WEBVTT 00:01.510 --> 00:02.080 Hey, everyone. 00:02.110 --> 00:06.790 So let's continue our discussion on training your own and meeting Martin. 00:07.240 --> 00:11.410 So in the last video, video created this model and we put product training purpose. 00:11.770 --> 00:13.780 Now that training is completed. 00:14.320 --> 00:17.100 Let me display this model logic. 00:21.480 --> 00:21.900 All right. 00:21.930 --> 00:25.180 So we got our model object in terms of work to work class. 00:25.530 --> 00:31.650 So our model is really we can do the production and we can see some of the applications of this over 00:31.650 --> 00:32.320 to Emory. 00:33.090 --> 00:36.720 So the first one is we should execute this one first. 00:37.080 --> 00:44.330 So model dot v w v function we are going to use and that function will return witness that particular 00:44.330 --> 00:47.850 of how it is represented or what is its word vector. 00:48.390 --> 00:49.740 So let me have Randy. 00:52.360 --> 01:00.570 All right, so sort of a token hammy and men has been represented by this 32 numbers simply. 01:01.060 --> 01:08.200 It has on their applications like a most similar so similar to men, which are the other 10 words, 01:08.260 --> 01:09.790 which is very much Close's. 01:10.030 --> 01:16.780 Now, not in terms of the actual characters contained inside those token, but in terms of the semantic 01:16.780 --> 01:22.800 representation that man is related to, which are the 10 closest one. 01:23.020 --> 01:24.520 So let me executing. 01:27.260 --> 01:33.190 And you can see Roumain is quite close to man, couple, girl, boy. 01:33.500 --> 01:34.730 So we got all those. 01:35.870 --> 01:41.210 Which is quite similar in nature compared to what they were worthy of supply man. 01:41.540 --> 01:47.660 So now you can understand the power of this word to work or avoid inbreeding, kind of modern, which 01:47.660 --> 01:49.500 is a prediction based model. 01:49.970 --> 01:56.330 Now, there is no similarity between, let's say bye and a man, but genetically try to indicate that 01:57.140 --> 01:59.090 these are the similar voice. 01:59.300 --> 02:03.560 So this thing just cannot be possible with a simple string matching. 02:04.040 --> 02:06.730 But this is possible with Delp of this war, Emily. 02:07.130 --> 02:10.550 So what inbreeding has given us power to represent? 02:11.750 --> 02:14.300 Every single token in a hide Amundsen's space. 02:14.390 --> 02:17.270 And the results of fixed type of Hyrum, in essence, space. 02:17.830 --> 02:21.080 Now let's see one more application and you will be amazed to see about. 02:23.720 --> 02:24.110 Let's see. 02:24.560 --> 02:30.350 We'll try to do some simple vector arithmetic like I designed in subsection, so I'm just going to stop 02:31.040 --> 02:35.960 the word vector of this king minus man, and I'm just adding to the women. 02:36.720 --> 02:38.920 And let's see what results we get. 02:41.560 --> 02:44.790 So Volman is close to movement itself. 02:46.380 --> 02:52.770 Because king and man will be canceled each other, then we are union, education, religious. 02:53.070 --> 03:00.600 So in this case, we didn't get a very good result because we are expecting no king minus man plus woman 03:00.630 --> 03:07.620 should be like a queen because King will cancel out man the same way women will cancel out the queen. 03:07.890 --> 03:11.310 But now why this thing happened because we're limited by the data. 03:11.670 --> 03:13.740 Of course we are limited by the data. 03:14.280 --> 03:19.440 Let's see one more like a Berlin is the capital of Germany as a country. 03:19.920 --> 03:21.690 Then Paris is the capital of France. 03:21.720 --> 03:23.830 So let's see whether we'll get to France or not. 03:26.990 --> 03:27.350 All right. 03:27.380 --> 03:29.550 So in this case, we got a very good result. 03:30.080 --> 03:38.090 So if we subtract the word vector of Germany minus Berlin, plus Paris will be equal to France. 03:38.330 --> 03:41.030 So we got excellent result in this particular case. 03:41.330 --> 03:47.450 Now, you can imagine that why in this case, it works fine, because we're taking the data that is 03:47.450 --> 03:48.740 avoid news data. 03:49.220 --> 03:51.230 You can see why news data from the rally. 03:51.520 --> 03:58.000 So was probably in those dataset, we found the words like Germany, Berlin together. 03:58.010 --> 04:00.140 Same with Paris and France together. 04:00.440 --> 04:06.740 That's why those kind of semantic really and see, however, vote to work model embedding technique, 04:06.950 --> 04:09.440 it try to cattery very much easy. 04:10.040 --> 04:12.570 Let's see about the Massee minus football. 04:12.620 --> 04:16.160 So same Leiker Massie's the guard of the football. 04:16.490 --> 04:20.000 Same way we are trying to find voids the guard of the cricket. 04:20.810 --> 04:21.260 So. 04:22.200 --> 04:28.080 Mostly district tournament is being played in a is in Pacific countries. 04:29.460 --> 04:30.900 And let's see what will come. 04:32.230 --> 04:34.860 So we got stalled and find a undergoes. 04:35.140 --> 04:40.030 So we don't have enough data corresponding to this cricket might be. 04:40.210 --> 04:41.950 And that's why this thing happened. 04:42.790 --> 04:45.010 We can try with, let's say, hockey weather. 04:46.210 --> 04:51.130 I'm not sure where the hockey is a token available in this. 04:52.200 --> 04:54.900 Hello, regional leaders hope so hockey doesn't exist. 04:54.930 --> 04:55.920 Hockey doesn't exist. 04:56.330 --> 04:58.380 So let's just keep it as it is. 04:58.950 --> 05:02.350 So now you can understand the power of this word to whack Mardle. 05:02.370 --> 05:03.300 And we try to. 05:04.840 --> 05:10.600 Playing our own Martin over the weekend, we got some very, very good result, but we are still limited 05:10.600 --> 05:11.330 by the data set. 05:11.830 --> 05:13.750 And of course, that will be OK. 05:13.810 --> 05:16.720 My idea and intention is to teach you that we do. 05:16.730 --> 05:18.020 But this didn't seem like library. 05:18.030 --> 05:21.470 Feel your own custom data set how you can train the model. 05:22.080 --> 05:26.990 Now, next, we are going to deal with that same word, two VEC model. 05:27.040 --> 05:32.170 But that will be a pre train work work model and we will see the same result. 05:32.630 --> 05:40.420 What I would say will apply those very standard updated model, which is being trained by some big companies 05:40.840 --> 05:45.270 and those particular model will give it excellent results. 05:45.350 --> 05:46.950 So see you in the next video.