Negative sampling is needed for loss function calculation, for example Margin Ranking Loss 

Negative sampling restores the missed negative edges in the graph, thus, improves the accuracy of the model

Negative sampling normalizes the embeddings of relations which improves the robustness of the model

Negative sampling is mostly used as data augmentation for better training convergence

probabilities of initial state emission

probabilities of moving between states

observation likelihoods

HMM

Decoding

Encoding

Learning

The efficiency of training algorithms in generating text

The complexity of a language model's architecture

The degree of uncertainty of a language model on unseen data

107

106

105

108

Window size

None of the above

Both of the above

Weights of the neural network

Antisymmetric, 1-N

Symmetric, 1-N

Symmetric, Composition

Antisymmetric, Inverse 

Hyponymy

Meronymy

Antonymy

Holonymy