We sometimes encounter ambiguous letters when reading handwritten words, but we can still interpret the words. For example, the same shape can be interpreted as an A in CAT but an H in THE. At what level of analysis does the feature net resolve this issue?

The bigram level

Explanation:

The bigram level of analysis is an example of the N-gram model(as we can also have trigram). It is used in statistical language models to calculate probability and interprete letters in words based on previous occurrence(preceding word). In other words a bigram(less commonly called digram) makes prediction using conditional probabilities that are based on previous word. A tigram would do just same thing but predicts based on two preceding words.