1 00:00:09,555 --> 00:00:14,352 In this video, we are going go in a little bit more detail about how 2 00:00:14,352 --> 00:00:19,063 some of what we have learned in this course can be all put together 3 00:00:19,063 --> 00:00:23,100 in this specific natural language processing tasks. 4 00:00:23,100 --> 00:00:29,499 That are really important in a lot of tasks of understanding free-text. 5 00:00:29,499 --> 00:00:32,870 Information is, as we have seen so many times, 6 00:00:32,870 --> 00:00:37,070 is hidden in free-text in very interesting ways. 7 00:00:37,070 --> 00:00:41,460 You have most traditional traditional information that is structured, 8 00:00:41,460 --> 00:00:45,840 while a lot of information now is in unstructured free text form. 9 00:00:48,900 --> 00:00:52,250 As we have seen earlier when I was introducing this whole course, 10 00:00:52,250 --> 00:00:56,720 that about 80% of data is now in unstructured form, 11 00:00:56,720 --> 00:01:00,320 in blogs, in websites, on websites, and so on. 12 00:01:01,510 --> 00:01:02,580 And it's growing. 13 00:01:02,580 --> 00:01:03,707 You have tweets and 14 00:01:03,707 --> 00:01:07,590 other social media that is kind of adding more information here. 15 00:01:09,390 --> 00:01:10,990 So the question then becomes, 16 00:01:10,990 --> 00:01:15,370 how do you convert this unstructured text to structured form? 17 00:01:15,370 --> 00:01:19,060 You don't necessarily have to convert it all, but more of what im trying to get 18 00:01:19,060 --> 00:01:25,680 here is, how do you extract relevent information from unstructured text. 19 00:01:25,680 --> 00:01:29,350 And if you want to make it searchable or make it usable later on you would probably 20 00:01:29,350 --> 00:01:32,960 put it in a structured form as it has been traditionally kept. 21 00:01:32,960 --> 00:01:36,880 So that way it is conversion from unstructured to structured form text. 22 00:01:38,440 --> 00:01:40,330 And that's where information extraction comes in. 23 00:01:41,470 --> 00:01:46,520 The goal of this task is to identify and extract fields of interest from free text. 24 00:01:47,880 --> 00:01:52,450 So let's take an example of a medical article. 25 00:01:52,450 --> 00:01:55,860 This one is from WebMD from quite some time back. 26 00:01:55,860 --> 00:02:01,189 But it says Erbitux helps treat advanced lung cancer. 27 00:02:01,189 --> 00:02:06,096 By looking at this, you know that one of the pieces of information that 28 00:02:06,096 --> 00:02:09,102 you want to extract would be the headline. 29 00:02:09,102 --> 00:02:13,925 But very importantly also, the components of that saying that 30 00:02:13,925 --> 00:02:17,930 Erbitux is a particular treatment for lung cancer. 31 00:02:19,730 --> 00:02:25,012 And the fact that advance lung cancer is a specific form of lung cancer. 32 00:02:25,012 --> 00:02:29,550 So you're going to understand which modifier could be dropped. 33 00:02:29,550 --> 00:02:33,100 So in this case, advanced could be taught but lung couldn't be. 34 00:02:33,100 --> 00:02:37,920 Because lung cancer treatments are different from let's say blood cancer or 35 00:02:38,980 --> 00:02:40,060 breast cancer and so forth. 36 00:02:42,230 --> 00:02:46,490 But this webpage has more information than just the headline. 37 00:02:47,650 --> 00:02:54,220 It has some information about who wrote this article or who reviewed this article. 38 00:02:55,670 --> 00:03:02,310 Or when was it published or the place it was published from and so on. 39 00:03:02,310 --> 00:03:07,460 And you can see that this happens all the time in news reports and other places 40 00:03:07,460 --> 00:03:13,520 when you see some piece of information, some news article or a tweet. 41 00:03:13,520 --> 00:03:18,440 It has more than just the text that is important, and 42 00:03:18,440 --> 00:03:20,470 are of some interest for people. 43 00:03:20,470 --> 00:03:23,730 So who wrote a particular news report? 44 00:03:23,730 --> 00:03:24,710 Where is it coming from? 45 00:03:26,320 --> 00:03:27,550 What does it say? 46 00:03:27,550 --> 00:03:31,062 Who or what are the other persons involved in this story? 47 00:03:31,062 --> 00:03:37,148 This is the typical representation of an information extraction task. 48 00:03:39,516 --> 00:03:43,230 The first thing to notice here is something about fields of interest. 49 00:03:44,680 --> 00:03:47,309 These fields of interest are named entities. 50 00:03:49,350 --> 00:03:54,504 If it's news, we are talking about people, and places, and date, and 51 00:03:54,504 --> 00:03:59,220 organizations, and other geopolitical entities. 52 00:03:59,220 --> 00:04:00,560 When you say White House, 53 00:04:00,560 --> 00:04:04,710 you mean something completely different than a white house. 54 00:04:04,710 --> 00:04:09,538 Or when you're saying 10 Downing Street it's a completely different 55 00:04:09,538 --> 00:04:13,189 meaning than the actual address 10 Downing Street. 56 00:04:13,189 --> 00:04:18,368 If you're talking about finance you're talking about money, or monetary values. 57 00:04:18,368 --> 00:04:23,987 About how much worth particular save for, or the names of companies, 58 00:04:23,987 --> 00:04:28,000 or the stock price index of a particular company. 59 00:04:29,820 --> 00:04:32,780 When talking about medicine and health, you're talking about diseases and 60 00:04:32,780 --> 00:04:34,200 drugs and procedures. 61 00:04:35,230 --> 00:04:40,030 And especially if you're talking about the protected information, protected health 62 00:04:40,030 --> 00:04:46,260 information, then you're talking about address and their unique identifiers. 63 00:04:46,260 --> 00:04:51,020 Sometimes even emails and URLs and provisions and so on. 64 00:04:51,020 --> 00:04:55,039 So you can see that these name can be very, 65 00:04:55,039 --> 00:05:01,178 very diverse as they don't all follow the same ideas and the same features. 66 00:05:01,178 --> 00:05:05,577 So for example, in news it's fairly common for people, 67 00:05:05,577 --> 00:05:08,338 name and places to be capitalized. 68 00:05:08,338 --> 00:05:13,402 So you have a title case there, you would see China 69 00:05:13,402 --> 00:05:17,871 with C capital more often than with small c. 70 00:05:17,871 --> 00:05:20,829 You have dates, they are typically of a particular format, 71 00:05:20,829 --> 00:05:23,680 we have seen some examples of that early on in this course. 72 00:05:24,790 --> 00:05:27,830 But those don't apply in medicine. 73 00:05:27,830 --> 00:05:31,850 So, for example, when you're talking about lung cancer, you don't have capital L and 74 00:05:31,850 --> 00:05:32,520 capital C. 75 00:05:34,410 --> 00:05:39,410 When you're writing it out, it's going to be all small case. 76 00:05:39,410 --> 00:05:44,620 So the particular rule that you might have for named entity being title cased is not 77 00:05:44,620 --> 00:05:47,229 applicable for some of the data sets, like in medicine. 78 00:05:49,650 --> 00:05:54,040 Once you have understood what named entities are, the next is relations. 79 00:05:54,040 --> 00:05:59,790 Relations are, basically is, what happened to who, when, where, and so on. 80 00:05:59,790 --> 00:06:05,700 So again, here even though the particular pieces of information are named entities, 81 00:06:05,700 --> 00:06:10,100 so who is basically a person or an organization. 82 00:06:10,100 --> 00:06:12,640 When is a particular time, or a date. 83 00:06:12,640 --> 00:06:18,790 Where is a place, but it also has this relationship between these named entities, 84 00:06:18,790 --> 00:06:21,250 to get meaningful semantics out of it. 85 00:06:22,700 --> 00:06:26,880 So let's take these two examples in more detail and see what that means. 86 00:06:26,880 --> 00:06:31,940 So named entity recognition relies on something called named entities. 87 00:06:31,940 --> 00:06:35,230 Named entities are noun phrases that are of specific type and 88 00:06:35,230 --> 00:06:38,650 refer to specific individuals, places, organizations, and so on. 89 00:06:39,740 --> 00:06:45,850 And the named entity recognition task is a set of techniques and methods 90 00:06:45,850 --> 00:06:52,170 that would help identify all mentions of predefined named entities in text. 91 00:06:52,170 --> 00:06:56,830 So for example, you are to identify the mention, 92 00:06:56,830 --> 00:06:59,680 the actual phrase where something is happening. 93 00:06:59,680 --> 00:07:04,280 So you need to know where does that mention start and where does it end? 94 00:07:04,280 --> 00:07:09,480 So this is the boundary detection subtask within named entity recognition. 95 00:07:09,480 --> 00:07:14,970 Once you understood the boundary, the next task is to type it, 96 00:07:14,970 --> 00:07:19,730 to tag it, or to classify one of the named entity types that you are interested in. 97 00:07:21,170 --> 00:07:26,593 So for example, if you have the word Chicago, 98 00:07:26,593 --> 00:07:34,060 it could be a place, it could be an album, it could be a font. 99 00:07:34,060 --> 00:07:37,190 So depending on what kind of variation it is, 100 00:07:37,190 --> 00:07:41,670 you need to know what is the label it should be assigned to this word. 101 00:07:41,670 --> 00:07:46,750 Once you identify that Chicago is your element phrase. 102 00:07:46,750 --> 00:07:49,710 So let's take an example of an named entity task 103 00:07:49,710 --> 00:07:52,100 in the context of a medical document. 104 00:07:53,850 --> 00:07:56,810 I'm going to leave this one and let you try out 105 00:07:56,810 --> 00:08:00,510 what are the different named entities here that you would be interested in. 106 00:08:03,300 --> 00:08:07,560 Okay now that you've had a chance to see what are the named entities, 107 00:08:07,560 --> 00:08:08,770 let's see some of them. 108 00:08:11,030 --> 00:08:16,168 One big one is the disease or the condition or the symptom, right? 109 00:08:16,168 --> 00:08:19,705 So you have bilateral hand numbness or 110 00:08:19,705 --> 00:08:24,467 occasional weakness or C5-6 disc herniation. 111 00:08:24,467 --> 00:08:27,800 Okay, but you also have age here. 112 00:08:27,800 --> 00:08:32,690 So you could say age is 63 or you could say age is 63 year old. 113 00:08:33,850 --> 00:08:36,880 You have gender, female. 114 00:08:36,880 --> 00:08:44,441 You have mention of a medical specialty in neurologist. 115 00:08:44,441 --> 00:08:49,874 You have a procedure that was done, so you have a workup or an MRI. 116 00:08:49,874 --> 00:08:55,882 And in relation to all of this, you also have body parts like feet and hand. 117 00:08:55,882 --> 00:09:00,654 And certainly, you come to a conclusion where you can see that these 118 00:09:00,654 --> 00:09:04,910 named entities don't have to be separate from each other. 119 00:09:04,910 --> 00:09:13,380 So for example, bilateral hand numbness is a completely valid named entity in itself. 120 00:09:13,380 --> 00:09:16,860 But hand is also a named entity that is of interest. 121 00:09:16,860 --> 00:09:18,190 So then the question becomes, 122 00:09:18,190 --> 00:09:22,390 what is the granularity with which you need to annotate this. 123 00:09:24,160 --> 00:09:26,460 And if you're doing hand as a separate annotation, 124 00:09:26,460 --> 00:09:32,350 why not C5-6 disc or cord because you're talking about spinal cord in this case, 125 00:09:32,350 --> 00:09:33,710 you're talking about an anatomical part. 126 00:09:34,970 --> 00:09:38,680 So these are the decisions that need to be made when you're defining this named 127 00:09:38,680 --> 00:09:40,050 entity task. 128 00:09:40,050 --> 00:09:43,970 So in this particular case, if you define the task to be one, 129 00:09:43,970 --> 00:09:49,880 where you say we are only interested in the condition, the diagnosis, 130 00:09:49,880 --> 00:09:56,190 the age, the gender, the procedures that were done, and the specialty. 131 00:09:56,190 --> 00:09:57,070 That's it. 132 00:09:57,070 --> 00:10:02,050 In this then, you're not going to focus on a body part, hand and feet, and 133 00:10:02,050 --> 00:10:03,641 even if they are there, that's fine. 134 00:10:03,641 --> 00:10:06,320 And some decision already was done that way. 135 00:10:06,320 --> 00:10:12,442 So for example, nobody said that patient is not a valid entity in this role, 136 00:10:12,442 --> 00:10:15,324 it could be, it very well could be. 137 00:10:15,324 --> 00:10:19,131 If neurologist is a valid entity, patient could be a valid entity, and 138 00:10:19,131 --> 00:10:23,650 that's a decision that has to be made when you're defining this named entity task. 139 00:10:24,720 --> 00:10:29,010 Suppose you have these goals then, you know that you have not only 140 00:10:29,010 --> 00:10:32,780 identified the boundary, but also colored it in some way, 141 00:10:32,780 --> 00:10:38,020 tagged it to see what is the type of this [INAUDIBLE] that we have identified. 142 00:10:39,850 --> 00:10:43,160 Okay, now that we know the task well, 143 00:10:43,160 --> 00:10:46,500 what are the approaches that people use to identify named entities? 144 00:10:47,600 --> 00:10:50,950 It really depends on the kind of entities that we have. 145 00:10:50,950 --> 00:10:55,090 So for example, if you have well-formatted fields like dates and 146 00:10:55,090 --> 00:10:58,930 phone numbers, then you would use regular expressions. 147 00:10:58,930 --> 00:11:02,650 Recall that we have done the regular expressions for date, for 148 00:11:02,650 --> 00:11:05,440 example, in week one of this course. 149 00:11:05,440 --> 00:11:10,010 And that is a very valid named entity identification task. 150 00:11:10,010 --> 00:11:13,500 In fact, the assignment was really asking you to 151 00:11:13,500 --> 00:11:18,330 do an information extraction task for dates from the given text file. 152 00:11:20,760 --> 00:11:24,870 For other fields, it's fairly common to use a machine learning approach. 153 00:11:24,870 --> 00:11:29,040 In fact, even for dates and phone numbers you might want to 154 00:11:29,040 --> 00:11:33,430 use a machine learning approach, where you use these regular expressions as features. 155 00:11:33,430 --> 00:11:37,930 But you have other features that help them define this particular 156 00:11:39,140 --> 00:11:40,720 extraction to be valid. 157 00:11:40,720 --> 00:11:44,850 So for example, phone numbers and fax numbers are very similar. 158 00:11:46,230 --> 00:11:50,720 The only thing that differentiates them is what is the number preceded by. 159 00:11:50,720 --> 00:11:54,593 So is it phone dot, or is it fax dot. 160 00:11:54,593 --> 00:11:58,165 So some partition information kind of helps doing that, and 161 00:11:58,165 --> 00:12:02,717 you could either include it as a regular expression or let those be features in 162 00:12:02,717 --> 00:12:06,656 a machine learning model that you then train over large data sets. 163 00:12:06,656 --> 00:12:16,250 The standard NER task in natural language processing is this four-class model. 164 00:12:16,250 --> 00:12:23,020 Person, organization, location, and everything else, or everything outside. 165 00:12:23,020 --> 00:12:28,790 So, if you have a sentence like, John met Brenda. 166 00:12:29,790 --> 00:12:31,660 You want to say that John and 167 00:12:31,660 --> 00:12:37,540 Brenda are named entities so you have those two persons. 168 00:12:37,540 --> 00:12:41,830 And then Matt is a word that is outside of any named entity, so 169 00:12:41,830 --> 00:12:43,928 you are going to use these two classes. 170 00:12:43,928 --> 00:12:48,460 So you're going to say John is a person, Matt is outside and 171 00:12:48,460 --> 00:12:49,669 Brenda is a person again. 172 00:12:51,910 --> 00:12:54,550 Once you have identified these, then you are going to go and 173 00:12:55,580 --> 00:12:58,090 ask the question about relation and relation extraction. 174 00:12:59,200 --> 00:13:02,142 The relation extraction task is identifying the relationship between 175 00:13:02,142 --> 00:13:02,737 named entities. 176 00:13:04,200 --> 00:13:06,640 So let's take an example of 177 00:13:06,640 --> 00:13:10,840 something that we have seen earlier today on Erbitux help treat lung cancer. 178 00:13:12,270 --> 00:13:18,170 This sentence has a relationship between two named entities. 179 00:13:18,170 --> 00:13:19,880 One is Erbitux. 180 00:13:19,880 --> 00:13:24,460 Erbitux is colored yellow here to represent that it is a diagnosis, 181 00:13:24,460 --> 00:13:27,330 sorry it is a treatment of some kind. 182 00:13:27,330 --> 00:13:30,830 And then you have lung cancer that is represented, 183 00:13:30,830 --> 00:13:34,300 that's in green to represent a different name identity. 184 00:13:34,300 --> 00:13:37,430 In this particular case, a disease or a diagnosis. 185 00:13:39,500 --> 00:13:41,860 And then you have a relationship between them. 186 00:13:41,860 --> 00:13:45,709 So Erbitux and lung cancer are linked by a treatment relation. 187 00:13:45,709 --> 00:13:50,675 Going from Erbitux to lung cancer, that says Erbitux is a treatment for 188 00:13:50,675 --> 00:13:51,677 lung cancer. 189 00:13:51,677 --> 00:13:56,750 You could easily have a link going the other way. 190 00:13:56,750 --> 00:14:02,880 Lung cancer and to Erbitux, lung cancer is treated by Erbitux. 191 00:14:02,880 --> 00:14:06,000 So this is the relation, a very simple relation, 192 00:14:06,000 --> 00:14:11,330 a binder relation between a disease and a treatment, or a drug. 193 00:14:14,370 --> 00:14:17,320 The other task is co-reference resolution. 194 00:14:18,580 --> 00:14:21,880 That is to disambiguate mentions in text and 195 00:14:21,880 --> 00:14:25,500 group mentions together if they are referring to the same entity. 196 00:14:26,760 --> 00:14:30,590 An example would be if Anita met Joseph at the market and 197 00:14:30,590 --> 00:14:37,270 he surprised her with a rose, then you have two named entities, Anita and Joseph. 198 00:14:37,270 --> 00:14:40,270 But then, the second sentence uses pronouns, 199 00:14:40,270 --> 00:14:44,110 he to refer to Joseph and her to refer to Anita. 200 00:14:45,175 --> 00:14:51,510 So in this case, it's pronoun resolution where you are making an inference 201 00:14:51,510 --> 00:14:56,910 that if Anita met Joseph at the market, Joseph surprised Anita with a rose. 202 00:14:59,360 --> 00:15:00,950 Why is that important? 203 00:15:00,950 --> 00:15:06,650 That is in a question answering task where you're given a question and 204 00:15:06,650 --> 00:15:11,650 you need to find the most appropriate answer from text. 205 00:15:11,650 --> 00:15:15,890 Now appropriateness is something that we can talk about in more detail. 206 00:15:15,890 --> 00:15:19,330 But let's say in this case, the correct answer, or 207 00:15:19,330 --> 00:15:22,600 the set of correct answers, would be one definition here. 208 00:15:24,715 --> 00:15:29,290 Andyou would need, both named entity recognition and 209 00:15:29,290 --> 00:15:31,770 relation extraction to answer these questions. 210 00:15:31,770 --> 00:15:35,470 So, for example, the question was Erbitux treat? 211 00:15:35,470 --> 00:15:39,550 You first have to identify that Erbitux is a treatment. 212 00:15:39,550 --> 00:15:41,392 The relation is treat and 213 00:15:41,392 --> 00:15:46,741 then somehow fill a slot in this relation to say Erbitux is a treatment for 214 00:15:46,741 --> 00:15:51,145 something and that comes from text and that is lung cancer. 215 00:15:51,145 --> 00:15:53,111 Or who gave Anita the rose, 216 00:15:53,111 --> 00:15:57,986 where you are doing some sort of pronoun resolution to then identify 217 00:15:57,986 --> 00:16:03,129 that Joseph was who was the person who gave Anita the rose at the market. 218 00:16:03,129 --> 00:16:05,844 So this builds on the named entity recognition task and 219 00:16:05,844 --> 00:16:09,990 the relation extraction and co-reference resolution that we have seen earlier. 220 00:16:12,890 --> 00:16:18,800 So, to conclude, we see that Information Extraction is important task for 221 00:16:18,800 --> 00:16:22,910 natural language understanding and making sense of textual data. 222 00:16:22,910 --> 00:16:26,970 It is the first step in converting this unstructured 223 00:16:26,970 --> 00:16:28,650 text in to more structured form. 224 00:16:30,010 --> 00:16:34,240 Name Entity Recognition becomes a key building block in addressing these 225 00:16:34,240 --> 00:16:38,535 tasks and these advanced NLP tasks. 226 00:16:38,535 --> 00:16:43,955 And named Entity Recognition systems use the supervised machine learning 227 00:16:43,955 --> 00:16:48,395 approaches and text mining approaches that we have discussed in the course. 228 00:16:48,395 --> 00:16:52,655 So for example, if the entity that you need to recognize is a date, 229 00:16:52,655 --> 00:16:58,710 you are using typically expressions that we've talked about in week one. 230 00:16:58,710 --> 00:17:03,710 If you are talking about extracting people name. 231 00:17:03,710 --> 00:17:08,220 So person name or organization name, you are not only using emotional learning 232 00:17:08,220 --> 00:17:12,640 model to identify what is an identity and 233 00:17:12,640 --> 00:17:17,360 what label you should get it, but also the features that you're going to use 234 00:17:17,360 --> 00:17:20,910 are coming from what we talked about in week two. 235 00:17:20,910 --> 00:17:25,240 So, for example, we want to know that, yes, if it is capitalized or 236 00:17:25,240 --> 00:17:28,840 not, but what is the part of speech of a particular word? 237 00:17:28,840 --> 00:17:30,271 Is it a noun or a verb? 238 00:17:30,271 --> 00:17:34,221 What is the semantic role that a particular 239 00:17:34,221 --> 00:17:38,730 word is playing in a given context in a sentence? 240 00:17:38,730 --> 00:17:42,650 And these could be features that you would then put in in a named entity 241 00:17:42,650 --> 00:17:43,873 recognition model. 242 00:17:43,873 --> 00:17:49,234 NLTK has an in-built NER model that does trained, 243 00:17:49,234 --> 00:17:54,595 or new datasets and so on, for the standard task for 244 00:17:54,595 --> 00:18:00,320 the person, organization, location task. 245 00:18:00,320 --> 00:18:05,320 But I have shown in this video the named entity recognition 246 00:18:05,320 --> 00:18:07,990 problem goes beyond just news. 247 00:18:07,990 --> 00:18:10,450 If your even extending it to finance or 248 00:18:10,450 --> 00:18:15,760 then if you go all the way to medicine, it's a completely open problem. 249 00:18:15,760 --> 00:18:18,600 Where the topics that we have talked about in this 250 00:18:18,600 --> 00:18:22,710 course kind of puts it all together and brings it all together. 251 00:18:22,710 --> 00:18:26,190 Hope you had fun with the course and I had fun presenting it to you.