I'm trying to find the average length of words by paragraph. Data is pulled from a text file in the format of 1|For more than five years... where each line features a paragraph number.
So far this is my code:
from pyspark import SparkContext, SparkConf
sc = SparkContext('local', 'longest')
text = sc.textFile("walden.txt")
lines = text.map(lambda line: (line.split("|")[0],line))
lines = lines.filter(lambda kv: len(kv[1]) > 0)
words = lines.mapValues(lambda x: x.replace("1|","").replace("2|","").replace("3|",""))
words = words.mapValues(lambda x: x.split())
words = words.mapValues(lambda x: [(len(i),1) for i in x])
words = words.reduceByKey(lambda a,b: a+b)
words.saveAsTextFile("results")
And the current output follows this format:
('1', [(2,1),(6,1),(1,1)..etc)]),('2', [(2,1),(6,1),(1,1)..etc)]),('3', [(2,1),(6,1),(1,1)..etc)])
Where '1'/'2'/'3' are the paragraph IDs, and the tuples follow (word length, 1) format.
What I need to do is sum the values of the tuples (by key/paragraph ID) so that (2,1),(6,1),(1,1) becomes (9,3) and then divide these values (9/3) to find the average length of words in each paragraph.
I've tried a bunch of different things but just can't get this to work. Your help is greatly appreciated!
For your rdd case, try this.
text = sc.textFile("test.txt")
lines = text.map(lambda line: (line.split("|")[0],line))
lines = lines.filter(lambda kv: len(kv[1]) > 0)
words = lines.mapValues(lambda x: x.replace("1|","").replace("2|","").replace("3|",""))
words = words.mapValues(lambda x: x.split())
words = words.mapValues(lambda x: [len(i) for i in x])
words = words.mapValues(lambda x: sum(x) / len(x))
words.collect()
[('1', 4.0), ('2', 5.4), ('3', 7.0)]
I use the dataframe and got this.
import pyspark.sql.functions as f
df = spark.read.option("inferSchema","true").option("sep","|").csv("test.txt").toDF("col1", "col2")
df.show(10, False)
+----+---------------------------------------+
|col1|col2 |
+----+---------------------------------------+
|1 |For more than five years |
|2 |For moasdre than five asdfyears |
|3 |Fasdfor more thasdfan fidafve yearasdfs|
+----+---------------------------------------+
df.withColumn('array', f.split('col2', r'[ ][ ]*')) \
.withColumn('count_arr', f.expr('transform(array, x -> LENGTH(x))')) \
.withColumn('sum', f.expr('aggregate(array, 0, (sum, x) -> sum + LENGTH(x))')) \
.withColumn('size', f.size('array')) \
.withColumn('avg', f.col('sum') / f.col('size')) \
.show(10, False)
+----+---------------------------------------+---------------------------------------------+---------------+---+----+---+
|col1|col2 |array |count_arr |sum|size|avg|
+----+---------------------------------------+---------------------------------------------+---------------+---+----+---+
|1 |For more than five years |[For, more, than, five, years] |[3, 4, 4, 4, 5]|20 |5 |4.0|
|2 |For moasdre than five asdfyears |[For, moasdre, than, five, asdfyears] |[3, 7, 4, 4, 9]|27 |5 |5.4|
|3 |Fasdfor more thasdfan fidafve yearasdfs|[Fasdfor, more, thasdfan, fidafve, yearasdfs]|[7, 4, 8, 7, 9]|35 |5 |7.0|
+----+---------------------------------------+---------------------------------------------+---------------+---+----+---+
I know this is really different approach but would be helpful.