Word embedding as input or raw text? I'm trying to implement a neural network for text recognition and I'm a little bit confused about text inputs. 
The goal of the network is to classify a comment, toy example:
I like that -> GOOD
I don't like that -> BAD

For now I'm tokenizing my comments as:
[i, like, that]
[i, dont, like, that]

and my questions are:


*

*Should I input that list of words into the network? how?? one-hot encoding?

*I have read and tried Word2Vec but once I have my model should I translate the comment as a list of vectors? I mean:


In my Word2Vec model imagine that my 'i' is [0.23, 0.45, 0.1], 'like' is [0.67, 0.15, 0.98] and 'that' is [0.43, 0.25, 0.72], My first input (which were [i, like, that]) should be now [[0.23, 0.45, 0.1],[0.67, 0.15, 0.98],[0.43, 0.25, 0.72]]?
In that case, I should be using Word2Vec to reduce dimensionality, right? so if with one-hot enconding I had a matrix of WordNumber * WordNumber now I should create vectors  in order to have WordNumber * Vectorlength, right?   
 A: 
Should I input that list of words into the network? 

If you want to. It depends on what your goals are. If all you care about is maximizing performance on some hold-out set, then perhaps you should do an experiment to determine which method works best.

how?? one-hot encoding?

Yes.

I have read and tried Word2Vec but once I have my model should I translate the comment as a list of vectors?

Usually recurrent neural networks (rnn, lstm, gru) are used in this scenario because they can naturally work with variable-length sequences.
Alternatively, you can just train the word embeddings simultaneously with the rest of the model. Pre-trained embeddings aren't strictly necessary, although if your training data is limited then maybe pre-trained embeddings could generalize better.
A: I would use pretrained word embeddings with a neural network as Sycorax mentioned (recurrent or transformer) if you're looking for a good performance as state of the art models for sentiment analysis are nn-based nowadays.
You need to use the lookup index (word2idx) of the pretrained word embeddings to transform each word into an index then feed it into an embedding layer nn.Embedding in the case of pytorch for example. This layer would be initialized with the pretrained word vectors.
You can also have a look at using a pretrained language model like Bert for your use case which could help you overcome the small size of your dataset https://github.com/huggingface/pytorch-transformers
