Details
-
Sub-task
-
Status: Resolved
-
Major
-
Resolution: Incomplete
-
2.2.0
-
None
Description
This is the implementation of `ngrams` aggregate expression which is also implemented by Hive. It takes use of n-gram concept in natural language processing to understand texts.
Currently, Spark doesn't support using Hive UDAF GenericUDAFnGrams, which is actually a feature missing.
An n-gram is a contiguous subsequence of n item(s) drawn from a given sequence. This expression finds the k most frequent n-grams from one or more sequences.
This expression has the pattern of : ngrams(children: Array[Array[String]](or Array[String]), n: Int, k: Int, accuracy: Int), it can be used in conjuction with `sentences` to split the column of String to Array. Among the parameters:
Children indicates the 'given sequence' we collect n-grams from;
N indicates n-gram's element number, size 1 is referred to as a "unigram", size 2 is a "bigram", size 3 is a "trigram"...
K indicates top k;
Accuracy is related to the memory used for frequency estimation, more memory will give more accurate frequency counts.
A simple example:
`SELECT ngrams(array("abc", "abc", "bcd", "abc", "bcd"), 2, 4);` will get
`[
,
{["abc","abc"]:1.0},
{["bcd","abc"]:1.0}]`. Because there are four 2-grams for the input which are `["abc", "abc"], ["abc", "bcd"], ["bcd", "abc"], ["abc", "bcd"]`, and `["abc", "bcd"]` occurs 2 times, the other two 2-grams occurs 1 time each, while `["abc","abc"]` is alphabetically before `["bcd","abc"]`, so the answer is like that.
