Learn Before
Flaw in a Naive Inference Strategy
During the inference phase for a question-answering model, a common but incorrect approach is to independently find the token with the highest probability of being the start of an answer and the token with the highest probability of being the end of an answer, and then combine them to form the final span. Explain why this approach is flawed and could lead to an invalid or nonsensical answer.
0
1
Tags
Ch.2 Generative Models - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Analysis in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
Span Prediction Inference Formula
Identifying the Optimal Answer Span
A language model has processed the context 'The capital of France is Paris.' and produced the following probabilities for each token being the start or the end of an answer span. To determine the most likely answer, you must find the start and end token pair that yields the highest combined score (calculated as start_probability * end_probability), with the constraint that the start token cannot appear after the end token. Given the table below, which span is the most likely answer?
Token Index Start Probability End Probability 'The' 1 0.05 0.05 'capital' 2 0.10 0.05 'of' 3 0.05 0.05 'France' 4 0.20 0.10 'is' 5 0.05 0.05 'Paris' 6 0.50 0.60 '.' 7 0.05 0.10 Flaw in a Naive Inference Strategy