For my BSc thesis I am trying to classify asset price direction (up/down/neutral) using numerical features and Swedish text. The text is short financial news (ca 50 words each) that have sentiment values (in most basic form negative/ positive/neutral). I have in total around 3500 daily observations.
Before diving into the problem initially I thought it would be sufficient to just label the news data and use it as feature in a ensemble machine learning model such as random forest. In parallel I would need to train a text classifier (using bag of words and tfidf) such as SVM in order to classify out-of-sample inputs. However, I discovered that labeling introduces a non-negligible amount of class noise, as some text is more ambiguous and cannot easily be classified (e.g. does slightly positive sentiment fall into the neutral or positive category?). For reference, I don't have access to a sentiment analysis tool for financial news in Swedish that could handle the task for me automatically.
I read about using deep learning models for automatic feature extraction from text and classification, but then comparing this setup performance to regular machine learning models such as RF (which is the goal of my thesis) becomes very problematic since the feature inputs are not the same.
Using the features extracted automatically by tfidf does not seem feasible, as this yields a size 2-3000 vector and I have a small dataset.
Therefore I need a solution for handling the text efficiently and uniformly in order for it to be used as feature for training my DL and ML models. So my question is what is the best way to approach this problem? Sentiment analysis with manual labeling as before? NLP approach? Other?
Your ideas and suggestions are much appreciated.