How to embed Sequence of Sentences in RNN?

Question

I am trying to make a RNN model (in Pytorch), that takes couple of sentences and then classifies it to be either Class 0 or Class 1.

For the sake of this question let's assume that the max_len of the sentence is 4 and max_amount of time steps is 5. Thus, each datapoint is on the form (0 is a value that used for padding padded value):

    x[1] = [
    # Input features at timestep 1
    [1, 48, 91, 0],
    # Input features at timestep 2
    [20, 5, 17, 32],
    # Input features at timestep 3
    [12, 18, 0, 0],
    # Input features at timestep 4
    [0, 0, 0, 0],
    # Input features at timestep 5
    [0, 0, 0, 0]
    ]
    y[1] = [1]

When I have just one sentence per target: I simply pass each word to the embedding layer and then to the LSTM or GRU, but I am a bit stuck on what to do when I have a sequence of sentences per target?

How do I build an embedding that can handle sentences?

Answer 1

The simplest way is to use 2 kinds of LSTM.

Prepare the toy dataset

xi = [
# Input features at timestep 1
[1, 48, 91, 0],
# Input features at timestep 2
[20, 5, 17, 32],
# Input features at timestep 3
[12, 18, 0, 0],
# Input features at timestep 4
[0, 0, 0, 0],
# Input features at timestep 5
[0, 0, 0, 0]
]
yi = 1

x = torch.tensor([xi, xi])
y = torch.tensor([yi, yi])

print(x.shape)
# torch.Size([2, 5, 4])

print(y.shape)
# torch.Size([2])

Then, x is the batch of inputs. Here batch_size = 2.

Embed the input

vocab_size = 1000
embed_size = 100
hidden_size = 200
embed = nn.Embedding(vocab_size, embed_size)

# shape [2, 5, 4, 100]
x = embed(x)

The first word-LSTM is to encode each sequence into a vector

# convert x into a batch of sequences
# Reshape into [2, 20, 100]
x = x.view(bs * 5, 4, 100)

wlstm = nn.LSTM(embed_size, hidden_size, batch_first=True)
# get the only final hidden state of each sequence

_, (hn, _) = wlstm(x)

# hn shape [1, 10, 200]

# get the output of final layer
hn = hn[0] # [10, 200]

The second seq-LSTM is to encode sequences into a single vector

# Reshape hn into [bs, num_seq, hidden_size]
hn = hn.view(2, 5, 200)

# Pass to another LSTM and get the final state hn
slstm = nn.LSTM(hidden_size, hidden_size, batch_first=True)
_, (hn, _) = slstm(hn) # [1, 2, 200]

# Similarly, get the hidden state of the last layer
hn = hn[0] # [2, 200]

Add some classification layers

pred_linear = nn.Linear(hidden_size, 1)

# [2, 1]
output = torch.sigmoid(pred_linear(hn))