1 00:00:00,000 --> 00:00:06,038 Let's now turn to the use of summarization in question answering. The goal of 2 00:00:06,038 --> 00:00:10,348 summarization is to produce an abridged version of the text that contains 3 00:00:10,348 --> 00:00:14,555 information that's important to or relevant for a user, a particular user 4 00:00:14,555 --> 00:00:18,471 need. And so that might mean abstracting any kind of document or article, 5 00:00:18,471 --> 00:00:22,649 summarizing email threads or meetings, simplifying text, we'll talk about 6 00:00:22,649 --> 00:00:27,645 compressing sentences. And we can talk about summarization in two ways. Single 7 00:00:27,645 --> 00:00:32,162 document summarization. Given a single document, we're gonna produce an abstract 8 00:00:32,162 --> 00:00:36,737 or an outline, perhaps a headline, a very short summary of a document. In multiple 9 00:00:36,737 --> 00:00:41,026 document summarization, we're given a group of documents, presumably related 10 00:00:41,026 --> 00:00:45,920 documents, and our job is to produce the gist of these documents. So if we have 11 00:00:45,920 --> 00:00:50,005 some news stories on the same event, we might summarize the event by picking 12 00:00:50,005 --> 00:00:54,808 information from different documents. And that could be true also for any kind of 13 00:00:54,808 --> 00:00:59,438 set of web pages that are focused on some topic or question. Now we distinguish 14 00:00:59,438 --> 00:01:03,769 between generic summarization and query-focused summarization. Generic 15 00:01:03,769 --> 00:01:08,224 summarization is what I've talked about so far. We have some document, or some set of 16 00:01:08,224 --> 00:01:12,788 documents, and our job is just to build a summary. Query-focused summarization is 17 00:01:12,788 --> 00:01:17,524 summarizing a document with respect to the information need the user expresses by 18 00:01:17,524 --> 00:01:21,403 their query. So you can think of query-focused summarization as just 19 00:01:21,403 --> 00:01:25,568 another kind of complex question answering. We're answering a question by 20 00:01:25,568 --> 00:01:30,018 summarizing a document that has the information in it to construct the answer 21 00:01:30,018 --> 00:01:34,601 that the user needs. Now we've all seen query focused summaries, because that's 22 00:01:34,601 --> 00:01:39,178 what search engines use to show you information about a page. So for example, 23 00:01:39,178 --> 00:01:44,610 Google will give you 156 characters about 26 words for each page as a summary of the 24 00:01:44,610 --> 00:01:49,374 page. So if I've asked a question to Google, what is die brücke, I get, 25 00:01:49,374 --> 00:01:54,830 let's say these three URLs back. We have that title of the URL, the URL, and then 26 00:01:54,830 --> 00:01:59,990 some text. And that text, that snippet, is some kind of summary of the page that 27 00:01:59,990 --> 00:02:04,434 helps the user understand what's in the page. So we can think of summarization in 28 00:02:04,434 --> 00:02:07,913 general, of single documents, as generating things like snippets. And we're 29 00:02:07,913 --> 00:02:12,873 going to talk about the more general task other than specifically just snippets for 30 00:02:12,873 --> 00:02:17,420 web search, but you can think of these snippets as a characteristic example. 31 00:02:17,420 --> 00:02:21,760 Now where a snippet is the answer to a question by taking information from a 32 00:02:21,760 --> 00:02:26,100 single document, we often want the answer from multiple documents. So creating 33 00:02:26,100 --> 00:02:30,891 answers to complex questions that involves summarizing from multiple documents, we're 34 00:02:30,891 --> 00:02:35,175 gonna create a cohesive answer that combines information from each of these 35 00:02:35,175 --> 00:02:39,376 documents. That's multiple document question answering. Now we can think of two 36 00:02:39,376 --> 00:02:43,246 ways of doing summarization. In extractive summarization we create a 37 00:02:43,246 --> 00:02:47,391 summary by taking particular words or phrases that are in the document and 38 00:02:47,391 --> 00:02:51,924 building a summary just out of those words or phrases. And a snippet from a search 39 00:02:51,924 --> 00:02:55,943 engine like Google is an extractive summary in this way. By contrast an 40 00:02:55,943 --> 00:03:01,090 abstractive summary is one in which we create our own words, different words than 41 00:03:01,090 --> 00:03:06,303 are in the text, to summarize the content of the text. And, we're going to be talking 42 00:03:06,303 --> 00:03:10,982 completely about extractive summarization today. Abstractive summarization is an 43 00:03:10,982 --> 00:03:16,704 important research goal, but very difficult. And one thing to think about 44 00:03:16,704 --> 00:03:20,684 whenever we're talking about summarizing from a document, whether it's a web page 45 00:03:20,684 --> 00:03:25,261 or anything else, is what our baselines are for summarization. And good writers 46 00:03:25,261 --> 00:03:28,901 often put their ideas right at the beginning, in the title or in the first 47 00:03:28,901 --> 00:03:32,603 sentence. So a simple baseline, whenever we're measuring any kind of summarization 48 00:03:32,603 --> 00:03:38,259 algorithm is just taking the first sentence. So, for example, in this Google 49 00:03:38,259 --> 00:03:43,313 query: what is Die Brücke, if we take the output from this first hit and we 50 00:03:43,313 --> 00:03:47,707 look at this sentence Die Brücke was a group of German expressionist artists 51 00:03:47,707 --> 00:03:54,660 formed in Dresden and so on. Let's look at this page where the snippet comes from. 52 00:03:54,920 --> 00:04:04,623 And we can see that this snippet is really just the first 156 characters of this 53 00:04:04,623 --> 00:04:10,126 page. So sometimes the best snippet is really just the beginning. And so we're 54 00:04:10,126 --> 00:04:14,007 going to use that not only as a baseline but later we'll see that that's a feature 55 00:04:14,007 --> 00:04:18,794 we could use in different kinds of summarization algorithms. We've 56 00:04:18,794 --> 00:04:23,371 introduced the task of natural language generation, producing shorter abstracts or 57 00:04:23,371 --> 00:04:27,446 summaries or even headlines from a document. And we've talked about query 58 00:04:27,446 --> 00:04:31,353 focused summarization, where we use summarization to answer a particular 59 00:04:31,353 --> 00:04:36,209 question. We build a summary of a document that is specifically designed to answer a 60 00:04:36,209 --> 00:04:40,618 question posed by the user. And we've talked about single and multiple document 61 00:04:40,618 --> 00:04:44,685 summarization. And now we'll see the details of how to do each of these tasks 62 00:04:44,685 --> 00:04:46,400 in the further lectures.