WEBVTT 00:00.960 --> 00:02.400 Hello, everyone, and welcome back. 00:03.240 --> 00:09.180 So the new topic, which we are going to learn in this accent is a word M80, and it is one of the very 00:09.180 --> 00:14.640 important and crucial step, you know, any kind of natural language processing, delicate task. 00:15.000 --> 00:19.870 So what I'm reading is you can see like it has complete political slides. 00:20.310 --> 00:21.830 There's an LP field. 00:22.260 --> 00:24.630 So what exactly avoid embedding is? 00:26.480 --> 00:29.000 Let's see if we get to straight to the computer. 00:29.360 --> 00:35.810 And if you tell computer to match character by character, I mean computer can easily do it. 00:35.930 --> 00:39.520 And computer can do very much at a faster rate. 00:40.820 --> 00:47.360 But suppose if you're searching for, let's say, massee on a Google Web search and you'll get the results 00:47.360 --> 00:51.760 related to football, also, you will get results related to internal Loiselle. 00:52.130 --> 00:54.800 So why such things happen as a human? 00:54.890 --> 00:59.390 We know that this message is related to football lives that aren't always related to football. 00:59.870 --> 01:02.340 And this three towns are interconnected. 01:02.840 --> 01:09.470 But that doesn't signifies with a complete string matching because these two strings are completely 01:09.530 --> 01:09.920 different. 01:10.220 --> 01:16.820 So how computer can understand that these three things are related to each other, even if there is 01:16.820 --> 01:17.840 no magic between them. 01:18.220 --> 01:23.960 This football itself is a in terms of computer representation, it's a completely different thing compared 01:23.960 --> 01:29.330 to what Partovi, Masti and other Nilo represented in a computer. 01:29.810 --> 01:33.740 So that string representation is not sufficient. 01:34.280 --> 01:40.220 While comparing this type of things, but as we see in Google Web site, it is quite smart enough. 01:40.580 --> 01:46.190 And they're using internally this voyle imaging technique so that we get the idea that when you search 01:46.190 --> 01:49.280 for Masoli, you will get the results related to football. 01:49.340 --> 01:52.070 And and so let's take one more example. 01:52.490 --> 01:54.740 Let's say Happel is a tasty fruit. 01:55.190 --> 02:03.240 Now, in this particular sentence, how can Computer understand that apple is a fruit which can be eaten? 02:03.650 --> 02:05.430 But that is not an organization. 02:05.900 --> 02:12.050 So that is the kind of intelligence this void embedding will bring into our natural language processing. 02:12.240 --> 02:12.620 Martin. 02:13.780 --> 02:20.150 Well, the meeting is all about understanding about your tax, the semantic represent and exist between 02:20.360 --> 02:24.020 individual votes, and let's try to formally define it. 02:24.950 --> 02:31.220 So the word I'm reading is all about the representation of your word, which captures the meaning of 02:31.220 --> 02:31.930 immediate word. 02:32.270 --> 02:35.480 The semantic relationship exists between the different types. 02:35.510 --> 02:36.230 So what? 02:36.650 --> 02:40.540 And the same word which is being used in a different kind of contexts. 02:41.030 --> 02:47.120 And all this thing, we are going to get implemented by this word embedding technique, which is nothing 02:47.150 --> 02:50.150 but a numerical representation of your text. 02:50.330 --> 02:57.350 So once you convert this text into kind of numerical representation, you can have a comparison between 02:57.350 --> 02:57.520 them. 02:57.830 --> 03:01.070 You can define some sort of distance measure criteria. 03:01.670 --> 03:08.150 How closely these two words are related to each other or how far they're situated in a full gamut of 03:08.540 --> 03:09.770 Englis Dictionary. 03:10.610 --> 03:12.830 But why does what I'm reading is required? 03:13.250 --> 03:18.320 Now, if you observe that many of the machine learning algorithm are almost all machine learning and 03:18.380 --> 03:24.590 deep learning architecture, I am indeed just cannot process the text directly in that raw form. 03:24.830 --> 03:33.210 So my idea here is to convert this plane taxied to kind of raw numbers and once we convert it a raw 03:33.210 --> 03:39.240 numbers so we can apply a different machine learning algorithms like a classification variation and 03:39.300 --> 03:42.270 listening because they require such a kind of number as input. 03:42.500 --> 03:46.280 They just cannot understand text statically as the input and void. 03:46.310 --> 03:51.800 Embedding plays a very important, vital role for converting this vote to a kind of text. 03:53.010 --> 03:55.900 Now, what are the different types of war and embedding techniques? 03:55.940 --> 04:00.690 So at a broad level, you can divide this technique into two different media categories. 04:01.020 --> 04:04.660 So like a frequency base, embedding and undermine is up Predix and Baizhang. 04:05.250 --> 04:09.780 Now, our main focus here will be to learn about this prediction base embedding. 04:10.380 --> 04:17.130 And we won't go into much detail about the critical part, but mainly Vokes based on the neural network 04:17.550 --> 04:18.180 architecture. 04:18.540 --> 04:25.370 So one is a CB or W that will be a continuous Baggot Ford model and undermining this. 04:25.410 --> 04:26.280 Keep Cramerton. 04:26.850 --> 04:33.930 So here what we are trying to do based on the context of war or the word which is surrounded by some 04:33.930 --> 04:39.720 particular war, we are trying to build some neural network model and those neural network model will 04:39.720 --> 04:47.880 come up with a vector representation of your any of the words that will be kind of very best representation 04:47.880 --> 04:54.380 you can see compared to this frequency base embedding like a contractor, DFI of Corkins. 04:54.600 --> 05:00.100 So mostly both of this countercurrent mean contractors and T.F. idea. 05:00.550 --> 05:06.930 Both of them behave a little seen in our text classification projects where we are trying to find the 05:07.230 --> 05:12.600 same numerical representation based on the frequency that how many times some particular work. 05:12.790 --> 05:14.460 So that will be a contractor. 05:14.760 --> 05:19.260 How many times the combination of two or three or multiple words occur together. 05:19.560 --> 05:24.480 So that will be idea behind this co-occurrence vector and a little bit better. 05:24.480 --> 05:26.550 An advanced technique like a PEF idea. 05:26.640 --> 05:32.190 So that will count further how many times some particular Kaarina document. 05:32.760 --> 05:37.720 And another component as the IDF like inverse document frequency stack. 05:37.800 --> 05:41.520 How many times some particular volke across all the documents. 05:41.970 --> 05:49.470 So these are all basic frequency based embedding technique and under cites is prediction based modeling 05:49.470 --> 05:49.860 technique. 05:50.190 --> 05:52.560 Let me give you a little bit more detail. 05:52.620 --> 05:58.980 So let me go to the call up and then in the next video, we will see how to implement VITTA, one of 05:58.980 --> 06:03.770 the NLB library, Jency Predix and base embedding technique. 06:04.890 --> 06:08.730 So for better understanding purposes, I created this KALEV file. 06:09.300 --> 06:15.750 And there are three examples I have given, like how you can represent the text into a kind of number. 06:16.290 --> 06:17.990 So let's have one heart encoding. 06:18.270 --> 06:22.040 So that will be kind of bag of words kind of model. 06:22.560 --> 06:28.150 So that says that, let's say rehabber, just the one sentence like a dog gets sat on the mat. 06:28.740 --> 06:36.290 So totally velho how unique words are available like doe cat sat on hand. 06:37.110 --> 06:44.070 So wherever that particular word occurs, the next position of that particular word will be given the 06:44.070 --> 06:47.190 value one meaning all values will be zero. 06:47.760 --> 06:49.760 So that is Koller's of one order encoding. 06:50.100 --> 06:57.900 Now for the what the vector will be zero zero zero zero one because at this particular date only occupies 06:57.960 --> 06:58.920 one remaining. 06:58.920 --> 06:59.970 All values will be zero. 07:00.360 --> 07:06.580 But now this approach is quite efficient because whatever than one hardcoded vector you will get, that 07:06.600 --> 07:08.080 is a very sparse. 07:08.430 --> 07:14.520 So let's see if you can just imagine the total words in a dictionary like a 10000 word dictionary. 07:14.820 --> 07:20.370 So in every single vector you will get almost ninety nine point ninety nine percent element will be 07:20.370 --> 07:20.640 zero. 07:20.680 --> 07:23.700 Only one place there will be a value one. 07:24.180 --> 07:26.570 So this approach will be quite inefficient. 07:26.970 --> 07:27.820 Let's go. 07:27.820 --> 07:32.460 We done under outputs like a code, each word with some unique number. 07:32.850 --> 07:34.530 So let's take a same example. 07:34.530 --> 07:36.560 Let's say the cats act on a map. 07:37.040 --> 07:40.770 Now we will give one single number to every single voice. 07:41.340 --> 07:46.800 So that will be assigned to one and two will be assigned to set simply some numbers will be assigned 07:46.800 --> 07:49.620 to every single unique one. 07:50.010 --> 07:57.210 Now, this particular approach is even more efficient, but that is better compared to this sparse vector 07:57.300 --> 07:57.710 approach. 07:58.530 --> 08:04.520 Now here the main problem is that when you convert this type of things into this, this is as even a 08:04.530 --> 08:04.980 label. 08:04.990 --> 08:05.810 Bayes indexing. 08:05.940 --> 08:06.360 Also. 08:07.410 --> 08:12.270 Let's say no has been presented with one and this two has been represented. 08:12.870 --> 08:15.410 I mean, said has been represented by 10 and 12. 08:15.900 --> 08:21.120 Now, when we supply this thing to the machine learning algorithm, the algorithm will understand that 08:21.460 --> 08:22.730 set is better than that. 08:23.280 --> 08:24.660 But that is not a case, actually. 08:24.810 --> 08:30.390 That is not a case, because now once we convert it into two number, it should also make sense how 08:30.390 --> 08:33.170 something is not better than other in terms of text. 08:33.390 --> 08:37.410 But when we convert it to number of automate, it will be something like that. 08:37.470 --> 08:39.360 And that control makes. 08:40.480 --> 08:46.410 Well, I saw it last one, an important one is a word emerging so you can read about things here. 08:46.830 --> 08:54.330 And I just want to show you that compared to earlier to approach this word embedding approach will try 08:54.330 --> 08:57.730 to give us some fixed line rector. 08:58.710 --> 09:03.090 So let's say in our case, the vector land will be, let's say, for Davidson. 09:03.360 --> 09:09.360 So care has been represented in such a kind of four day image that Matt has been represented by this 09:09.360 --> 09:10.020 phone number. 09:10.080 --> 09:12.900 And on has been represented by this phone number. 09:13.140 --> 09:20.350 So this with a four dimensional embedding will create for every single voice or I would say unique words. 09:20.470 --> 09:20.700 Really? 09:20.700 --> 09:22.110 Well, you know, what a dictionary. 09:22.110 --> 09:27.390 And that is what the word Embling is and how we are going to create this thing based on the neural network. 09:27.840 --> 09:28.170 All right. 09:28.440 --> 09:29.940 So see you in the next video. 09:29.940 --> 09:37.380 We'll get my hands on practical tech, how we can create our own war and maybe based on some simple 09:37.650 --> 09:38.080 dataset. 09:38.670 --> 09:38.980 All right. 09:39.120 --> 09:40.440 See you in the next video.