexible won't get bent out of shape.
• 3:15
  SPEAKER_03: So I think you're you look relaxed.
• 3:17
  SPEAKER_01: I get bent out of shape a lot.
• 3:20
  SPEAKER_01: That's probably why she sent my flexibility.
• 3:24
  SPEAKER_03: Well, uh, you know, onward and and upward with uh great stuff.
• 3:29
  SPEAKER_03: So I'm excited about today.
• 3:30
  SPEAKER_03: So I'll let you uh take it.
• 3:32
  SPEAKER_01: Absolutely.
• 3:33
  SPEAKER_01: We have a wonderful guest today.
• 3:35
  SPEAKER_01: She is a true polymath with expertise in library and information science, programming, machine learning, AI ontologies, near and dear to my heart, information architecture, also near and dear to my heart, and has a background in research and academia.
• 3:52
  SPEAKER_01: She has multiple patents for her work in natural language processing and has developed innovative AI-powered semantic search tools for scientific and patent research.
• 4:01
  SPEAKER_01: Her work has been adopted by a range of institutions, and she is at the leading edge of cognitive AI.
• 4:07
  SPEAKER_01: She is the founder and CEO of Artificial Researcher, a startup based in Vienna, Austria.
• 4:13
  SPEAKER_01: Please welcome Linda Anderson.
• 4:16
  SPEAKER_03: Yay, Linda.
• 4:17
  SPEAKER_03: It's good to have you with us.
• 4:18
  SPEAKER_03: Hi.
• 4:19
  SPEAKER_03: Thanks.
• 4:19
  SPEAKER_01: We're waiting for the audience to pause.
• 4:21
  SPEAKER_04: It's all different.
• 4:25
  SPEAKER_04: Uh thank you.
• 4:26
  SPEAKER_04: And I'm very happy to be here.
• 4:28
  SPEAKER_04: And thank you for the nice introduction.
• 4:31
  SPEAKER_04: Um and I'm also honored to be the first to be on your podcast this year.
• 4:38
  SPEAKER_03: Yep.
• 4:39
  SPEAKER_03: You know, we we chose you specifically.
• 4:42
  SPEAKER_03: I mean, besides your track record, it was uh also um we were grateful that you would uh, you know, given the time zone that you're in, stay up at all hours to actually help us.
• 4:52
  SPEAKER_03: So we're like, yeah, let's let's give her the first one.
• 4:56
  SPEAKER_04: Yeah, yeah.
• 4:58
  SPEAKER_04: Thank you.
• 4:59
  SPEAKER_03: It's fun.
• 5:00
  SPEAKER_03: So you know what, Linda?
• 5:01
  SPEAKER_03: I think um, I think it's super interesting because of your background and given you know that it seems like you know, the the different the low the degrees and the focus and everything else has just been building blocks one after another after another after another.
• 5:17
  SPEAKER_03: I'd love to get a sense from you as to, you know, we'll we'll get into the you know origin of how you start and everything else, but I'd love to get a sense for you know why that path and maybe you know how that's been helping, right, in terms of the work that you've been doing.
• 5:33
  SPEAKER_04: Yeah, I mean, I've always been interested in information and knowledge, and it could have to do with I was a late late starter reading books.
• 5:42
  SPEAKER_04: So uh when I when I have cracked that encryption of how to read, I my idea was if I could stay in a library my entire life and just read books, it would be lovely and structure them, structure the books in different ways to extract knowledge.
• 6:00
  SPEAKER_04: That is my uh truly interest.
• 6:04
  SPEAKER_04: And there's so much things you can learn from actually extracting knowledge from different books or different uh mediums, and then join them together and come up with new solutions which you have not thought about uh before.
• 6:17
  SPEAKER_04: And that's always been something that I don't, regardless if I've been doing legal as a 12-year-old or my first programming, which I did when I was nine, and that was back in 1984.
• 6:29
  SPEAKER_04: So then I programmed in BASIC.
• 6:33
  SPEAKER_04: Um so I always like to do different stuff and extract information and knowledge, structure them and see how I can use it.
• 6:42
  SPEAKER_04: So that that's part of my life.
• 6:45
  SPEAKER_03: So let's double-click on one thing that you said on um, you know, because I don't know that I remember cracking the code on uh on reading and opening up books, right?
• 6:54
  SPEAKER_03: Because it's it's something that's done from from a young um from just a young age, right?
• 7:00
  SPEAKER_03: And by the way, the beauty of of our pack podcasts is that not only do I look like I'm in my 20s, so people, you know, will you know, they they but I do I do I do remember writing go-to statements as well.
• 7:14
  SPEAKER_03: So I'm right there with you.
• 7:16
  SPEAKER_03: Now, my go-to statements were to build images, right, in in basic and stuff and and save them on a on a you know, either a a floppy drive or you know, or whatever it was at the time, right?
• 7:28
  SPEAKER_03: And making sure that we loaded and and then save.
• 7:31
  SPEAKER_03: But um let let me let me, I think, because I think it's really, really important to double click on on what you said in terms of of your when you actually really knew and understood, like, you know, to want to just sit there and read and understand and ingest information.
• 7:48
  SPEAKER_03: Because apparently, from what you've told us in the past, that that was a really difficulty for you in the beginning, right?
• 7:54
  SPEAKER_03: So will you take us on that journey a little bit and kind of you know tell us how that how that you know that clipped for you?
• 8:00
  SPEAKER_04: Yeah, okay.
• 8:02
  SPEAKER_04: So when we started reading around when we were seven in Sweden.
• 8:05
  SPEAKER_04: So that's when you go start in school.
• 8:07
  SPEAKER_04: Now it's six, but when I started back in the 80s, it was seven, and everyone could read after just a few months, and I I couldn't even understand how these different uh I would say symbol, I didn't under know it was letter, how these different symbols actually could be joined together and come out as a word, and then you could go uh and then you can connect all these words into sentences and so on.
• 8:33
  SPEAKER_04: For me, that was a puzzle.
• 8:35
  SPEAKER_04: And um and I I really struggled with that for a long time, and that's because of my dyslexia.
• 8:43
  SPEAKER_04: So everything when when people, when my teacher said, okay, uh, you hear what's right, you hear how you should pronounce this, and I didn't hear, I didn't understand what it is I had perfectly here, I could hear, but I didn't hear how these were connected.
• 9:00
  SPEAKER_04: Uh so I just uh started to trying to encrypt as decrypt how a different sequence of letters are connected to a specific sound and how it's semantically connected to a meaning.
• 9:15
  SPEAKER_04: So I seen is saw in patterns instead of putting uh letters together and form a word.
• 9:23
  SPEAKER_04: I know that if it's these four sequences of letters, then this is this word.
• 9:28
  SPEAKER_04: So I had to jump uh directly to that.
• 9:32
  SPEAKER_04: And everything I had to learn, how to join a sentence together in order to do it grammatical correctly, that was rules.
• 9:39
  SPEAKER_04: That was explicit rules that I had to learn exactly as a computer.
• 9:45
  SPEAKER_04: So uh one thing I usually joke about when I talk about my dyslexia is that I'm as stupid as a computer because I need everything explicit.
• 9:56
  SPEAKER_04: And this is what I find very fascinating that if you can decode language and make all of these things that people know is correct just because they know it, they hear it.
• 10:09
  SPEAKER_04: But for computer, the computer can't know it's correct.
• 10:13
  SPEAKER_04: You need to tell them why it's correct, you need to give them the rules.
• 10:17
  SPEAKER_04: And that is how I work also.
• 10:19
  SPEAKER_04: I need to have all the rules, all the exceptions, and then I can do a grammatically correct uh sentence or write or and these actually have given me an edge because I look at a text as a computer would look at the text.
• 10:35
  SPEAKER_04: It's just a lot of symbols that you join together and you need to extract them in order to see what type of information you can get out.
• 10:45
  SPEAKER_04: Um, so when I see a sentence, I usually see a tree structure where the noun phrase is, where the part where the verb phrase is, and I see the structure and know that okay, the noun in the beginning and the noun in the end has a connection, and the verb is actually saying what connection it has to between each other.
• 11:06
  SPEAKER_04: So if I give Seth a book, that means that I'm the person doing something, and the book is something that I'm giving to Seth.
• 11:16
  SPEAKER_04: So that's these are the things, and you you see it as uh like a tree structure.
• 11:22
  SPEAKER_04: And you can't say I give Seth a book.
• 11:28
  SPEAKER_04: I need to see say something more, uh you know, how it should be uh uh connected.
• 11:35
  SPEAKER_03: So what age then did you were you able to, you know, derive out the narrative and the context, right?
• 11:41
  SPEAKER_03: To carry over, because it's one thing to look at all those, but you know, when we get into some of these, the bias of the narrative or the bias of the context changed the meaning of the sentence, sentence structure and the me and the content, right?
• 11:54
  SPEAKER_03: So what point did that kind of you know start to be a reality for you where you could see the force from the trees?
• 12:00
  SPEAKER_04: Um around 12.
• 12:02
  SPEAKER_04: Before that, it was more or less me struggling with one putting one letter before another letter.
• 12:08
  SPEAKER_04: So I cracked them.
• 12:10
  SPEAKER_04: I actually was doing programming before I could read.
• 12:13
  SPEAKER_04: So the so the math was much easier for me because it was number and just logic.
• 12:19
  SPEAKER_04: Meanwhile, this with what why should you use um uh S on a third person in English when it's present?
• 12:29
  SPEAKER_04: That why should I?
• 12:31
  SPEAKER_04: And and you know, all of these question, why should I do that?
• 12:36
  SPEAKER_01: I'm I'm not a rule follower, but why should I want to do it?
• 12:40
  SPEAKER_04: So so around 12 I started reading a lot of books, so I probably read one book uh per week or um something like that.
• 12:49
  SPEAKER_04: Um sometimes it was more, and I read every type of book.
• 12:53
  SPEAKER_04: So of right now I'm reading crime, uh crime a lot in German because that's so my way of practicing German.
• 13:02
  SPEAKER_04: Uh but uh when I started when I was 12, it was all over the place.
• 13:07
  SPEAKER_04: It was Shakespeare, it was everything that I could come across.
• 13:11
  SPEAKER_04: Um so uh yeah, and uh from that point, it has always been interesting for me to extract knowledge.
• 13:19
  SPEAKER_04: And also one thing when you read when you're 12, you need to start doing your exams, and you need to uh actually also extract information from books.
• 13:29
  SPEAKER_04: And I'd never had the time to read all of these big books to get all the information right.
• 13:35
  SPEAKER_04: So I needed to learn how to extract information quickly from books because I didn't have time to read the entire book.
• 13:43
  SPEAKER_04: So if uh if it was some kind of uh physics or chemicals or something, I just needed to identify what's the most important part, which will be on the exam.
• 13:53
  SPEAKER_04: So that so I started actually doing information extraction manually at the 12th because I needed to.
• 14:01
  SPEAKER_04: Exactly.
• 14:02
  SPEAKER_04: Yeah.
• 14:03
  SPEAKER_01: But the framework and the mindset and the the approach is really what's very analogous to what you're doing now.
• 14:09
  SPEAKER_01: Uh talk a little bit about um artificial uh researcher and the kinds of semantic work uh combined with AI.
• 14:17
  SPEAKER_01: What I liked about uh when we first started conversations is you're really at that intersection of ontology's information architecture and artificial intelligence.
• 14:26
  SPEAKER_01: And there's a lot of folks out there that are laboring under the misconceptions, maybe a little bit less right now, uh, but but certainly uh that oh, you just throw the you just point the AI to all the data and all the content and it's figures it out, right?
• 14:41
  SPEAKER_01: And um and so you know our philosophy is that you can do that to some degree, but you really do need to think about that core architecture and the semantics.
• 14:50
  SPEAKER_01: Do you want to talk a little bit more about your philosophy when it comes to that intersection?
• 14:56
  SPEAKER_04: Uh so my um my philosophy when it comes to this is that domain knowledge is the key.
• 15:04
  SPEAKER_04: And domain knowledge uh exists in ontology, some kind of structure knowledge.
• 15:09
  SPEAKER_04: And if you want to start mining, text mining uh a new area, you need to have this type of thing.
• 15:18
  SPEAKER_01: Do you want to provide your quick definition?
• 15:21
  SPEAKER_01: I have a quick one as well of ontology.
• 15:24
  SPEAKER_04: Yeah, so I will provide you the linguistic definition, which is probably a bit different.
• 15:30
  SPEAKER_04: So if you have, for example, WordNet, where you have um you have phrases, so domestic pet is a phrase which is apparent to, for example, dogs, cats, and so forth.
• 15:43
  SPEAKER_04: So you have some kind of structure of that.
• 15:46
  SPEAKER_04: So you know that um how how different phrases are connected and related, and that's also how you explain certain concepts.
• 15:55
  SPEAKER_04: Um, what's what's a synonym to cat?
• 15:59
  SPEAKER_04: Is it a feline synonym?
• 16:01
  SPEAKER_04: Yeah, it's partly a uh a synonym, but it's actually a parent, it's uh it's a broader term for all the felines.
• 16:10
  SPEAKER_01: So taxonomies are certainly part of that parent, child, whole part relationship.
• 16:14
  SPEAKER_01: And then we have multiple taxonomies uh that describe any knowledge domain, and then we have the relationships between them, right?
• 16:23
  SPEAKER_01: So when I think of ontology, it's it's just it's the know I call it the knowledge scaffolding of the organization.
• 16:28
  SPEAKER_01: I think of it as the framework in which you can hang your knowledge, and so it's taxonomies, the source structures, and and the source just has the equivalent terms and as and related terms, and then all those relationships between them.
• 16:42
  SPEAKER_01: You know, if I have products and I have services, here's the services and that product.
• 16:45
  SPEAKER_01: If I have a problem and a solution, I can have two taxonomies, and then here are the solutions for those problems, that type of thing.
• 16:53
  SPEAKER_01: So that was my quick definition.
• 16:55
  SPEAKER_01: Did you want to uh continue on and talk about how that relates to uh AI and how that relates to some of the work that you're doing in knowledge extraction?
• 17:05
  SPEAKER_04: So mostly for the knowledge extraction, we are using these broad and specific terms.
• 17:09
  SPEAKER_04: So we actually go in and take a huge, large amount of data from a specific domain because different words have different meanings in different domains.
• 17:18
  SPEAKER_04: So if I take the word bus, uh we we here, all three of us, will think about different words, uh so different association.
• 17:28
  SPEAKER_04: But um if you think about in computer, it would be bus slot card.
• 17:34
  SPEAKER_01: Bus was the term you said, right?
• 17:36
  SPEAKER_01: B US.
• 17:40
  SPEAKER_01: And if you if you think you a bunch of people get in to go from one place to another, a bus.
• 17:46
  SPEAKER_04: Exactly, exactly.
• 17:48
  SPEAKER_04: But uh that is not a bus slot card though, that is a double wheel bus.
• 17:53
  SPEAKER_04: So it shows that your semantic actually goes to the double wickel bus, actually the transportation.
• 17:58
  SPEAKER_04: Meanwhile, uh a bus, the bus meaning in bus slot card is rather that it's still a transportation, but it's digits that uh something uh like bin codes that go, or let's say numbers goes from a position A to a position B.
• 18:17
  SPEAKER_04: So there is it's still the same function, it's actually coming down to what we call semantic primitives that bust have the function of being uh um uh something that takes people, numeric numbers, whatever from a position A to a position B.
• 18:36
  SPEAKER_04: Okay, that's that's that's that's there.
• 18:39
  SPEAKER_04: And um so that is actually also when you think about semantic ontologies and special lexico-semantic ontologies, what is the function of this word?
• 18:51
  SPEAKER_04: And when it comes when it comes to bus, it's quite interesting because it actually can mean it can be a double wiki bus, a transportation, but it can also be if you train your algorithm, for example, on a patent and you forgot to do a good balance in your training your deep learning algorithm, it could be that this word bus would only be associated with memory and register because that is actually the context where they occur, and that's what the computer learns gets teached, or that's what we uh teach it.
• 19:25
  SPEAKER_04: And that's where we come back to this explicit and implicit information.
• 19:29
  SPEAKER_04: So if you want to make something explicit for a computer, you need to provide it with all the knowledge that bus can also be a double-whickled bus, and you need to provide it in a certain statistical level, because if you just have one occurrence where it's a double-whickled bus and 250 occurrence where it's uh like a bus slot card, then it will be only associated with re-istere and memory and so forth.
• 19:55
  SPEAKER_04: And that's where we have this key thing.
• 19:57
  SPEAKER_04: Why do we need ontologies?
• 20:00
  SPEAKER_04: Ontologies is some kind of manually curated or semi-manually curated, which can help the computer, the deep learning algorithm, to understand that okay, there is different types of bus.
• 20:14
  SPEAKER_04: I'm looking for, I just found one meaning of bus.
• 20:17
  SPEAKER_04: I need to find the other.
• 20:19
  SPEAKER_04: Otherwise, if you let it go by itself, it will only learn what was most occurring.
• 20:25
  SPEAKER_04: The the more newer algorithm can differentiate a bit better, but it's still so they need to occur.
• 20:33
  SPEAKER_04: Or you need to show to the algorithm that this is what you should be looking for.
• 20:43
  SPEAKER_03: On the we'll see the the bias that can occur that they don't know that they don't know.
• 20:50
  SPEAKER_03: That makes sense.
• 20:50
  SPEAKER_03: Which which is which?
• 20:52
  SPEAKER_03: Uh and so when they when when these organizations get in there, to your point, they're looking for bus and they put that out as an initial thing, but they don't know that they should be looking for, let's say, computer bus versus transportation bus.
• 21:04
  SPEAKER_03: Right?
• 21:05
  SPEAKER_03: So how do you how do you um coach them in the in these kind of things?
• 21:09
  SPEAKER_03: Because it seems to me that we don't get into really good with organizations, you know, in general.
• 21:16
  SPEAKER_03: We don't get into this level of discussion in kind of the first go-round.
• 21:21
  SPEAKER_03: Like you, like Seth said, here's a bunch of data, throw an AI model at it and see what comes out, which okay, that's one approach.
• 21:28
  SPEAKER_03: But you know, how would you how would you instruct an organization then to be cognizant of that in the beginning?
• 21:36
  SPEAKER_04: So uh uh I can take the example I did in my PhD.
• 21:40
  SPEAKER_04: Uh so I worked with uh patent engineers and patent examiners, and they have just this problem with the word of boss or cell or whatever of these uh words that has several different meanings.
• 21:52
  SPEAKER_04: Uh, but the primates, the primates and uh uh the semantic primates is still similar, is that okay, we to use the taxonomy and now I'm not talking about this uh not the taxonomy of of of words but rather of what can which field technical field are you in and then you look at what's the meaning of this word in that field so that's the first thing you do you actually need to apply structure data so you need to take whatever taxonomy you have if you have uh for example international patent classification which is like a um have um eight uh sections and then it goes down to being uh it goes down to class subclass um uh and then you have uh main group and group subgroup so it's quite broad in or narrow grounded in several millions examples in the end but this is a ways for so saying okay when it's when the word occurs in this context it belongs to transportation when the word occurs in this context which you have labels on it is referring to some kind of memory some kind of type of uh micro shape something in that and that is the first thing to actually say you can't chunk everything into one pot and hope it comes out good that I'm not for generalized models there are cases where you can actually do that but you need to actually be aware in which context are you going to use this model.
• 23:34
  SPEAKER_04: So if you're going to work on developing uh searching on a transportation patent you need to make sure that you actually have the if you want to build a search engine using distribution and semantic you need to make sure that the training of that algorithm is on that domain.
• 23:53
  SPEAKER_04: And then people saying this is where it becomes very particular.
• 23:57
  SPEAKER_04: Okay then I only train it on that domain.
• 24:00
  SPEAKER_04: That's not good either because you have a general understanding of language.
• 24:04
  SPEAKER_04: So you need to have a balanced corpus that's what we say when you picked uh from all types of areas of a language that is covering the basic grammatical explicit information that you have the semantic information that you can derive from a collection and on top of that you put the target area you want to go to but only training it on a specific domain means that it will not understand what the word dog is if it's not existing there.
• 24:37
  SPEAKER_04: It's used probably in in patent will be an acronym for something and it won't understand that it's actually an animal.
• 24:44
  SPEAKER_01: So you you you lose that type of training and and and that makes a lot of sense you know when you're talking about begin with the structure the architecture the concept space the concept space has to be related to the domain in which you're working in which you're trying to retrieve information.
• 25:06
  SPEAKER_01: So that's kind of a the architectural perspective this is the reference data right this is telling the system what's important and what terminology means and what the relationship that terminology to other terminology and that's very helpful for search.
• 25:20
  SPEAKER_01: Well let's also talk about so that that's one way of contextualizing queries and and uh uh retrieval and then we have to look at the content itself the knowledge base the actual assets of the the the corpus the body that we're going after and a lot of folks are thinking well do we really need to think about structuring and curating that uh can't we just you know again with uh our algorithms and maybe with a reference architecture do you know extraction classification and so on but do we need to worry about the curation and the structure of the knowledge base and of course I have my philosophy on that and I think it's aligned with yours.
• 26:04
  SPEAKER_01: Would you want to talk a little bit about that?
• 26:06
  SPEAKER_01: Again, you know when people think about cognitive agents or cognitive assistance if you say chatbot or chatbot's a channel it's a channel since data content and knowledge right so we need to make make sure that's the right structure for us.
• 26:19
  SPEAKER_01: Do you want to talk a little bit about that?
• 26:20
  SPEAKER_04: Yeah I I will probably go back and do it where you're talking about pre-processing not just coming to the high level as you have but actually the importance of pre-processing data because that's it's really the key thing.
• 26:35
  SPEAKER_04: Usually when I say developing an AI uh system if you want to develop a chatbot 70% you spend on curating cleaning and structuring the data and then hopefully you can send it for the algorithm to do the first learning.
• 26:52
  SPEAKER_04: And if you haven't structured your data before you just send in whatever you have into the algorithm and then you come out and say it doesn't work.
• 27:00
  SPEAKER_04: Yeah it doesn't if you just send in a lot of unstructured and no no way for it to actually go through this collection and give it guidance what it should be looking for, you you won't get any good example out.
• 27:14
  SPEAKER_04: You won't get a shepherd that actually can predict uh understand what you want to how you want to for example I did uh for a company um which works a lot with allergies I did cooking recipes where um we have a lot of allergies in Sweden so 10% of us has some kind of food allergy.
• 27:34
  SPEAKER_04: So for us it's very common that okay I want to do this but I'm allergic to all of these what can I use instead?
• 27:41
  SPEAKER_04: So in in order for doing that you first need to collect a lot of recipes and see how they are different.
• 27:48
  SPEAKER_04: And then you need to collect information about all the ingredients and see how they can be substituted.
• 27:54
  SPEAKER_04: And if you're working in a country like Sweden where you have a lot of allergy food allergy you'll need also to know which ones are allergenic the ones that you usually have it's usually some kind of nuts, a white flower you all of these ones and what have people use instead so you then need to collect the uh information about what have people used instead when they were doing pancake if you are uh allergic to eggs a banana is something they are using okay so that is actually the knowledge that you need to start collecting and these needs to be structured so so actually you have different types of data and then if we go in in particular in this you need also to see how are these recipes structured by itself just how is it structured in the files and how can I extract when and where things are cooked and how it should be cooked and things like that.
• 28:52
  SPEAKER_04: So um I don't know if it was your answer but I this is a good example because everyone has been cooking at some point like yeah so that's a great it's a great example and and I think what you're don't you know what you're suggesting is that we're curating the the knowledge sources the relationships between these concepts in order for that type of an application to work correctly and as I have mentioned a lot of organizations are not really clear on this idea that they need to spend a lot of time and energy on the knowledge sources themselves on curating them and structuring them appropriately for that retrieval.
• 29:32
  SPEAKER_01: So I think you you certainly answered the question with a with a great example.
• 29:37
  SPEAKER_02: Okay.
• 29:38
  SPEAKER_01: Oh by the way before we uh before we jump back in I just want to take a moment and uh reintroduce uh reintroduce you Linda our guest is uh Linda Anderson who is the founder and CEO of Artificial Researcher which is a startup based in Vienna Austria uh and also want to uh thank again our our sponsors CMS Wire or Simpler Media it's a parent company uh the Marketing AI Institute and uh early information science so just wanted to take a moment to do that again.
• 30:07
  SPEAKER_03: Hey Linda let me ask you a question because we've got you know definitely because you sit in the in the crossroads of a very um competitive world of language and speech especially for AI right it's probably we'll say probably let's in the in the maturity model on the more mature of of AI related you know technologies.
• 30:29
  SPEAKER_03: And given you know the data classification and how important that is you you know you can use things like BERT models and stuff and there's been a lot of like enhancements to those open source models and things for you know bidirectional for NLP.
• 30:42
  SPEAKER_03: I would love to get a take from you and maybe this is is more centered around you know what the initial hypotheses of an organization is to set some type of an outcome but supervised versus semi-supervised models right one is a lot of data in the front I'm just gonna come up with what I need because in a language construct like we were talking about there's so much data there that we need in order to understand the context of let's say all the information there as opposed to a very selective supervised model that actually has a classifier up front, right?
• 31:17
  SPEAKER_03: I'd love to get a take for when and why one versus another or are they mutually exclusive or what does it look like in your world?
• 31:27
  SPEAKER_04: Oh so since I've worked with domain specific mostly so I work with pharmaceutical I work with um patent a lot in different areas of patent and I work a lot also with research papers.
• 31:40
  SPEAKER_04: So mining research paper, finding out have they data sets, where can I find the code they have used and all of these ones we mine every information out of our research article, not just for reading but actually what did they do?
• 31:56
  SPEAKER_04: Why did they do it what was the result what was the data can this be applicable on what I want to do as a PhD or as a master things like that.
• 32:06
  SPEAKER_04: So so if we if we then come back to how how do you what what model can you do using these ones?
• 32:14
  SPEAKER_04: I usually say you need to use both okay in beginning you start with some kind of rule-based more supervised machine learning and see if you can extract enough information to train a seed of a bird where you start doing semi training as a semi annuity is supervised.
• 32:36
  SPEAKER_04: And then there is a definition okay when it comes to pre-trained models so so for example when you train a distribution semantic models so a word for word to work or something just to learn the similarity between different words in a collection you actually don't learn the semantic similarity but rather the collection similarity as the attributional so in this patent if we only had a bus a computer patent that will be bus and memory register and so forth and no transportation vehicles and so forth.
• 33:14
  SPEAKER_04: So those ones where you just want to learn attributional semantics, those ones you can feed into Word to wake or uh word to doc uh document vector system, whatever you can feed them in and train them on finding similarity of this sort.
• 33:34
  SPEAKER_04: So you have for example Cybert and you have a PatentBird from Google.
• 33:38
  SPEAKER_04: Those are really great they are trained on sentences and they can do similarities on sentences, even phrases I noticed but then they go into a bit of a problem but those those are great ones.
• 33:50
  SPEAKER_04: But if you want to train it to identify a specific part of a sentence like you want to find a chemical compound that's something else.
• 33:59
  SPEAKER_04: Then you need to start okay with usually an ontology or some kind of list of a lot of chemical compounds and then start marking that out so that for example you can use bird to train okay start learning what a chemical compound is is in these sentences.
• 34:18
  SPEAKER_04: You use something that is called IO system so you mark out these are the these are the letter sequences that are chemical compounds and then you train it and then you let it predict further.
• 34:34
  SPEAKER_04: So you give the new data and ask it to predict and then you curate this one and at certain time it can actually start predicting quite well by itself but you need to feed in some kind of um manually annotated data or extract it from ontology but there need to be some manual start uh usually I call it seed you need to have a seed sentence if you're doing a classifier.
• 35:02
  SPEAKER_04: And yeah I've done uh some really good classifier with uh using BERT for example summarization what's important in a document or uh identifying technical terms or identifying data sets inside scientific articles so I can know oh who used my data set which I produce which is very important for a citation for a researcher to know oh everyone use my data set and then you have citation not only data citation but you also have the article citation and things like that.
• 35:33
  SPEAKER_04: So that's really important.
• 35:35
  SPEAKER_04: And so I usually say if I have enough data and I want to do target I start with supervised or semi-supervised trainer bird and then I let it go.
• 35:48
  SPEAKER_04: Okay.
• 35:49
  SPEAKER_04: If I'm doing some kind of just similarity computation and not sentence similarity but only word similarity I can more or less use a word to work very unsupervised method.
• 36:01
  SPEAKER_04: But if I want to learn it give it to learn specific semantic similarities so the difference between bus and bus, then I need to give it some hints.
• 36:12
  SPEAKER_04: I need to then model the data I feed into this system.
• 36:15
  SPEAKER_04: And that's also supervised because usually if you just take everything in there you don't know what you have what you're teaching the computer or the algorithm.
• 36:25
  SPEAKER_04: So if you then sample the data so you know you have a good sequence of bus, double wicker bus and bus the bus slot card, then you know it will learn both.
• 36:37
  SPEAKER_03: Good.
• 36:37
  SPEAKER_04: And be able to distinguish between it.
• 36:40
  SPEAKER_03: Well yeah for me you pass the test.
• 36:42
  SPEAKER_03: So for the for the interwebs out there in the world definitely employ Linda.
• 36:45
  SPEAKER_03: She'll help you get your stuff squared away and uh she knows what she's talking about in terms of making sure she you do you know she'll she'll give you the right type of perspective information and approach as opposed to you know no you only use this one model this one classifier this one thing and it's going to solve all your problems which is you know selling the mystic arts of what this all is so nice.
• 37:09
  SPEAKER_04: Yeah.
• 37:10
  SPEAKER_04: Sadly I tried that before in my master so I don't know those ones because I did one of the first master on um uh patent retrieval in the world and I did it on Swedish uh decompounding should you decompound Swedish word because we have very lengthy word and we put everything together and and I wanted to detect if we should decompound in order to have a better retrieval system.
• 37:36
  SPEAKER_04: And this was together with the Swedish patent office and I thought I have the perfect algorithm I will just and this was back in 2001 so I would just do a vector space models and I will use my algorithm which I developed in my bachelor and it will be great.
• 37:52
  SPEAKER_04: Yeah three years later said okay yeah I think I got it right now and I had to do all of this before I could actually get it running.
• 38:02
  SPEAKER_03: Yep.
• 38:05
  SPEAKER_03: I would love to get your take too Linda one one question then and then maybe maybe Seth if yours is a predecessor to this one but you know given out there in the world especially in consulting you know we run into these scenarios where it's an advisory like hey what where do we start what do we do where do we get going all that kind of stuff right and that's one that's one thing.
• 38:27
  SPEAKER_03: Where do you believe like in your mind I'd love to get your um your thoughts around where the let's say academia is missing in terms of educating people to get them started in the right direction because we we get you know by in in the professional we get all these downstream really bright and brilliant people however it's like they're unequipped to deal with these types of scenarios.
• 38:50
  SPEAKER_03: And even I mean it's it's crazy and sad that we even have to talk about ontologies given this but where do you believe like academia could do a better job?
• 39:00
  SPEAKER_04: I think it's very important to actually put forward the interdisciplinary because especially in a natural language processing you can come from different areas you can be computer scientist in and then you're expert on all the algorithm and the math or you can come from computation linguistic what I did I did both.
• 39:19
  SPEAKER_04: So I did computer science as well as a computation linguistics.
• 39:23
  SPEAKER_04: So for my master I worked with computation linguistics and then in my PhD I did actually my PhD at uh the technical university of Vienna so so that was computer science uh and that uh to to really understand that these uh these are not independent from each other you need to learn both what a noun is in order to actually be able to do text mining identifying technical terms that's because technical terms in English it's at least 80% uh nouns so that's and they are actually multi word multi-nouns so it's uh compound nouns so they're more than one word and this is things that you would need to be aware of so just because you speak a language write a language can read a language doesn't mean that you can do computation linguistic or text mining on that language even if you're a native speaker because you need to understand the structure of the language and that's that's where I'm very unfortunate because I well I am and was as stupid as a computer so I needed to get everything explicit to teach the computer the same thing the same thing I taught myself.
• 40:35
  SPEAKER_04: So this was really this was really something that was easy for me.
• 40:40
  SPEAKER_04: But I would say promote interdiscipline and also make sure that the students actually take this in that it's not just one solution that will solve the problem.
• 40:51
  SPEAKER_04: You would probably have to go first to the left and then to the right and then you have to do a roundabout because there is things that layers in language or in speech technology that you need to identify which you didn't think about if you're not taught this.
• 41:09
  SPEAKER_04: So I think broader more broadness and preferable also not just computer like you have you have math you have all this stem but also the uh the more like social science is also important not just only for for that it's important um discipline I've done linguistic journalism library and science and computer science and math and all of these I've done the entire with if I put it like that and so I I can understand what a computer science talks about and I can understand what a general linguistic talk about when he or she talks about topology.
• 41:49
  SPEAKER_01: So that is where I I would like to have a broadness um yeah that is what I would recommend yeah that's great and and uh as Chris you started to mention this uh about the getting started so I think you know getting the right talent uh in place but looking for broad interdisciplinary background uh in most organizations you know they're still getting their minds around these areas especially around cognitive AI and being able to uh build out you know virtual assistance and uh the ability to offload um uh you know call center uh tasks and and to apply uh allow uh customers to sell service when you look at the organizations that are trying to do these things where where would you advise them to start their process start their journey and what do they need to think about regardless of the um final or the end application and you know I look at search and and uh retrieval I look at the virtual assistance as just part of the continuum of search and retrieval so that's why you know I really appreciate your semantic search perspective in organizations today you know they they've I like to say that a lot of the applications for AI and machine learning and cognitive and semantics are are are due to to make are trying to make up for their past sins in poor information curation.
• 43:16
  SPEAKER_01: All right so you know if we did things right in the first place maybe we wouldn't have so many problems.
• 43:22
  SPEAKER_01: But talk a little bit about where you think organizations should be focusing their time and their effort at this point when they're early in the the journey.
• 43:30
  SPEAKER_04: So so if you're starting you know you have a lot of data you because you have collected it and you know that it's some kind of machine readable data.
• 43:38
  SPEAKER_04: Let's say that we we we are there and it's not based on OCR so it is actually somewhere machine readable.
• 43:46
  SPEAKER_04: And you know it was after the paradigm and because I need to point point this out in Swedish for example you have this um laut so you have A with two dots on or a circle and so forth.
• 43:59
  SPEAKER_04: And in beginning Of computer science time, we actually reduce them to only being A instead of the two day dots.
• 44:08
  SPEAKER_04: In best case scenarios, we put AE, for example, and we could actually know it.
• 44:13
  SPEAKER_04: But uh, so you need to actually start knowing what is the condition of your data, can you even use it?
• 44:21
  SPEAKER_01: That's the first point.
• 44:22
  SPEAKER_01: Great, great first point.
• 44:24
  SPEAKER_01: Yep, yeah.
• 44:25
  SPEAKER_01: That's the first thing.
• 44:26
  SPEAKER_04: Okay.
• 44:26
  SPEAKER_01: Where is it and is it usable?
• 44:28
  SPEAKER_01: Yeah, sorry.
• 44:29
  SPEAKER_04: Yeah.
• 44:29
  SPEAKER_04: And then you start, what is the vision that you want to use it for?
• 44:34
  SPEAKER_04: So I usually go for uh what is the condition of your data, and then I ask, what is your vision?
• 44:40
  SPEAKER_04: And I already know the condition of the data because we already find it out, and then I know the vision, and then I can see is it even possible to match it or not?
• 44:51
  SPEAKER_04: If it's not, I will tell you, okay, given the condition, this will take you two two years to curate.
• 44:57
  SPEAKER_04: I would rather say that let us take a smaller set and collect it from somewhere else and simulate it, and then we try to slowly uh do a curation of your data.
• 45:09
  SPEAKER_04: For example, right now, I think actually using um the uh distribution semantic models is really good for OCR correction because you can actually really do that very well, because they OCR has like OCR correction or spelling uh errors, they have of certain characteristics.
• 45:28
  SPEAKER_04: What you are doing in incorrect spelling.
• 45:31
  SPEAKER_04: OCR is special because you always know that E uh or I will be a problem, so it can become uh like a L or something, depending on what what type of OCR uh it was and what type of uh data it was.
• 45:45
  SPEAKER_04: So that's where you can do it, but you know that the context will be the same, so you can actually detect OCR and reduce that.
• 45:52
  SPEAKER_04: One of the problems we have in patent in patent collection is that we have millions of words, and there is not millions in old languages.
• 46:01
  SPEAKER_04: We have a few words that not milli a few, but we have a set of words, it's not that we have millions, but due to the poor OCR that is done on patent data, we have millions of words that actually are meaning uh you need to join together to find normalize.
• 46:20
  SPEAKER_04: And here is where distribution semantic can actually curate it.
• 46:23
  SPEAKER_04: So if it's such a problem, I would actually say, okay, you want to have a product in one year, let us take a smaller set and do a simulation where we have clean data.
• 46:34
  SPEAKER_04: And then we do this curation to increase the curator.
• 46:38
  SPEAKER_04: And here it comes into therefore, you need to do supervised machine learning in the beginning.
• 46:43
  SPEAKER_02: Right.
• 46:44
  SPEAKER_04: And when we have curated this data, we can introduce it and start doing semi-supervised to almost unsupervised.
• 46:51
  SPEAKER_01: So that's great.
• 46:54
  SPEAKER_01: We're right at the top, we're almost at the end of our time.
• 46:57
  SPEAKER_01: And uh, I think I summarize that in terms of knowing what your business objective is, uh, and understanding your data, right?
• 47:05
  SPEAKER_01: What's the state of the data?
• 47:06
  SPEAKER_01: Do we have the right data to the right structure, the right quality, quality, and uh and what is our objective?
• 47:12
  SPEAKER_01: Well, we are at the end of the time.
• 47:13
  SPEAKER_01: It's really been a pleasure talking to you Linda.
• 47:16
  SPEAKER_01: Um, I've had a lot of fun, Chris.
• 47:19
  SPEAKER_01: As usual, it's it's always a blast.
• 47:21
  SPEAKER_01: And uh I do want to remind uh folks that uh Linda is the founder and CEO of Artificial Researcher.
• 47:28
  SPEAKER_01: Um there's some information about her company and her services in the uh show notes.
• 47:35
  SPEAKER_01: And uh again, uh want to thank our sponsors, uh Simpler Media, uh CMSWAR and uh Marketing AI Institute, Early Information Science, and of course, Sharon for uh doing all the production work in the background.
• 47:48
  SPEAKER_01: And and again, Chris, thank you so much.
• 47:50
  SPEAKER_01: Of course, of course, it's been a pleasure.
• 47:53
  SPEAKER_03: Linda, thank you so much.
• 47:54
  SPEAKER_03: It's always a pleasure to you know meet really intelligent colleagues that are come on, let's say you're smarter than a damn computer for sure.
• 48:02
  SPEAKER_04: Thank you, thank you.
• 48:04
  SPEAKER_04: Yeah, and uh for our algorithm is definitely smarter than a computer.
• 48:11
  SPEAKER_01: All right, everyone.
• 48:12
  SPEAKER_01: Thank you again, and we'll see you next time.
• 48:14
  SPEAKER_01: Okay.
• 48:15
  SPEAKER_03: Okay, talk soon.
• 48:16
  SPEAKER_03: Thanks, guys.
• 48:21
  SPEAKER_00: Thanks for listening to the Early AI podcast with your host, Seth Early and Chris Featherstone.
• 48:27
  SPEAKER_00: We hope you enjoyed the program and took away some ideas that you can start to implement in your organization right now.
• 48:33
  SPEAKER_00: If you're stuck and need help, check out www.early.com for information about early information science statistics services.
• 48:42
  SPEAKER_00: And of course, check the show notes for links to any resources we mentioned today.
• 48:46
  SPEAKER_00: That's all for now.
• 48:47
  SPEAKER_00: See you next time.