Question: mean aggregation in Encoder

[yichengchen24]

Thanks for the excellent work!

However, I have a question about the mean aggregation operation in Encoder: why divide by the length of the chunked sentence len(chunked_instances[index]) instead of the number that the sentence is chunked to?

The corresponding code is

# aggregate the embeddings
embeddings = np.zeros((len(sentences), self.hidden_size))
for i, index in enumerate(chunked_indices):
    if aggregate_method == 'mean':
        embeddings[index] += chunked_embeddings[i] / len(chunked_instances[index])
    elif aggregate_method == 'sum':
        embeddings[index] += chunked_embeddings[i]
    elif aggregate_method == 'max':
        embeddings[index] = np.maximum(embeddings[index], chunked_embeddings[i])