It decreases perplexity by providing clearer sentence boundaries

It has no impact on the perplexity calculation

It increases the perplexity by introducing additional tokens in the vocabulary

We do not know how exactly, it only adjusts the total count of word tokens considered in the perplexity calculation

A syntactic parsing method focusing on triplets of words

A probabilistic model based on three-word combinations

A sequence of three words occurring together in a text

A non-projective tree can be drawn with no crossing edges

Incompleteness of all treebanks automatically derived from phrase-structure treebanks

transition-based

graph-based

both

As a target in the loss function

As an evaluation metric

As an embedding

No

BPE tokenizer is not a machine learning model, so it cannot be trained

Yes

Can be used only with CRF, because CRF is discriminative 

Can be used with both types of PoS models

Can be used only with HMM, because HMM is a generative model

No

It depends

Yes

Permanent changes in natural languages

Natural languages have a regular nature

NLP requires large corpora