For an upcoming project I'm trying to write a text classifier for the IMDb sentiment analysis dataset. This needs to vectorize words using an embedding layer and then reduce the dimensions of the output with global average pooling. This is proving however to be very difficult for my low experience level, and I am struggling to wrap my head around the dimensionality involved, bearing in mind I must avoid libraries such as tensorflow that would make it very basic exercise. I am hoping that I could make it easier by encoding each word in the reviews as a one-hot vector, and passing it through a few regular dense layers. Would this work and yield decent results?
One hot encoding is a good strategy to apply with categorical variables that assume few possible values. The problem with text data is that you easily end up with corpora with a really large vocabulary. If I remember correctly the IMDb dataset contains around 130.000 unique words, which means that you should create a network with an input matrix of size 130.000 x max_length where max_length is the fixed maximum length allowed for each review. Apart from the huge size, this matrix would also be extremely sparse, and that's another big issue in using one-hot encoding with text.
For these reasons, I really doubt you would achieve any good results with a simple one-hot encoding. Embeddings where actually designed precisely to overcome all these issues, they have fixed reasonable size, they assume continue values between 0 and 1, which is desirable for deep neural networks, and they can be treated as "extra" trainable weights of a network.
If you really want to avoid embeddings I would suggest you to use (or implement, I don't think it will be so hard) a term frequency–inverse document frequency vectoriser. It is closer to one-hot encoding in the fact that it is based on the creation of a huge co-occurances matrix between words, but at least the values are continuous and not dichotomous. Nevertheless I would not expect high performances with the tf-idf either, simply because this type of encoding works best with shallow models like the Naive Bayes rather than deep models.