How to retrain classification models

I have training data that is a set of transaction descriptions and the category of that transaction.

I am trying to build a spend analyzer using this information to help me classify my transactions into categories. I built a model by training a custom word2vec model on my data and got good accuracy with it.

The problem is that every few weeks, as I get newer data with newer words, the quality of my predictions starts dropping. I could retrain my model but I’ll have to manually sit and label these transactions to get an updated training set. Labeling the transactions manually seems the opposite of what I am trying to achieve and takes more effort than building the model.

So my question is how do I retrain my model with newer descriptions of transactions without actually having the true labels?