Google

In a multi-round dialogue with $$K$$ turns, the generation of a response $$\mathbf{y}^k$$ at any given round $$k$$ is conditioned on the entire preceding conversational history. This history includes all prior user requests and model responses up to the current request. For a conversation with sequence $$\mathbf{x}^1, \mathbf{y}^1, \dots, \mathbf{x}^K, \mathbf{y}^K$$, the conditional log-probability of generating the $$k$$-th response is expressed as: $$ \log \mathrm{Pr}_{\theta}(\mathbf{y}^k|\mathbf{x}^1, \mathbf{y}^1, \dots, \mathbf{x}^k) $$. This value is a key component in defining the overall training objective for dialogue models.

Conditional Log-Probability of a Response in Multi-Round Dialogue

The training objective for a multi-round dialogue model is to find the optimal parameters, $$\tilde{\theta}$$, that maximize the cumulative log-probability of all model responses across the entire conversation. For a $$K$$-round dialogue, this is achieved by summing the conditional log-probabilities for each response, where each response is conditioned on the full conversational history up to that point. The formal optimization problem is: $$ \tilde{\theta} = \arg\max_{\theta} \sum_{k=1}^{K} \log \mathrm{Pr}_{\theta}(\mathbf{y}^k|\mathbf{x}^1, \mathbf{y}^1, \dots, \mathbf{x}^k) $$. A straightforward implementation of this objective involves calculating the conditional probability for each of the $$K$$ turns separately, but it requires running the language model $$K$$ times.

Training Objective for Multi-Round Dialogue Models

A dialogue model is engaged in a conversation that has reached its third turn. The full history of the conversation so far is `{x_1, y_1, x_2, y_2, x_3}`, where `x_k` is the user's input and `y_k` is the model's response at turn `k`. Write the mathematical expression for the conditional log-probability that the model uses to generate its next response, `y_3`.

Formulating Conditional Probability in Dialogue

Based on the principle of how a response at any given turn is generated based on the preceding conversational history, analyze the provided dialogue snippet. Identify the specific turn where the model's response indicates a failure and explain which components of the conversational history were likely ignored during the generation of that faulty response.

Diagnosing Conversational Memory Failure

Consider a dialogue model engaged in a three-turn conversation, represented by the sequence `{x_1, y_1, x_2, y_2, x_3, y_3}`, where `x_k` is the user's input and `y_k` is the model's response at turn `k`. When the model calculates the probability of generating the third response (`y_3`), how does the set of information it conditions on relate to the set of information used to calculate the probability of the second response (`y_2`)?

Learn Before

Related