Hello, this is Kane. Today, I would like to explain a package called Scikit-Learn LLM instead of discussing LangChain. This package utilizes LLM to help make NLP analysis a bit easier. By using Scikit-Learn LLM, you can easily perform key themes in text analysis, such as sentiment analysis, sentence summarization, and even sentence translation. I think it would be good to mention that this package makes all of these tasks quite simple.

Now, I will explain in more detail how these features can help us. First, Scikit-Learn is a representative package in machine learning, known as Scikit-Learn. By combining the capabilities of Scikit-Learn and LLM, it serves as a package that provides assistance in analyzing text more conveniently and quickly. Therefore, it can be seen as a tool that aids in classification analysis that was previously conducted using Scikit-Learn.

In addition, the combination of vectorization and tasks required for text analysis, such as LLM, can make text modeling easier and at a higher level. Therefore, it is meaningful. Let’s take a closer look at how we conducted text analysis previously with Scikit-learn. 

First, when performing text classification analysis with Scikit-learn, the first thing we need to do is the preprocessing step to prepare the text we have in a format suitable for the classification model. As you can see here, the stemming process is involved, and there is also the assumption of vectorization using a count vectorizer. Most of the time, when vectorizing text, we use the TF-IDF vectorization method. We can utilize the feature extraction text package from Scikit-learn for this purpose.

The text will undergo a process of converting into numerical values. Once we refine the text we have and complete the conversion process, the final step involves training a classification analysis model. The model we train could be a logistic classifier or combined with well-known classification analysis models like a random forest classifier. When we do this, the part that requires the most time and effort is the text preprocessing.

However, to perform text classification analysis through a cycle run, we can significantly simplify the process using Scikit-learn. With just a line or two of code, we can achieve the classification analysis. When we use Scikit-learn for text classification analysis, running the code below will enable the classification.

The analysis is completed. So, in the Scikit-Learn model, we import the zero-shot GPT classifier from GPT classification. Based on this, we input the data, and by placing predictions directly into the X data, we obtain our classification analysis results as labels. You can think of it as eliminating all the previous processes we went through. Therefore, it is a significant advantage of Scikit-Learn that we can utilize classification without text preprocessing and text vectorization. This means that using Scikit-Learn makes the process much easier and less labor-intensive than before. However, there are still some drawbacks. When we previously ran our deep learning models for NLP, there were relatively fast models available. This is mainly because L&M models have significantly fewer parameter sizes.

LLM must have at least 7 billion parameters to perform these tasks. However, previously, models like BERT and others have shown that such models, which are specialized for downstream tasks, do not need to have such a large parameter size. They typically do not exceed several billion parameters; even models with hundreds of millions or tens of millions of parameters can perform these tasks. Therefore, the combination of L&M is significantly faster in performing classification and analysis compared to using a library like Scikit-learn. 

Now, regarding what is possible with Scikit-learn, we can consider tasks like sentiment analysis, news category classification, and text classification analysis, as well as text summarization, such as writing article summaries. Lastly, text translation is also possible, as we generally do with chat models.

When we perform tasks in PyTorch, they are executed more precisely through Scikit-learn. For instance, when we analyze text sentiment, the reason we need to use Scikit-learn is that when we provide a sentence and ask it to classify the sentiment, we input a prompt. However, if the prompt is not well-formulated, the output of the sentiment analysis may not only present the sentiment but also include unnecessary lengthy explanations about the context behind the sentiment. Therefore, when we need to conduct sentiment analysis on many texts and reviews, if we require the output to consist solely of the sentiment for each word, we can utilize Scikit-learn effectively.

It is much easier to analyze. In the case of text summarization, there are many summarization models available, including extractive summarization models and generative summarization models, as well as summaries generated through BART. Compared to those, I think GPT summarization performs very well. Personally, I believe this because when a document comes in, existing models may struggle to summarize new data if it is not part of their training dataset. This is because they do not have as much knowledge as large language models. On the other hand, large language models that have been trained on an enormous amount of text data can summarize any text very effectively. Therefore, we can easily use these features in Psyche. Let's practice and learn this, and finally, we will also cover text translation.

We want to translate a large volume of documents and overseas texts into Korean all at once. In this case, we can easily utilize Cycle LM. Now, let's move on to the practical demonstration. 

First, we will install Cycle LM. Once the installation is complete, the next step is to retrieve our OpenAI API key and set it up. You should access the Cycle LM configuration and set the OpenAI key with the API key we received. There is a section labeled OpenAI org, which you can find by going to the OpenAI API page. From there, click on settings, and when you go to the organization section, you will receive an organization ID even if you registered as an individual.

You can simply copy this and paste it into the values we set here. By doing this, you will see how we have configured the L&M model using the OpenAI API key. If, for instance, I want to use a different API, like the Gemini API, or if I want to work with an open-source model I have, you can visit the SciKit-Learn GitHub page to learn how to load your own model here or how to use the Gemini API. If you need further assistance, feel free to leave a comment, and I will respond to you.

First, we will conduct a classification analysis, and we will start with one of the most common tasks: sentiment analysis of reviews. I have gathered 12 example movie reviews for this purpose. Let's analyze the sentiment.

When conducting an analysis, we can categorize sentiments into positive and negative. I want to demonstrate how well the LLM performs, particularly when using Psyche LM. Since LLMs are integrated, I will show you how they can make high-difficulty tasks easier compared to previous methods. To do this, I will classify sentiments into six categories and ask you to try it out. What you need is to call our zero-shot GPT classifier. You can proceed with your own review texts. I will compile a list of reviews containing a specific emotion and run the GPT classifier on it. Please declare the zero-shot GPT classifier and the clf Pro, and then go through the fitting process. The reason we include fitting in zero-shot classification is that we want to set it up so that it provides these outputs even though we are not training it on specific data.

You can think of it this way. So, regardless of what we put in, we set it to output values like these. I will instruct it to classify and analyze these movie reviews into one of six emotions: joy, surprise, fear, sadness, and anger. Once we have completed the fitting with the clf Pro settings, we can proceed to the prediction process. Now, let's perform sentiment classification on the review data we have. I will provide 12 reviews, and you can see the loading process happening. Through these 12 API calls, we will perform sentiment analysis. It has now been completed, and at this moment, I have not set which L&M model to use below. So, by default, GPT-3.5 is set.

Now, let's review the results of the classification analysis conducted with GPT-3.5. As you can see, the first review expressed a feeling of happiness, and the sentiment was well classified as joy. Additionally, there was a comment about being surprised by an unexpected plot twist in the movie, which was also categorized as surprise. I can see that when I classify and analyze these simple emotions, it performs very well. Compared to the time it takes when we use models from Hugging Face or when we analyze with fine-tuned models specialized for our dataset, we can analyze sentiments much faster in this way. The advantages of Scikit-learn are clearly evident in this regard. First, we have classified one emotion.

I conducted a classification analysis, and secondly, I will perform a multi-label classification task when there are multiple emotions present in these sentences. Again, I will utilize the multi-label zero-shot GPT classifier in the same way as before. This time, we have structured an example of a review for sentiment analysis in this manner. In a single sentence, there are various emotions present. For instance, you can see joy, surprise, and being touched. In the middle, there are also elements of fun and fear. This shows that a single sentence does not contain just one emotion but multiple emotions mixed together. I believe this kind of multi-emotion analysis can be very useful in practical sentiment analysis work.

We do this because whether we analyze reviews or the words of any individual, it is very rare for a sentence to be completely composed of a single sentiment. Therefore, it would be beneficial to extract all the sentiments that a sentence possesses. By doing this, we can diversify the direction of our projects. So, this time we will proceed with multi-label classification. We will retain the existing emotions, and this time we will work with mixed reviews. We will use a multi-label zero-shot GPT classifier without any additional training. For this, we will utilize the GPT-4 model. While I believe GPT-3.5 would perform well enough, given that we are dealing with multi-sentiment analysis, I think it’s essential to have a model that understands context better, which is why I will bring in GPT-4. And Max...

I set the label to 3, and if I set the max label to 4, I can extract up to four emotions. After that, I declare the classifier. This time, we go through the same fitting process, but I will wrap it with the emotion labels list that has been saved. Since we will only output one result when we provide a multi-label response, it gets classified as an emotion and outputs that single result. However, this time we are doing multi-emotion analysis, so when we output multiple emotions, it will be displayed as a list. Therefore, you can think of it as wrapping the emotion labels in a list for the output. I will save it as a label again and let's take a look at the results. Since I used GPT-4 this time, you may notice that it takes a bit longer compared to before.

There are existing deep learning models that we are familiar with, which can accomplish this much faster. Therefore, I believe it is important to understand that there is definitely a disadvantage in terms of time when utilizing this. Now that we have completed this, let's check if the review was analyzed correctly for sentiment. First, we see that joy, surprise, and emotion are included, and it appears that joy, surprise, and emotion have been classified well. There were instances where surprise was felt, but it was classified as anger, and it handled those surprises well. There were also scary scenes that were surprising. However, this classification analysis has some inaccuracies. While there were scary scenes, when the protagonist's cheerful attitude was described as surprising, the emotions of the person writing the review were both scared and surprised. So, it’s not that I was cheerful, but rather that the protagonist was.

It was surprising how well it worked. I should have extracted only the emotion of surprise, but when analyzing the sentiment based on this sentence, you can see that fear, excitement, and surprise were all extracted. In fact, no matter how good our model is, it's very difficult to create a model that can accurately check for such nuances during sentiment analysis. Despite that, I hoped that GPT-4 would handle this better, but it seems that it still struggles with these aspects. It might be due to the Korean language; if it had been done in English, it might have performed better. These are areas that need to be filtered out while using the model, so it would be good to make adjustments as you utilize it. Next, we will try text summarization. For text summarization, we will use the summarize function from the GPT text-to-text model.

Now, instead of defining the dataset myself, I will use an example dataset provided by Scikit-learn. I will save the dataset as a get ration dataset and declare GPT summarize, utilizing the GPT-3.5 Turbo model this time. Interestingly, there is a parameter called max words, which allows you to specify the maximum number of words. For instance, if I set it to 15, the prompt will generate a summary consisting of a maximum of 15 words. Once declared, I will define it with the required parameters, and then I will apply the summarize transform function to the dataset I have. I will input my dataset into the summarize function to output the summary. Let's output the first original text from the example dataset and its corresponding summary. The original text will look something like this.

You can see that it took some time to summarize the total number of sentences in this three-line sentence. It has now been completed, and you can see the summary of this three-line sentence condensed into a single line. It would be interesting to check if we adhered to the defined maximum words. One, two, three, four, five, six, seven, eight, nine, ten, eleven, twelve, thirteen. We can see that we did not exceed the defined maximum of 15 words. In fact, while it is possible to use existing deep learning models for a maximum of 15 words, it is a bit more convenient for us. We can utilize it without needing to adjust hyperparameters or reconfigure the existing architecture. This ability to extract such results with just a few lines of code is very appealing.

It seems that many of you may already have some expectations about what we are currently conducting, which is sentiment analysis. This summary can be viewed as a package designed to receive specific types of responses by passing it through a GPT prompt template. For example, in the case of sentiment analysis, if we look at the zero-shot GPT classifier function, the prompt template could involve applying the fit function to classify the first set based on the second value. Therefore, when I input the reviews I have and run the predict function, I will input one review and use the predetermined prompt template to conduct the classification analysis on that review. However, the label must be one of the six options provided in the prompt template.

When we do that, we create an output that matches one of the six words using tools like a list parser. You can think of it as producing the results we are currently observing. The summary is the same. We provide the original text and input a prompt template. The GPT summarize prompt template will ask to summarize the given text without exceeding the maximum word limit. When we do that, the result will be outputted as a transformed version. I think that's how you should understand it. Currently, we have a total of ten original texts, and let's take a look at all of them. If we look at the second sentence, it seems to be a single line with three lines and fewer than fifteen words. If we count, it appears to be one, two, three, four, five, six, seven.

We went over 15, but we still tried our best to write a summary according to the guidelines we set. It's a significant advantage to have a tool that allows us to create summaries so easily. For example, many of you might be familiar with Naver's summarization bot. When you enter an article, you may have seen a summary written in one to three lines. To create something like that, the entire text of the article needs to be processed by a deep learning model, which then displays the summary. If this process can be simplified and run in such a straightforward manner, it would save a lot of resources. In that sense, it seems to be a very good library. Finally, let's try some translation work.

I will proceed with the text translation. This time, I will use GPT to translate text directly. We have brought in an example dataset for the Psyche RAM, and we will use this example dataset for translation. When defining the GPT translator, we will utilize the 3.5 turbo model and set the output language to Korean, asking it to translate into Korean.

After that, when we transform the original text using the GPT translator, we will create a system where the translated sentences are stored as Korean text. If we check the output, we can see that it has been translated well into Korean. Since the example dataset is not in English, I am not sure if it has been translated correctly, but anyway, whenever I have a sentence, I will...

그대로 어 리스트로 담은 다음에 여기다가 넣어 주면 되겠죠. 내가 가진 영어 문서들이나 영어 리뷰들을 번역하고 싶다면, GPT 트랜슬레이터를 선언한 다음에 트랜스폼 안에 리스트 형태로 담아서 그걸 출력해 보면, 저희가 번역을 그대로 활용할 수 있다고 볼 수 있겠습니다. 

그러면 오늘은 사이킷을 통해서 간단하게 NLP 분석을 수행할 수 있는 실습을 진행해 봤는데요. 여러분들이 실무에서 진행하시면서 NLP 작업으로 굉장히 고생을 많이 하셨을 텐데, 사이킷으로 조금 더 쉽게 접근해 보시고 다양한 이점을 많이 가져가셨으면 좋겠습니다. 오늘도 강의 들어 주셔서 감사하고요. 댓글로 또 듣고 싶은 강의가 있으시면 댓글 달아 주시면 제가 다음 강의에 한번 넣어 보도록 하겠습니다. 감사합니다.

LLM으로 NLP 분석을 더욱 쉽게, Scikit-LLM

LLM으로 NLP 분석을 더욱 쉽게, Scikit-LLM

스크립트