AI-ordlista
Den kompletta ordlistan över AI
Conditional Random Field
Discriminative statistical model used for structured prediction, which directly models the conditional probability P(y|x) of labels y given observations x.
Potential Factor
Non-negative function that quantifies the affinity between variable configurations in a CRF, transformed into probabilities by normalization.
Energy Function
Function that assigns a scalar value to each possible variable configuration, where low energy configurations correspond to high probabilities.
Linear-Chain CRF
Specialized type of CRF where output variables form a linear chain, widely used for sequence labeling like NER or POS tagging.
Partition Function
Partition constant that ensures the sum of probabilities of all possible configurations equals 1, calculated as the sum of exponentials of energies.
Conditional Likelihood
Objective function maximized during CRF training, measuring the probability of correct labels given training observations.
Message Passing
Inference algorithm that propagates information between nodes in a factor graph to compute marginal beliefs or MAP assignments.
High-Order CRF
Extension of linear-chain CRFs that models dependencies between non-adjacent variables in the sequence, capturing more complex relationships.
Décodage MAP
Processus d'inférence qui cherche la configuration d'états maximisant la probabilité a posteriori (Maximum A Posteriori) dans un CRF.
CRF-Semi-Cache
Variante de CRF où les états peuvent être partiellement observés, combinant apprentissage supervisé et non-supervisé pour améliorer la performance.
Facteur de Transition
Composant du CRF qui modélise la probabilité de transition entre états successifs dans une séquence, capturant les dépendances temporelles.