I'm currently analyzing Korean social media text. The below are the steps of the analysis.
- Collect/crawling text data from social media (e.g. Twitter, Facebook), which are related to specific topics.
- Analyze the data by using BERT. This includes text classification and sentiment analysis.
However, I've faced some complicated problems.
In step 1, I searched with some relevant keywords to the topic, but indeed, the collected data are not related to the topic. Some of them are related to the topic but some of them are not. In this case, how can I get these topic-related data? Should I annotate the data and training a model to classify the relevant/non-relevant data? Is there any unsupervised approach such as topic modeling? Is there any method that lets me collect topic-related data?
In step 2, I actually got documents, sentences, or paragraphs rather than just one sentence. Therefore the topics and sentiments are mixed in the documents and I'm really struggling to annotate them. I tried to separate the documents into smaller ones but there are no thresholds to distinguish them since they are social media text. Even if I could, this makes another problem which is the context of the document is disappeared and this kills the topic or sentiments.
Here's the summary:
- Collecting steps: Collect data related to a specific topic
- Problems: The collected data are actually not related to the topic
- Analyzing steps: Analyze (classification) the data by using BERT
- The labels are mixed in the collected data. How can I split them into some meaningful pieces (big enough to maintain the context of the documents, small enough to analyze them easily)