Hello everyone. Welcome to Big Data and Language. This is the last lecture of the first week. Today, we will talk about the recent research. So you might be curious about, okay, so now you understand the importance of the big data and relationship with big data and Language, but what kind of research going on? You might be curious about this. So the final chapter is about recent research. So are you ready? Let's get started. Here is the first research-based on big data to understand language or vacillate your language understanding. The first one as Machine Translation. You may not understand clearly what machine translation. So I will give you the image. So for example, if you do not understand the sign because you don't know that certain language, then if you use the application or in machine translation, then your machine such as your cell phone, or your iPad, like translated that certain language to your own language or language that you can understand. So let me give you the video clip. Though very short video clip as the example of machine translation. So here is the examples, tienda cerrada, which means store are closed. So even though you don't understand the Spanish, but you understand the English. Then you can't understand that's sign based on the machine translation. All right. So this is a very interesting example, but you may already know or use that application already. Let me give you another example of machine translation. Have you ever wondered how Facebook can make this? So you may notice that, you will see, see translation. So even though you don't understand Korean for example, Korean language, but if you click "See translation", then automatically that language or that sentences, words are translated in English, so you understand the gist of that expression written in Korean on Facebook. So you might wonder how it worked. So let's look at one by one. So in 2018, machine translation from Facebook Engineering Team, actually develop the machine translation. They focus on large monolingual corpora in each language. Corpora means, the Big Data Group. So for example, one Big Data Group Set, we call it corpus but more than one, the plural is corpora. I will explain about more about the corpus later in the future lecture. So they focus, they created and they used the corpora which means the multiple data sets and in each language. So the data set of Korean and data set of English, data set of Spanish, something like that. Using several methods for optimal result. So for example, here is the example that I cannot read or I cannot understand this sentence because I don't know this language, however the Facebook actually try to multiple different methods in order to find the optimal result. So for example, the machine actually translated those sentence word by word level. But still we cannot understand quite fully, and some words are not interpreted in English. So we can use like the phrasal-based machine translation. Then it's better to understand. But still it's not really quite. So the Machine Translation Team on Facebook, they combine these two different methods. Then finally, we can get the pretty meaningful and understandable sentence in English. So here is though one example that we can use or we already used machine translation. It's pretty interesting and you're probably now notice that, oh, okay, I already use machine translation. So it's not really nothing new in our really daily life. The other research based on big data to understand the language is question answering. So what is it? Let me explain one by one. So if you have any question online, then the question interface actually classified that question. The query for me after the query formation based on the data set, that answer is extracted. They finally display. The computer display the answers. So if you have any questions on, for example, Bank of America, then you may asks that questions on online. Then not only just a human being's computer AI, they also analyze the data and analyze the answer, and based on the data, they actually answer, full answer your question accordingly or in a right way based on the data. How this one work? Let's explain. Let me explain briefly. So answering questions which requires reasoning over multiple effects. This one is the how it work of answering question answering. So reducing the original query to a more informed query through layers. So let me give you an example table. This table, we can see that the all the red marks. The darker one is the more accurate information. So based on that, the computer actually can answer the question that human being or even like other a computer asks. The third technology is caption generation. So what is it? Caption generation means, if there is any picture, then AI or computer automatically produced the sentence level description. For example, the first picture, you can see that like man with a guitar. So the computer can produce the sentence as, men in black t-shirt is playing guitar. So not really grammatically perfect, however you can understand and it's pretty accurate to describe this picture. Let me give you one more example. Though the second one is, we can see the beautiful nice weather and nice weather on the nice weather. We see the little girl is jumping. So the computer actually describes this picture as girl in pink dress is jumping in the air. So it's pretty accurate. How would this caption generation work? Using this study, actually using data set of 82K images for training and 42K for testing. So what that mean is, this research team collect the huge amount of images and training, for training computers or training AI. Then, they also get 40 another, the little bit smaller set of pictures and they test whether the computer understands the caption accurately or not. So achieving results comparable to human evaluation, that's how they access. So for example, they have this pictures and one the same picture, the AI see that picture and at the same time, human being also see that picture and the both, they are asked to describe that picture and then compare in order to check that how AI is accurately describing that picture. But in this moment, it's not really quite perfect. So let me give you an example. So usually, it's pretty accurate. However, look at the first example. There are three dogs, right? They're running. But the AI only recognized two dogs only. So that's why computer describe this picture as two dogs play in the grass, which is not correct. We can see three dogs instead of two. So those kind of mistake is actually taken. Also, you can see another example. There is just a one hockey player, but the computer describes that the two hockey players are fighting over the puck, which one is not really correct. So how we can improve this capturing generation technology? We actually using more crowdsource which means like getting more dataset from human and also training the computer. So at the end as a result, the computer can produce more accurate description. Now, let's move to another technology which is text classification. We have bunch of articles, then the computer automatically categorize all the texts based on different genres. For example, if we have a bunch of news articles then the AI or computer, they can actually categorize all the texts data as technology, sports entertainment, something like that. So how it works? Let me give you one example of the research. KAIST Data Science Lab of Professor Meeyoung Cha actually detect dissatisfied customers based on text data of Samsung's live chat services, which means like if the customer has any complaints and they can call or they can go to online and talking about their complaint and then the online chat based on AI, actually they react those kind of complaints. So based on their research, it show that the LSTM-time and the GRU-time, those are the most accurate which means this one is all the new unfamiliar terminology. So let me explain one by one. RNN, LSTM, and GRU means the types of deep learning natural network. So this research group, the research team actually compare and check all the different types of deep learning neural network in terms of accuracy, precision, and recall. However, not only just these kind of types of deep learning neural network. Also the time is another factor which refers to taking time gap between real reply into account. So they can capture those kind of these two types and time which is LSTM-time and GRU-time, those two captured and memorize to look at far distant past in a chat conversation compared to RNN. So this research team actually found that what would be the most efficient and accurate chat system or data compared to other networking, other deep learning neural network. So how we know that these two are the most efficient or most accurate? Because like all the values in this table, it shows that accuracy, precision, and the recall values of these two are higher than any other vectors or any other independent variables. So now, let's move to the speech recognition which is another recent research about the big data and language. So speech recognition means we already used, right? So if you are the iPhone user and as I mentioned already the previous lecture, if you ask Siri like asking, "Where will be the good Korean restaurant?" Then maybe Siri collect, first of all, analyzing and understand your question and then giving the answers. So how it works? As I mentioned auditing signal, once the computer accepts the auditing signal and that signal actually analyzed, and then that one is produced as a language level as well and then the translation. So there are several steps that the computer follow for speech recognition. I will give you the good features. So you probably understand better based on this feature and I'll give you the resource information as well. So yeah, and also, let me give you an example. Are you familiar with this device? This is Amazon Alexa. This one is automatic speech recognition. This one use the automatic speech recognition technology. We call it ASR as well, stands for automatic speech recognition. This device detects the spoken sounds and recognize them as words. So basically, you can communicate with this device. It's pretty wonderful and amazing. So far, we've talked about the recent research of big data and language; machine translation, question answering, caption generation, text classification, speech recognition. So not only this, there are more different types of technology already for big data and language. This technology is not really perfect yet, however dramatically actually improved a lot based on the previous remarkable studies. So there are many more research, many more studies in NLP such as like language modeling, named-entity recognition, but I will stop here and we will talk about more later. So today, we've talked about recent research and the next which is week 2, we will focus in learning about spoken and written data. Thank you for your attention.
