Google

The log-probability of an ordered list or ranked sequence of preferences, $$\mathring{Y}$$, given an input $$\mathbf{x}$$, can be defined as the sum of the conditional log-probabilities at each stage of selection. For an ordered list $$\mathring{Y}$$ given by $$\mathbf{y}_{j_1} \succ \mathbf{y}_{j_2} \succ \cdots \succ \mathbf{y}_{j_N}$$, the overall log-probability is expressed as:
$$ \log \Pr(\mathring{Y} | \mathbf{x}) = \log \Pr(\mathbf{y}_{j_1} \succ \mathbf{y}_{j_2} \succ \cdots \succ \mathbf{y}_{j_N} | \mathbf{x}) $$
$$ = \log \Pr(\mathbf{y}_{j_1} | \mathbf{x}, \{\mathbf{y}_{j_1},\mathbf{y}_{j_2},...,\mathbf{y}_{j_N}\}) + \log \Pr(\mathbf{y}_{j_2} | \mathbf{x}, \{\mathbf{y}_{j_2},...,\mathbf{y}_{j_N}\}) + \cdots + \log \Pr(\mathbf{y}_{j_N} | \mathbf{x}, \{\mathbf{y}_{j_N}\}) $$
$$ = \sum_{k = 1}^{N} \log \Pr(\mathbf{y}_{j_k} | \mathbf{x}, \mathring{Y}_{\ge k}) $$
where $$\mathring{Y}_{\ge k}$$ represents the subset of the list of outputs that remain unselected at the $$k$$-th stage, i.e., $$\mathring{Y}_{\ge k} = \{\mathbf{y}_{j_k},...,\mathbf{y}_{j_N}\}$$.

Log-Probability of a Ranked Sequence

In a sequential ranking process, the notation $$\mathring{Y}_{\ge k}$$ represents the set of items that are available for selection at step $$k$$. This set consists of the item chosen at step $$k$$ and all subsequent items that have not yet been ranked. Formally, it is defined as: $$ \mathring{Y}_{\ge k} = \{\mathbf{y}_{j_k}, \dots, \mathbf{y}_{j_N}\} $$

Set of Remaining Items in a Ranked Sequence

The Plackett-Luce loss function is derived directly from the log-probability of a ground-truth ranked sequence. It is defined as the negative log-likelihood of observing this correct sequence, a formulation rooted in the principle of maximum likelihood estimation. The training objective is to minimize this loss, which effectively maximizes the model's probability of predicting the correct rankings. This loss is typically averaged over all samples in a dataset by taking the expectation of the negative log-probability.

Plackett-Luce Loss Function

A language model is tasked with ranking a set of three candidate responses {R1, R2, R3} for a given prompt. The model has computed the following conditional log-probabilities for selecting the best response from a given set of available options:

*   `log Pr(R1 | {R1, R2, R3}) = -0.6`
*   `log Pr(R2 | {R1, R2, R3}) = -1.1`
*   `log Pr(R3 | {R1, R2, R3}) = -1.5`

*   `log Pr(R2 | {R2, R3}) = -0.4`
*   `log Pr(R3 | {R2, R3}) = -1.3`

*   `log Pr(R1 | {R1, R3}) = -0.2`
*   `log Pr(R3 | {R1, R3}) = -1.8`

*   `log Pr(R1 | {R1, R2}) = -0.8`
*   `log Pr(R2 | {R1, R2}) = -0.9`

Note that the log-probability of selecting the single remaining item from a set of one is always 0 (e.g., `log Pr(R3 | {R3}) = 0`).

Based on these values, calculate the total log-probability for the specific ranked sequence `R1 > R3 > R2`.

Ranked Sequence Log-Probability Calculation

A model is designed to rank three candidate responses {A, B, C} for a given prompt. The model determines the final ranking by making a sequence of choices. At each step, it computes the probability of selecting the best response from the set of options not yet ranked.

Given the following conditional probabilities for the sequence A > B > C:
- The probability of selecting A first from the set {A, B, C} is 0.5.
- The probability of selecting B second from the remaining set {B, C} is 0.7.
- The probability of selecting C third from the final set {C} is 1.0.

Calculate the total log-probability (using the natural logarithm, ln) for the complete ranked sequence A > B > C. Provide your final answer rounded to two decimal places and show the summation formula you used.

Calculating Log-Probability for a Ranked List

A model is designed to rank a set of three documents {Doc A, Doc B, Doc C} for a given user query. To calculate the log-probability of the specific ranked sequence 'Doc A > Doc B > Doc C', a developer proposes calculating the total log-probability as the sum of the log-probabilities of each document being chosen first from the full set of three documents. Why is this approach fundamentally flawed for modeling a sequential ranking process?

Learn Before

Related