1 00:00:11,090 --> 00:00:16,129 So in this lecture, we will be introducing the next section of this course, which is all about how 2 00:00:16,129 --> 00:00:18,110 to build summaries out of text. 3 00:00:18,680 --> 00:00:22,700 Now, as always, we're going to answer the question why is this practical? 4 00:00:23,150 --> 00:00:27,470 And I hope that this is another one of those tasks where the answer is very obvious. 5 00:00:28,400 --> 00:00:30,320 So why should we summarize text? 6 00:00:31,190 --> 00:00:33,950 Well, notice how we already do this all the time. 7 00:00:34,490 --> 00:00:40,250 For example, when you read a scientific paper, it always comes with an abstract and abstract is a 8 00:00:40,250 --> 00:00:41,550 summary of that paper. 9 00:00:42,230 --> 00:00:46,340 So you don't need to read the whole paper if you just want to understand the basic idea. 10 00:00:46,880 --> 00:00:52,550 You can get away with just reading the abstract and then read the paper in full if you want more details. 11 00:00:53,600 --> 00:00:59,240 Professional documents also have a similar concept, although instead of calling it an abstract, it's 12 00:00:59,240 --> 00:01:01,040 sometimes called an executive summary. 13 00:01:01,760 --> 00:01:07,280 So clearly, this is a useful concept in the real world since we do this already, even without the 14 00:01:07,280 --> 00:01:08,390 help of computers. 15 00:01:09,260 --> 00:01:12,560 And obviously, we can extend this concept to our own lives. 16 00:01:12,920 --> 00:01:18,950 I know that for myself, my reading list is longer than the time I have just been reading, so it would 17 00:01:18,950 --> 00:01:24,620 be very useful if I could reach the same outcome or reach the same conclusions by reading only the most 18 00:01:24,620 --> 00:01:27,350 important parts of each document that I have to read. 19 00:01:32,000 --> 00:01:37,190 Another reason we might want to build tech summarisation systems is for pushing forward the field of 20 00:01:37,190 --> 00:01:38,600 artificial intelligence. 21 00:01:39,200 --> 00:01:41,660 Now you might think this seems kind of trivial. 22 00:01:41,690 --> 00:01:46,160 Build a algorithms to develop the field of A.I., but that's not what I mean. 23 00:01:47,000 --> 00:01:51,260 What I mean is that tech summarization is kind of a critical aspect of learning. 24 00:01:51,950 --> 00:01:55,670 For instance, think of everyday conversations that you have with other people. 25 00:01:56,240 --> 00:02:01,310 Sometimes when people explain things to you, you will paraphrase or summarize what they've told you 26 00:02:01,460 --> 00:02:02,420 in response. 27 00:02:02,840 --> 00:02:06,410 Then the other person might say, Yes, you're correct or no, you're not correct. 28 00:02:07,010 --> 00:02:12,980 So summarization is a way for learning systems, whether that's yourself or a machine, to demonstrate 29 00:02:12,980 --> 00:02:14,720 their understanding of a concept. 30 00:02:15,800 --> 00:02:21,140 If you think about it further, you'll realize that whenever you summarize an article or write an abstract, 31 00:02:21,500 --> 00:02:25,850 it really requires you to have a good understanding of what you are trying to summarize. 32 00:02:26,330 --> 00:02:29,420 In other words, it seems to require a true intelligence. 33 00:02:33,930 --> 00:02:37,140 Here's another application for building summaries out of text. 34 00:02:37,710 --> 00:02:42,030 In fact, you see this so often you probably don't even realize that it happens. 35 00:02:43,170 --> 00:02:49,170 As you recall, our sort of window to the internet is search engines such as Google, Bing, DuckDuckGo 36 00:02:49,170 --> 00:02:49,980 and so forth. 37 00:02:50,670 --> 00:02:54,510 Now, without looking, think about what you see after you type in a query. 38 00:02:55,710 --> 00:03:00,780 And while the answer is you see a bunch of search results and how are these search results presented? 39 00:03:01,710 --> 00:03:07,200 Well, typically you have a webpage title, a link to the webpage and the corresponding summary, which 40 00:03:07,200 --> 00:03:10,110 is these days usually customized your query. 41 00:03:10,770 --> 00:03:13,110 So this is another very important application. 42 00:03:17,770 --> 00:03:22,960 So now that you know why tech summarization is useful, let's discuss one detail before we outline this 43 00:03:22,960 --> 00:03:23,470 section. 44 00:03:24,580 --> 00:03:29,410 This detail is that we can actually categorize different methods of summarizing text. 45 00:03:30,010 --> 00:03:33,020 These categories are extractive and obstructive. 46 00:03:33,670 --> 00:03:34,600 So what do these mean? 47 00:03:36,040 --> 00:03:40,210 Well, extract of summaries or summaries that are built out of the document itself? 48 00:03:40,810 --> 00:03:43,960 So the summary is a subset of the document that was given. 49 00:03:44,920 --> 00:03:48,220 In other words, the summary was extracted from the document. 50 00:03:49,330 --> 00:03:55,060 Abstract of summaries, on the other hand, are novel sequences of text that are not just partial copies 51 00:03:55,060 --> 00:03:56,200 of the given document. 52 00:03:56,980 --> 00:04:02,770 As you can imagine, extractive summaries are much easier to generate compared with abstract of summaries 53 00:04:03,280 --> 00:04:04,660 with extractive summaries. 54 00:04:04,690 --> 00:04:09,820 We only need to identify the most relevant parts of the input, but with abstract of summaries, we 55 00:04:09,820 --> 00:04:12,190 need to develop a deeper understanding of the input. 56 00:04:12,520 --> 00:04:16,390 And we also need to convert that understanding into coherent text. 57 00:04:18,040 --> 00:04:23,170 Note that this section of the course will focus on extractive summaries, abstract of summaries are 58 00:04:23,170 --> 00:04:28,000 more suited to deep neural networks, such as sequence to sequence models and transformers. 59 00:04:32,550 --> 00:04:37,560 So now that you understand our general approach to the SUMMARISATION task, let's look at a section 60 00:04:37,560 --> 00:04:38,190 outline. 61 00:04:39,120 --> 00:04:41,100 This section will look at two methods. 62 00:04:41,640 --> 00:04:46,290 The first is pretty simple and only requires knowledge you learned in the first part of this course 63 00:04:46,560 --> 00:04:47,970 on vector based methods. 64 00:04:48,600 --> 00:04:51,450 The second is more complex and builds on the first. 65 00:04:51,930 --> 00:04:55,290 This one is called Tex Rank, which is based on Google's Page Rank. 66 00:04:55,710 --> 00:04:59,430 So if you've learned about Page Rank in the past, then this should be familiar. 67 00:05:00,390 --> 00:05:05,670 This one is more advanced and requires a bit of knowledge about probability and linear algebra, including 68 00:05:05,670 --> 00:05:11,520 Markov chains, although the way we use Markov chains is different from the probability section of this 69 00:05:11,520 --> 00:05:12,090 course. 70 00:05:13,320 --> 00:05:18,570 Note that if you were a beginner student and you only wants to know how to use text rank without going 71 00:05:18,570 --> 00:05:24,450 through any advanced math, this is possible after the advanced lectures will have one more beginner 72 00:05:24,450 --> 00:05:28,950 lecture where we look at libraries that implement text rank, which you can use directly.