WEBVTT 00:00.780 --> 00:03.600 -: One obvious thing you might like to do when using ChatGPT 00:03.600 --> 00:05.580 is to get it to classify data for you 00:05.580 --> 00:07.770 into specific different types of labels. 00:07.770 --> 00:09.720 For example, if you're looking at reviews, 00:09.720 --> 00:11.760 maybe that's whether the review was positive 00:11.760 --> 00:13.530 or whether the review was negative. 00:13.530 --> 00:15.060 Let's have a look at how we can do this 00:15.060 --> 00:18.420 using LangChain and the ChatGPT API. 00:18.420 --> 00:20.336 First, you're gonna have to install LangChain 00:20.336 --> 00:21.750 and Langchain OpenAI. 00:21.750 --> 00:24.390 And then what we're gonna do is import the chat models, 00:24.390 --> 00:28.020 ChatOpenAI, SystemMessage, and HumanMessage. 00:28.020 --> 00:29.340 After running that as well, 00:29.340 --> 00:32.220 you'll also need to import your OPENAI_API_KEY 00:32.220 --> 00:34.170 if you haven't done so already. 00:34.170 --> 00:35.003 As you'll see here, 00:35.003 --> 00:36.960 we've got a list of different types of reviews 00:36.960 --> 00:39.240 where each review is based on a movie. 00:39.240 --> 00:41.640 It says things like I love or I hate this movie, 00:41.640 --> 00:43.350 or this movie's a waste of time. 00:43.350 --> 00:48.000 And so what we wanna do is build a classifier using ChatGPT 00:48.000 --> 00:50.940 that allows us to get structured labeled data 00:50.940 --> 00:52.530 from this review text. 00:52.530 --> 00:53.790 So we've got our Python list, 00:53.790 --> 00:55.380 which is a list of dictionaries. 00:55.380 --> 00:57.960 We can instantiate our ChatOpenAI model, 00:57.960 --> 01:01.050 and then we can store our classifications in a Python list. 01:01.050 --> 01:03.390 We'll loop over each individual review 01:03.390 --> 01:04.290 and then we'll just print 01:04.290 --> 01:06.300 that we're gonna classify that review. 01:06.300 --> 01:08.490 And then inside each individual review, 01:08.490 --> 01:10.657 we're gonna have a system prompt which says, 01:10.657 --> 01:13.770 "You're responsible for the classification of movie reviews. 01:13.770 --> 01:15.540 Please classify the following review 01:15.540 --> 01:17.340 as positive or negative. 01:17.340 --> 01:19.590 You must use only the following words: 01:19.590 --> 01:21.540 negative or positive." 01:21.540 --> 01:22.890 And then we've got the HumanMessage 01:22.890 --> 01:26.610 which contains the content of specifically what that review. 01:26.610 --> 01:29.220 We're also saying if the content 01:29.220 --> 01:31.980 that we get back from the LM is not negative or positive, 01:31.980 --> 01:32.850 then we're throwing an error 01:32.850 --> 01:34.590 because we've got an invalid classification. 01:34.590 --> 01:36.840 And that's something we'd look at fixing 01:36.840 --> 01:39.120 in terms of on an individual basis. 01:39.120 --> 01:41.100 And so I can run this 01:41.100 --> 01:42.570 and you'll see that we're actually generating 01:42.570 --> 01:43.950 different types of classifications 01:43.950 --> 01:45.750 for all of these types of reviews. 01:45.750 --> 01:47.580 And then we can go and have a look 01:47.580 --> 01:49.140 at the response of this afterwards. 01:49.140 --> 01:50.940 So it's gonna print this out so you can see 01:50.940 --> 01:53.610 it's looping through each individual classification. 01:53.610 --> 01:55.800 And then what we're gonna have at the end of this 01:55.800 --> 01:58.503 is a way for us to classify different reviews. 01:59.580 --> 02:00.780 So we finally finished 02:00.780 --> 02:02.730 and we've got all of our different reviews. 02:02.730 --> 02:06.734 One thing we could also do is we could import pandas 02:06.734 --> 02:08.820 and then we could have a date frame 02:08.820 --> 02:13.820 and we could zip up the reviews with the classifications. 02:14.940 --> 02:16.230 And so what you'll see here now 02:16.230 --> 02:18.090 is we are just gonna import pandas 02:18.090 --> 02:19.650 and this is gonna make a data frame 02:19.650 --> 02:21.330 where for each of the reviews, 02:21.330 --> 02:23.520 we're then gonna zip that up with the classifications 02:23.520 --> 02:24.450 that came back. 02:24.450 --> 02:26.130 And now you can see we've got an easy 02:26.130 --> 02:28.740 nice pandas data frame that we can save, 02:28.740 --> 02:30.030 so I'm gonna save that, 02:30.030 --> 02:31.290 and then we've got this data frame 02:31.290 --> 02:32.820 that we can then manipulate. 02:32.820 --> 02:34.530 And then we can say, this is the review 02:34.530 --> 02:35.940 and this is the classification. 02:35.940 --> 02:38.340 And we could do things like count the number of, 02:39.360 --> 02:41.040 count the number of these classifications, 02:41.040 --> 02:44.040 so we can see that we had a total of 32 classifications. 02:44.040 --> 02:46.500 We could do various different types of operation in pandas. 02:46.500 --> 02:49.110 But the point I want to stress across to you is 02:49.110 --> 02:51.540 think about the way that you can use ChatGPT 02:51.540 --> 02:53.250 to label some of your data. 02:53.250 --> 02:57.090 And you can use this on text, you can use this on JSON data, 02:57.090 --> 02:59.310 anything that you want in terms of that. 02:59.310 --> 03:01.140 And it's also possible to use the Vision API 03:01.140 --> 03:02.610 with this type of information as well. 03:02.610 --> 03:06.780 So think about using ChatGPT as a classification engine, 03:06.780 --> 03:09.303 which you can use to get structured labeled data.