ChatGPT applications (NLP perspectives)

ChatGPT, again?

I know, I know. As much as I hate to have a say with what everybody is talking about, I'm only going to talk about on the topic, NLP.

ChatGPT and NLP

As you know, ChatGPT or simply GPT for that matter, is based on a (very) large language model. And ChatGPT truly opened the door for many opportunities for greater public (including companies like Microsoft or NVIDIA).

At its core is the resourceful language context and as a NLP practitioner, I've wondered if it really helps me with routine text analytics I had implemented with painstaking preprocessing (NER, News type classification and so on).

The short answer is: Absolutely!

In fact, its outcome was surprisingly good enough to be skeptical of myself for the time I had spent for those tasks.

While not perfect nor free to use (my API credits have expired without me being noticed), ChatGPT clearly satisfies its core colleagues, text practitioners.

Selected applications

Among every possible application (too many if I had to mention all), I found the following relevant to my area of expertise and interest.

Topic Classification (BERTopic)

These days, I use BERTopic most frequently (on Google Colab). With version 0.14.0, BERTopic started to support ChatGPT in its topic representations. By this addition, users can enhance the resultant topics into more sensible sentences.

This is the easiest thinkable application of ChatGPT without much effort. It says it supports ChatGPT and you can just take advantage of it.

Extracting company names (NER)

This is one of the classic NLP examples which require context awareness. The easiest example would be Apple and apple in pretty different contexts.

Apple launched newer version of Airpods a few weeks ago.
An apple is on the table.

I admit that this may be a too simple example. Anyway, the point is that without a context, we aren't so sure about the validity of NER.

This is where ChatGPT shines too. LLMs have been trained with plenty of texts with their contexts, therefore they are able to distinguish better which are companies of interests and others aren't.

Filtering out promotional news

My day job is to analyze news text. Many of the articles I come across are not purely informative or do not serve my purpose.

In Korean articles, though my experience, higher percentage among them are promotional news for specfic companies compared to news in English.

As before, an easy task for human is not for a computer without considering the context.

However, ChatGPT can do the job easier albeit not entirely as impressive as the previous task.

Closing

Currently, I am in the stage of experiment for these tasks leveraging ChatGPT.

While it makes our job easier, one has to keep in mind that due to its high demand and rate limit, processing large volume of data might not be ideal. NER task is especailly burdensome as it usually has to read the whole text to know which to extract.

Even if we can bear the time delay using ChatGPT, how about the bill?

This brilliant tool clearly has a lot of advantage to be taken advantage of, however, how to leverage it in the best way is left to users.