1 00:00:11,080 --> 00:00:16,120 So in this lecture, we will be summarizing this section of the course, which was on vector models 2 00:00:16,120 --> 00:00:16,850 in NLP. 3 00:00:17,650 --> 00:00:22,990 This section was essentially about how to make the connection between text, which is represented as 4 00:00:22,990 --> 00:00:27,650 a string and numbers, which is what we required to do any kind of analytics. 5 00:00:28,480 --> 00:00:32,650 Thus, this section was about how to convert text into vectors. 6 00:00:34,140 --> 00:00:39,870 Part of this section was about text preprocessing, since this is a necessary step before conversion 7 00:00:39,870 --> 00:00:41,130 into vectors can happen. 8 00:00:41,790 --> 00:00:47,910 In particular, we learned about tokenization, which is the process of converting text into tokens. 9 00:00:48,540 --> 00:00:52,230 We learned that tokens can be words, characters or sub words. 10 00:00:52,860 --> 00:00:58,650 We also learned about the bag of words concepts where we ignore the ordering of the tokens in the document. 11 00:00:59,970 --> 00:01:02,190 We also learn that not all words are equal. 12 00:01:02,580 --> 00:01:09,480 Some words like and it and is can appear in any kind of document and thus are likely not informative. 13 00:01:10,230 --> 00:01:11,700 We call these stop words. 14 00:01:12,420 --> 00:01:17,670 We then learned about stemming in limited zation, which involves converting a word into its root. 15 00:01:18,360 --> 00:01:23,730 This is useful because we don't want to have to independently learn about each variation of a word. 16 00:01:24,300 --> 00:01:29,580 We already know that words like run runs, running and ran all mean the same thing. 17 00:01:30,960 --> 00:01:34,800 Furthermore, this helps us to reduce the dimensionality of our vectors. 18 00:01:35,430 --> 00:01:40,770 One simple reason we want to do this is because the more dimensions we have, the more time it takes 19 00:01:40,770 --> 00:01:41,910 to do computation. 20 00:01:46,570 --> 00:01:50,260 We then learned about several techniques for converting text into vectors. 21 00:01:50,620 --> 00:01:55,540 The simplest being counts of tokens where each token is a separate vector component. 22 00:01:56,320 --> 00:02:01,030 Using this method, we were able to build a text classifier with pretty good performance. 23 00:02:01,720 --> 00:02:06,850 We then learned about TFI Taf, which handles one of the problems with count vectors, which is that 24 00:02:07,150 --> 00:02:12,910 it doesn't account for non informative words that may have high counts but show up in many documents. 25 00:02:13,750 --> 00:02:18,760 We also looked at the concept of Vector Similarity, which we applied to build a movie recommendation 26 00:02:18,760 --> 00:02:19,300 system. 27 00:02:20,380 --> 00:02:25,450 As an advanced exercise, we then looked at how to implement TFI Taf from scratch. 28 00:02:25,990 --> 00:02:30,160 This helped to give us deeper insight on how TFI Taf actually works. 29 00:02:31,480 --> 00:02:36,640 Furthermore, it introduced us to an important concept, which was the word to index mapping. 30 00:02:37,360 --> 00:02:43,000 This is especially useful in deep learning where we work with the indices and not the words themselves. 31 00:02:44,750 --> 00:02:50,120 The final topic we looked at was neural word embeddings, which are word vectors used in deep learning 32 00:02:50,120 --> 00:02:51,260 and neural networks. 33 00:02:51,800 --> 00:02:57,140 Specifically, we did a little preview of words of and glove, which, unlike the other methods we've 34 00:02:57,140 --> 00:03:01,100 discussed, convert words into vectors instead of hold documents. 35 00:03:01,910 --> 00:03:06,320 We then saw how these vectors can do interesting things like word analogies. 36 00:03:07,010 --> 00:03:12,440 They showed us that the simple process of converting words into vectors does something very meaningful. 37 00:03:13,640 --> 00:03:16,880 Words really represent concepts on multiple dimensions. 38 00:03:17,330 --> 00:03:19,070 For example, take the word king. 39 00:03:19,700 --> 00:03:22,880 King represents the ruler of a society in one dimension. 40 00:03:23,390 --> 00:03:25,730 King represents mail in another dimension. 41 00:03:26,270 --> 00:03:28,310 King is a noun in another dimension. 42 00:03:28,910 --> 00:03:32,600 In fact, these dimensions are not abstract, but rather numeric. 43 00:03:33,260 --> 00:03:38,870 This is actually a very interesting idea, since if we can represent concepts with numbers, then it 44 00:03:38,870 --> 00:03:45,080 may also be the case that thinking and reasoning are simply mathematical operations on those numbers.