7.4 Chatbots

Chatbot Tutorial

Created Date: 2025-07-07

In this tutorial, we explore a fun and interesting use-case of recurrent sequence-to-sequence models. We will train a simple chatbot using movie scripts from the Cornell Movie-Dialogs Corpus.

Conversational models are a hot topic in artificial intelligence research. Chatbots can be found in a variety of settings, including customer service applications and online helpdesks. These bots are often powered by retrieval-based models, which output predefined responses to questions of certain forms.

In a highly restricted domain like a company’s IT helpdesk, these models may be sufficient, however, they are not robust enough for more general use-cases. Teaching a machine to carry out a meaningful conversation with a human in multiple domains is a research question that is far from solved. Recently, the deep learning boom has allowed for powerful generative models like Google’s Neural Conversational Model, which marks a large step towards multi-domain generative conversational models. In this tutorial, we will implement this kind of model in PyTorch.

7.4.1 Preparations

To get started, download the Movie-Dialogs Corpus zip file.

The next step is to reformat our data file and load the data into structures that we can work with.

The Cornell Movie-Dialogs Corpus is a rich dataset of movie character dialog:

  • 220,579 conversational exchanges between 10,292 pairs of movie characters;

  • 9,035 characters from 617 movies;

  • 304,713 total utterances.

This dataset is large and diverse, and there is a great variation of language formality, time periods, sentiment, etc. Our hope is that this diversity makes our model robust to many forms of inputs and queries.

First, we’ll take a look at some lines of our datafile to see the original format.

7.4.1.1 Create Formatted Data File

For convenience, we’ll create a nicely formatted data file in which each line contains a tab-separated query sentence and a response sentence pair.

The following functions facilitate the parsing of the raw utterances.jsonl data file.

7.4.1.2 Load and Trim Data

Our next order of business is to create a vocabulary and load query/response sentence pairs into memory.

Note that we are dealing with sequences of words, which do not have an implicit mapping to a discrete numerical space. Thus, we must create one by mapping each unique word that we encounter in our dataset to an index value.

For this we define a Voc class, which keeps a mapping from words to indexes, a reverse mapping of indexes to words, a count of each word and a total word count. The class provides methods for adding a word to the vocabulary (addWord), adding all words in a sentence (addSentence) and trimming infrequently seen words (trim). More on trimming later.

Now we can assemble our vocabulary and query/response sentence pairs. Before we are ready to use this data, we must perform some preprocessing.

First, we must convert the Unicode strings to ASCII using unicodeToAscii. Next, we should convert all letters to lowercase and trim all non-letter characters except for basic punctuation (normalizeString). Finally, to aid in training convergence, we will filter out sentences with length greater than the MAX_LENGTH threshold (filterPairs).

Another tactic that is beneficial to achieving faster convergence during training is trimming rarely used words out of our vocabulary. Decreasing the feature space will also soften the difficulty of the function that the model must learn to approximate. We will do this as a two-step process:

  1. Trim words used under MIN_COUNT threshold using the voc.trim function.

  2. Filter out pairs with trimmed words.

7.4.1.3 Prepare Data for Models

7.4.2 Define Models

The brains of our chatbot is a sequence-to-sequence (seq2seq) model. The goal of a seq2seq model is to take a variable-length sequence as an input, and return a variable-length sequence as an output using a fixed-sized model.

Sutskever et al. discovered that by using two separate recurrent neural nets together, we can accomplish this task. One RNN acts as an encoder, which encodes a variable length input sequence to a fixed-length context vector. In theory, this context vector (the final hidden layer of the RNN) will contain semantic information about the query sentence that is input to the bot. The second RNN is a decoder, which takes an input word and the context vector, and returns a guess for the next word in the sequence and a hidden state to use in the next iteration.

7.4.2.1 Encoder

The encoder RNN iterates through the input sentence one token (e.g. word) at a time, at each time step outputting an “output” vector and a “hidden state” vector. The hidden state vector is then passed to the next time step, while the output vector is recorded. The encoder transforms the context it saw at each point in the sequence into a set of points in a high-dimensional space, which the decoder will use to generate a meaningful output for the given task.

At the heart of our encoder is a multi-layered Gated Recurrent Unit, invented by Cho et al. in 2014. We will use a bidirectional variant of the GRU, meaning that there are essentially two independent RNNs: one that is fed the input sequence in normal sequential order, and one that is fed the input sequence in reverse order. The outputs of each network are summed at each time step. Using a bidirectional GRU will give us the advantage of encoding both past and future contexts.

Bidirectional RNN:

Note that an embedding layer is used to encode our word indices in an arbitrarily sized feature space. For our models, this layer will map each word to a feature space of size hidden_size. When trained, these values should encode semantic similarity between similar meaning words.

Finally, if passing a padded batch of sequences to an RNN module, we must pack and unpack padding around the RNN pass using nn.utils.rnn.pack_padded_sequence and nn.utils.rnn.pad_packed_sequence respectively.

Computation Graph:

  1. Convert word indexes to embeddings.

  2. Pack padded batch of sequences for RNN module.

  3. Forward pass through GRU.

  4. Unpack padding.

  5. Sum bidirectional GRU outputs.

  6. Return output and final hidden state.

7.4.2.2 Decoder

The decoder RNN generates the response sentence in a token-by-token fashion. It uses the encoder’s context vectors, and internal hidden states to generate the next word in the sequence. It continues generating words until it outputs an EOS_token, representing the end of the sentence. A common problem with a vanilla seq2seq decoder is that if we rely solely on the context vector to encode the entire input sequence’s meaning, it is likely that we will have information loss. This is especially the case when dealing with long input sequences, greatly limiting the capability of our decoder.

To combat this, Bahdanau et al. created an “attention mechanism” that allows the decoder to pay attention to certain parts of the input sequence, rather than using the entire fixed context at every step.

At a high level, attention is calculated using the decoder’s current hidden state and the encoder’s outputs. The output attention weights have the same shape as the input sequence, allowing us to multiply them by the encoder outputs, giving us a weighted sum which indicates the parts of encoder output to pay attention to. Sean Robertson’s figure describes this very well:

Luong et al. improved upon Bahdanau et al.’s groundwork by creating “Global attention”. The key difference is that with “Global attention”, we consider all of the encoder’s hidden states, as opposed to Bahdanau et al.’s “Local attention”, which only considers the encoder’s hidden state from the current time step. Another difference is that with “Global attention”, we calculate attention weights, or energies, using the hidden state of the decoder from the current time step only. Bahdanau et al.’s attention calculation requires knowledge of the decoder’s state from the previous time step. Also, Luong et al. provides various methods to calculate the attention energies between the encoder output and decoder output which are called “score functions”:

Overall, the Global attention mechanism can be summarized by the following figure. Note that we will implement the “Attention Layer” as a separate nn.Module called Attn. The output of this module is a softmax normalized weights tensor of shape (batch_size, 1, max_length).

7.4.3 Define Training Procedure

7.4.3.1 Masked loss

Since we are dealing with batches of padded sequences, we cannot simply consider all elements of the tensor when calculating loss. We define maskNLLLoss to calculate our loss based on our decoder’s output tensor, the target tensor, and a binary mask tensor describing the padding of the target tensor. This loss function calculates the average negative log likelihood of the elements that correspond to a 1 in the mask tensor.

def maskNLLLoss(inp, target, mask):
    nTotal = mask.sum()
    crossEntropy = -torch.log(torch.gather(inp, 1, target.view(-1, 1)).squeeze(1))
    loss = crossEntropy.masked_select(mask).mean()
    loss = loss.to(device)
    return loss, nTotal.item()

7.4.3.2 Single Training Iteration

The train function contains the algorithm for a single training iteration (a single batch of inputs).

We will use a couple of clever tricks to aid in convergence:

  • The first trick is using teacher forcing. This means that at some probability, set by teacher_forcing_ratio, we use the current target word as the decoder’s next input rather than using the decoder’s current guess. This technique acts as training wheels for the decoder, aiding in more efficient training. However, teacher forcing can lead to model instability during inference, as the decoder may not have a sufficient chance to truly craft its own output sequences during training. Thus, we must be mindful of how we are setting the teacher_forcing_ratio, and not be fooled by fast convergence.

  • The second trick that we implement is gradient clipping. This is a commonly used technique for countering the “exploding gradient” problem. In essence, by clipping or thresholding gradients to a maximum value, we prevent the gradients from growing exponentially and either overflow (NaN), or overshoot steep cliffs in the cost function.

7.4.3.3 Training iterations

It is finally time to tie the full training procedure together with the data. The trainIters function is responsible for running n_iterations of training given the passed models, optimizers, data, etc. This function is quite self explanatory, as we have done the heavy lifting with the train function.

One thing to note is that when we save our model, we save a tarball containing the encoder and decoder state_dicts (parameters), the optimizers’ state_dicts, the loss, the iteration, etc. Saving the model in this way will give us the ultimate flexibility with the checkpoint. After loading a checkpoint, we will be able to use the model parameters to run inference, or we can continue training right where we left off.

7.4.4 Define Evaluation

After training a model, we want to be able to talk to the bot ourselves. First, we must define how we want the model to decode the encoded input.

Greedy decoding is the decoding method that we use during training when we are NOT using teacher forcing. In other words, for each time step, we simply choose the word from decoder_output with the highest softmax value. This decoding method is optimal on a single time-step level.

To facilitate the greedy decoding operation, we define a GreedySearchDecoder class. When run, an object of this class takes an input sequence (input_seq) of shape (input_seq length, 1), a scalar input length (input_length) tensor, and a max_length to bound the response sentence length. The input sentence is evaluated using the following computational graph:

Now that we have our decoding method defined, we can write functions for evaluating a string input sentence. The evaluate function manages the low-level process of handling the input sentence. We first format the sentence as an input batch of word indexes with batch_size==1. We do this by converting the words of the sentence to their corresponding indexes, and transposing the dimensions to prepare the tensor for our models.

We also create a lengths tensor which contains the length of our input sentence. In this case, lengths is scalar because we are only evaluating one sentence at a time (batch_size==1). Next, we obtain the decoded response sentence tensor using our GreedySearchDecoder object (searcher). Finally, we convert the response’s indexes to words and return the list of decoded words.

7.4.5 Run Model

7.4.6 Conclusion

That’s all for this one, folks. Congratulations, you now know the fundamentals to building a generative chatbot model! If you’re interested, you can try tailoring the chatbot’s behavior by tweaking the model and training parameters and customizing the data that you train the model on.

Prev: 7.3 Language Modeling Next: 8.1 Speech Feature Extraction