Suggesting and refining user input based on original user input

Construct a feature vector for each word or term w in a corpus,
including a count for each feature /in the feature vector

I

Determine the value of each feature/in the feature vector as the pointwise mutual information Ml between the word w and feature/

I

Determine a similarity value between two words or phrases w; and w2 as the cosine of the angle between their feature vectors using the values of the features in the feature vectors

Segment queries in query logs into one or more query word sequences that maximize overall probability for the query

I

Determine frequent n-grams (n-word sequences) and count the query word sequences where all adjacent pairs of words in the sequence are frequent n-grams

I

Filter out non-compound or non-phrasal word sequences by requiring a compound/phrase to appear at both the beginning and the end of some queries (but not necessarily in the same query)

I

Construct a feature vector for each n-gram in a corpus,
including a count for each feature/in the feature vector

I

Determine value of each feature/in the feature vector as the point-wise mutual information MI between the n-gram and the feature/

70

71

72

73

74

75

Determine a similarity value between two n-grams as the cosine of the angle between their feature vectors using the values of the features in the

feature vectors

I

Generate extraction/contraction table of pairs of compounds where one compound is a substring of another compound with respective counts, e.g., from query logs, or similarity values

76

77

FIG. 8