Deep Learning is one of the most common buzzwords in tech nowadays but, what does it mean for your company?
Deep Learning has revolutionized AI in the last decade. It has made AI algorithms much more powerful, and today we all use it on a daily basis: in search engines, in recommendations of your movie or music app, speech recognition (Alexa, Siri, etc), machine translation (Google Translate), and now in Talent Acquisition.
In this webinar, Vincent Slot will explain how we teach computers to understand human language through deep neural networks. We will explain it in Layman’s terms using real-life examples from the TA domain.
This webinar is designed for Talent Acquisition managers and directors who have an interest in AI and want to understand its practical applications.
No prior knowledge required.
Watch the recording now!
Webinar | Recorded on 10 June | 30 Minutes
Don't want to watch the entire webinar?
Vincent: Welcome, everyone, to this webinar on AI for talent acquisition, A Deep Dive into Deep Learning. Thank you very much for joining this webinar. Today’s topic will be deep learning, and how it can benefit talent acquisition. It’s a very exciting topic. It’s a very powerful tool. And it has many, many applications within or outside talent acquisition. It’s also a very complex topic, so I’m going to explain it on a very high level so that everyone in the audience, including the non-technical viewers will understand what deep learning is about, how it works, and how it can benefit a talent acquisition.
My name is Vincent Slot. I lead the Search R&D team at Textkernel. Basically, the Search R&D team is responsible for the quality of your searching and matching results when using the Textkernel search and match product. My background is in AI. I have a master’s degree from the University of Groningen. And I’ve been at Textkernel for about 2 and a half years.
Textkernel is an Amsterdam-based company. It started out as an all-purpose R&D lab in 2001. But over the years grew into an international leader in AI in the domain of HR and recruitment and talent acquisition. We do R&D for all kinds of applications within a recruiter or within the talent acquisition workflow, basically starting from CV parsing to searching and matching candidates, and everything in between.
So, today’s agenda will look like this. First, I will start with giving an introduction on AI for talent acquisition. We’ve had a previous AI webinar by Textkernel, which covered the topic what is AI, and how can it benefit talent acquisition? I’m going to summarize that as basic knowledge. So, it’s not required to have attended that webinar. I will start from the basics, start from scratch.
Today’s topic is deep learning and how that works and how we can benefit from that. So, that’s the second section. And finally, we will look at some applications of deep learning for talent acquisition. At the end, there will be room for questions. Zoom has a QA button. And you can use that to answer questions at the end of the webinar, or during and we’ll answer them at the end.
So, let’s start with the introduction. Well, actually, let’s start with the most basic question, what is AI? What exactly is that? AI is a very broad definition. In a very, very broad definition, it’s any program, any system that can reason or act. So, that even includes a simple rule-based systems even that just have a rule to make a certain decision or something like that. So, that’s not very intelligent, but it exhibits some kind of basic form of intelligent behavior.
Machine learning is actually when we talk about AI in this modern digital age, we’re actually talking about machine learning, which is a subset of these intelligent systems. Let’s say what defines machine learning is that these decision rules are not programmed by the developer, but they are actually learned from data. And then a more recent approach to machine learning is deep learning. And that’s even more powerful where the algorithms learn from vast amounts of data and find all kinds of hidden patterns.
Now, this will probably sound very abstract. So, we’re going to look at some examples, and we’re going to dive a bit deeper into this to make it more clear. So, this is a typical 3-step machine learning process, data collection, model training, and prediction. If we want to use machine learning, we need data. So, we’re going to collect, in this example, pictures of cats and pictures of dogs. And we label them accordingly. So, we give this data to an algorithm. And that algorithm can learn what a cat looks like, what a dog looks like. So, once it’s done training, we can use this in actual systems, and we can show it a picture, in this case, of a cat. It has learned from the data that a cat has pointy ears and whiskers, and can give us a prediction that that picture that we’re showing is likely a cat.
Now, in the HR domain, we’re not generally interested in cats and dogs, but rather in people, so our data will more likely be CVs. In this example, we’re looking at machine learning for CV parsing. So, our training data will be annotated CVs. These the CVs, a human has actually annotated them. Like, this is date of birth, this is work experience, and these are some skills and so on.
So, if we give all this data, this can be thousands of documents to a machine learning model, that model will learn that this input sentence is most likely a date, a job, and a company, because it has seen from the training data that that is generally what this sentence means. So, this is very briefly, machine learning. And in this example, how we can use that for CV parsing.
There are many, many more applications within HR and TA. For example, for mining knowledge, finding a related terms, that’s one example that we’ll dive deeper into. Later on in this webinar, we can find job categories with machine learning. So, if we, for example, need a system that understands that a Java developer is a programmer, but a PHP developer’s also a programmer, we can use machine learning to infer those kinds of relations.
We can use machine learning for analytics. I’m not going into every example. But this is just to show that there are many, many applications where machine learning fits in the talent acquisition domain. We can use it to make semantic search. So, a search engine that actually understands what you’re searching for, that can interpret the language as you type it. We can use machine learning for automatic matching of people in jobs. So, you can get relevant candidates for a job with just one click.
So, this allows for many use cases within the domain of HR. For example, your internal mobility, you can shortlist applicants. You can even things like upskilling, you can find out which skills your employees need to learn in order to do their job most efficiently. So, again, I’m not going into detail on these use cases. But I just want to illustrate that this is a very powerful technique that has many applications within the HR domain.
So, today’s topic is deep learning. Deep learning, as I said before, is basically a category of machine learning algorithms. But instead of telling these types of algorithms, “These are cats. These are dogs,” we just give it data, and it can ingest that data and find hidden patterns in there.
So, in this example, in this webinar, we are going to look at how deep learning can help us understand human language. That’s the area of AI research we call natural language processing. It’s basically, yeah, having computers understand natural language.
So, if we want to teach an algorithm a language, then we have to ask ourselves the question, what gives a word meaning? What does the word mean? If I were to teach what a car, what the word ‘car’ means to a child, I would simply point outside my window at a car and say, “That’s a car.” Through the years, this kid will learn that a car is something we commute with, that requires fuel to drive, that you need a driving license for, and all those kinds of basic things about a car that’s implicit in the way the world works. However, the problem is these algorithms, they don’t have embodiment in the real world. They don’t live in the real world. They just have the data.
So, if we look at these 4 sentences, can we say something about the meaning of words by just looking at the text? And the answer is, yes, we can, fortunately. So, let’s go through the very simple exercise of just counting which words co-occur in a sentence. Then we can draw this table where each word is represented by a row of numbers. And this row of numbers tells us something about what that word means. Because, for example, ‘driver’s license’ is related to the word ‘car’, but not to any of the other words. ‘Cat’ and ‘kitten’ are related to each other, but not to any of the other words.
So, actually, this table can tell us something about what the words mean. And it’s just based on these 4 sentences of training data. The nice property of representing the words as these rows of numbers is that the rows that look similar, are actually similar in meaning. So, for example, the rows, if we look at these rows for cat and for kitten, these rows look very similar because they appear in similar contexts as each other, which implies that they have a shared meaning or they belong together, they are similar.
Now, if we were to extend this table to contain input data from, let’s say, tens of thousands of documents, or hundreds of thousands of documents, and it can contain tens of thousands of words, well, that’s just not very useful. Then this matrix will explode, and yeah, that’s hardly usable.
So, what deep learning does, that’s actually constructing a numerical representation for these words, that says something about the meaning of those words, but not necessarily includes all words as an explicit dimension of these words. Again, it sounds very abstract, so let’s look at an example.
This is, for example, the word ‘English’ appearing in a number of HR documents. It can be CVs, can be vacancies. But these are typical contexts for the word ‘English’ that the language skill English to appear in.
Now, interestingly enough, if we replace the word ‘English’ by ‘Dutch’ or ‘French’ or ‘German’ or any other language skill, then these contexts, these sentences are still valid. So, if we just ignore the actual word and only look at the context, we can actually say that, at these spots where the Xs are, there belongs a language skill, most likely.
So, if we can represent this in numbers, and we can have words that appear in similar contexts, close by, nearby each other in this numerical representation, and we can say something about which words are similar or not.
So, let’s have an example where we represent each word as just 2 numbers so we can show them in this graph. Then what we want is that these skills, we’ve just shown that these skills are very likely to appear in the same context. The same words are around it. So, they should be close together in this semantic space. However, Dutch cities, they appear in very different contexts as language skills, but probably appear among themselves in similar context. We could say the same about these programming languages. This model is a very powerful representation of what words mean. That’s what deep learning for natural language excels at. So, it can really make rich semantic representations of language.
So, now let’s make it a bit more concrete. This last example was, of course, an HR related example. But if we extend that a bit further, we can look at this skill cloud. This is a project done by one of my colleagues for Textkernel’s innovation week. And in this cloud, every dot represents a skill from our skill database. So, we’ve used deep learning to come up with, let’s say, the meaning of those skills, which skills are related to each other? Which skills are similar? If we look at the skill, WordPress, for example, we can see that it is related to content management systems, creative writing, publishing articles. So, that’s because these words these terms that they appear in similar contexts. Right? We get imagine that WordPress and content management systems might appear in the same vacancies, for example. So, this space is a very powerful way to find, for example, related skills.
Another example is artificial intelligence. If we look at what’s related to that, we see deep learning, natural language processing, exactly today’s topic, so that’s right. And these skills are all related. And this can be very useful if we want, for example, to find the right candidate for a job. So, if we want to find someone in the domain of AI, maybe we might consider searching for deep learning, natural language processing as well, because well, apparently, they are related terms. And someone might have written either of these terms in their CV instead of artificial intelligence. So, this can help you find the right skills to look for in your candidate pool.
So, this is one example of how this would look. For example, this Java Hadoop cloud is just a 3-word query. But if we look at all the related terms that come with those words, then that allows us to get much more relevant results, because we also search for all those skills that are very likely to appear in the same context. So, they may appear in the same documents.
Another very interesting example of these kinds of representations these embeddings, as they’re called, these vectors that we that we let represent knowledge about words. Another example is the deep learning matcher. That’s a very interesting project that that actually is a bit different from the applications that we just looked at. Because in this case, we don’t represent skills or words in this semantic space, but we actually embed entire documents. So, each.in This picture is not just a word or skill, but it’s an entire vacancy or CV.
And that has interesting properties, because that means that vacancies that are similar in meaning are close together in this space. Same goes for CVs, or candidates. Candidates that are similar can occur together in this 3-dimensional space in this cloud.
But more interesting, also, candidates which are a good match, valid match for a given job will appear close in this space. So, if we look at if we want to find relevant candidates for a given job, we can just simply look around it in this vector space, and we can see which candidates are nearby. So, here we have a cluster, we’re looking at a cluster of financial managers. These are jobs and CVs mixed. But it means that these, because they are close together in this space, means that they have a similar meaning. This is a representation that shows a similarity between documents. That can be very useful if we want, for example, to recommend similar vacancies to a to a user, or if we want to recommend similar candidates to a user. Or if we want to find relevant candidates for a given vacancy, or vice versa.
So, to summarize, deep learning algorithms, the core of what I want to convey, the message that I want to convey is that deep learning algorithms have a numerical representation of words that capture the semantics, the meaning of that word. And that’s based on the context that this word appears. This can help us with all kinds of processes and help us discover all kinds of patterns in the data. And we’ve looked at some examples like finding similar skills, finding related documents, or actually doing matching on people and jobs.
Yeah, so we’re just looking at the tip of the iceberg here. So, there are many more, many, many more applications. And that’s, I think, what I want to show today to the people who are not that invested in deep learning yet, that it’s actually a very powerful tool and can help you understand your data much better.
Right, that’s what I wanted to say. I’d like to draw your attention to more upcoming webinars that may be of interest. One is about matching career portal visitors, and one about skill-based analytics. You can visit our website at textkernel.com If you want to find more information about this, or if you want to join these webinars.
Alright, I will leave a couple of minutes to answer some questions.
Question: Sure, I got some questions here for you. Can you hear me?
Vincent: Yeah, I can hear you.
Question: Good. So, one of them, it’s, “In what ways can I use deep learning to gain insights in the skills present in my candidate data or talent pool?”
Vincent: Okay. Yeah. So, as I’ve shown, we can see what skills are related. So, if you know which skills are prominent in your talent pool, and you know which skills are prominent in, let’s say, the vacancies that either your company has or that appear around the web, then this can give an interesting insight, because we can learn what skills are related to satisfy certain criteria. So, let’s take the example, of artificial intelligence again. So, if you know that your people are mentioning a lot of times artificial intelligence on their profile, then… but maybe in reality, in the vacancies, where people are actually looking for deep learning specialists, and not artificial intelligence, then deep learning can tell you that these concepts are very much related.
So, you’re actually looking at similar skills, but maybe that means that your employees or your candidates need to be upskilled and learn about deep learning in this in this example. So, yeah, that’s one of the ways that it can give more insight into how skills relate and how they kind of, yeah, co-exist.
Question: Another question here. it’s, “Can you name other additional examples to deep learning in the talent acquisition?”
Vincent: Yeah, yeah, I can. So, for example, I can name one example of a problem that we’re solving with deep learning, is actually that we’re looking at, for example, how can we recognize when terms are important in a search query? So, if you type a number of words as a search query, then each of them will have a similar weight in finding the results. However, we can use deep learning to find out based on a lot of data, what are important skills and what are less important skills for finding the good matches? So, we can improve the quality of your search results using deep learning.
Question: One other question here is, “Can I use deep learning for analytics?”
Vincent: Yes, yeah. That’s also a good question. For analytics, for example, one of the ways that deep learning can benefit analytics is through better normalization of whatever data you have. So, if you want to, for example, know, how many skills, how many times a skill occurs in your database, you need a normalized version of that skill. Because if you have a skill can be written in 100 different ways, probably. And they all mean the same thing. So, deep learning can also help you find the things that this… yeah, the ways to describe a skill that is different, that looks different on the surface, but actually means the same thing. So, in that sense, you can do better analytics, if you have good recognition of what the skills mean, and if they are maybe the same thing, or if they are different, if 2 things are different.
Question: Very good. Due to the time, this was the last question here to be answered. However, we do have the other questions here so we can answer to them via email.
Vincent: Yeah, sure. Yeah, I can answer some questions via email. If a question pops up after the webinar, don’t hesitate to write me an email. My email addresses here on the slide. I’d be happy to answer more questions or explain more, if that would be helpful. Thank you.
About Vincent Slot
Vincent graduated from the University of Groningen with a master’s degree in the field of Artificial Intelligence, focusing on Information Retrieval. Specialized in solving matching and ranking problems in search engines, he has been working for Textkernel since 2017. Vincent currently leads the Search R&D team and is in charge of quality and innovation for Textkernel’s matching technology.