This lecture is about the methods for text categorization. So in this lecture we're going to discuss how to do text for categorization. First, there're many methods for text categorization. In such a method the idea is to determine the category based on some rules that we design carefully to reflect the domain knowledge about the category prediction problem. So for example, if you want to do topic categorization for news articles you can say well, if the news article mentions word like a game and sports three times. That we're going to say it's about sports things like that and this would allow us to deterministically decide which category a document that should be put into. Now such a strategy would work well if the following conditions hold. First the categories must be very well defined and this allows the person to clearly decide the category based on some clear rules. A certainly the categories as half to be easy to distinguished at the based on a surface features in text. So that means some official features like keywords or punctuations or whatever, you can easily identify in text to data. For example, if there is some special vocabulary that is known to only occur in a particular category. And that would be most effective because we can easily use such a vocabulary or padding of such a vocabulary to recognize this category. Now we also should have sufficient knowledge for designing these words, and so if that's the case then such a can be effective. And so it does have a in some domains and sometimes. However, in general, there are several problems with this approach. First off, because it's label intensive it requires a lot of manual work. Obviously, we can't do this for all kinds of categorization problems. We have to do it from scratch for a different problem. problem because given the rules, what they need. So it doesn't scale up well. Secondly, it cannot handle uncertainty in rules, often the rules Aren't 100% reliable. Take for example looking at occurrences of words in texts and trying to decide the topic. It's actually very hard to have 100% correct rule. So for example you can say well, if it has game, sports, basketball Then for sure it's about sports. But one can also imagine some types of articles that mention these cures, but may not be exactly about sports or only marginally touching sports. The main topic could be another topic, a different topic than sports. So that's one disadvantage of this approach. And then finally, the rules maybe inconsistent and this would lead to robustness. More specifically, and sometimes, the results of categorization may be different that depending on which rule to be applied. So as in that case that you are facing uncertainty. And you will also have to decide an order of applying the rules, or combination of results that are contradictory. So all these are problems with this approach. And it turns out that both problems can be solved or alleviated by using machine learning. So these machine learning methods are more automatic. But, I still put automatic in quotation marks because they are not really completely automatic cause it still require many work. More specifically we have to use a human experts to help in two ways. First the human experts must annotate data cells was category labels. And would tell the computer which documents should receive which categories. And this is called training data. And then secondly, the human experts also need to provide a set of features to represent each text object. That can potentially provide a clue about the category. So, we need to provide some basic features for the computers to look into. In the case of tax a natural choice would be the words. So, using each has a feature is a very common choice to start with, but of course there are other sophisticated features like phrases or even parts of ancients tags or even syntax to the structures. So once human experts can provide this then we can use machine running to learn soft rules for categorization from the training data. So, soft rules just means, we're going to get decided which category we should be assigned for a document, but it's not going to be use using a rule that is deterministic. So we might use something similar to saying that if it matches games, sports many times, it's likely to be sports. But, we're not going to say exactly for sure but instead, we're going to use probabilities or weights. So that we can combine much more evidences. So, the learning process, basically is going to figure out which features are most useful for separating different categories. And it's going to also figure out how to optimally combine features to minimize errors of the categorization of the training data. So the training data, as you can see here, is very important. It's the basis for learning. And then, the trained classifier can be applied to a new text object to predict the most likely category. And that's to simulate the prediction of what human Would assign to this text object. If the human were to make a judgement. So when we use machine learning for text categorization we can also talk about the problem in the general setting of supervisement. So the set up is to learn a classifier to map a value of X. Into a map of Y so here X is all the text objects and Y is all the categories, a set of categories. So the class phi will take any value in x as input and would generate a value in y as output. We hope that output y with this right category for x. And here correct, of course, is judged based on the training data. So that's a general goal in machine learning problems or supervised learning problems where you are given some examples of input and output for a function. And then the computer's going to figure out the, how the function behaves like based on this examples. And then try to be able to compute the values for future x's that when we have not seen. So in general all methods would rely on discriminative features of text objects to distinguish different categories. So that's why these features are very important and they have to be provided by humans. And they will also combine multiple features in a weight map with weights to be optimized to minimize errors on the training data. So after the learning processes optimization problem. An objective function is often tied into the errors on the training data. Different methods tend to vary in their ways of measuring the errors on the training data. They might optimize a different objective function, which is often also called a loss function or cost function. They also tend to vary in their ways of combining the features. So a linear combination for example is simple, is often used. But they are not as powerful as nonlinear combinations. But nonlinear models might be more complex for training, so there are tradeoffs as well. But that would lead to different variations of many variations of these learning methods. So in general we can distinguish two kinds of classifiers at a high level. One is called generative classifiers. The other is called discriminative classifiers. The generative classifiers try to learn what the data looks like in each category. So it attempts to model the joint distribution of the data and the label x and y and this can then be factored out to a product of why the distribution of labels. And the joint probability of sorry the conditional probability of X given Y, so it's Y. So we first model the distribution of labels and then we model how the data is generate a particular label here. And once we can estimate these models, then we can compute this conditional probability of label given data based on the probability of data given label. And the label distribution here by using the Bayes Rule. Now this is the most important thing, because this conditional probability of the label can then be used directly to decide which label is most likely. So in such approaches objective function is actually likelihood. And so, we model how the data are generated. So it only indirectly captures the training errors. But if we can model the data in each category accurately, then we can also classify accurately. One example is NaÃ¯ve Bayes classifier, in this case. The other kind of approaches are called discriminative classifies, and these classifies try to learn what features separate categories. So they direct or attack the problem of categorization for separation of classes. So sorry for the problem. So, these discriminative classifiers attempt to model the conditional probability of the label given the data point directly. So, the objective function tends to directly measure the errors of categorization on the training data. Some examples include a logistical regression, support vector machines, and k-nearest neighbors. We will cover some of these classifiers in detail in the next few lectures. [MUSIC]