python - Efficient string matching in Apache Spark

Question

Using an OCR tool I extracted texts from screenshots (about 1-5 sentences each). However, when manually verifying the extracted text, I noticed several errors that occur from time to time.

Given the text "Hello there ??! I really like Spark ??!", I noticed that:

1) Letters like "I", "!", and "l" get replaced by "|".

2) Emojis are not correctly extracted and replaced by other characters or are left out.

3) Blank spaces are removed from time to time.

As a result, I might end up with a string like this: "Hello there 7l | real|y like Spark!"

Since I am trying to match these string against a dataset including the correct text (in thise case "Hello there ??! I really like Spark ??!"), I am looking for an efficient way how to match the string in Spark.

Can anyone suggest an efficient algorithm for Spark which allows me to compare the extract texts (~100.000) against my dataset (~100 million)?

Answer 1

I wouldn't use Spark in the first place, but if you are really committed to the particular stack, you can combine a bunch of ml transformers to get best matches. You'll need Tokenizer (or split):

import org.apache.spark.ml.feature.RegexTokenizer

val tokenizer = new RegexTokenizer().setPattern("").setInputCol("text").setMinTokenLength(1).setOutputCol("tokens")

NGram (for example 3-gram)

import org.apache.spark.ml.feature.NGram

val ngram = new NGram().setN(3).setInputCol("tokens").setOutputCol("ngrams")

Vectorizer (for example CountVectorizer or HashingTF):

import org.apache.spark.ml.feature.HashingTF

val vectorizer = new HashingTF().setInputCol("ngrams").setOutputCol("vectors")

and LSH:

import org.apache.spark.ml.feature.{MinHashLSH, MinHashLSHModel}

// Increase numHashTables in practice.
val lsh = new MinHashLSH().setInputCol("vectors").setOutputCol("lsh")

Combine with Pipeline

import org.apache.spark.ml.Pipeline

val pipeline = new Pipeline().setStages(Array(tokenizer, ngram, vectorizer, lsh))

Fit on example data:

val query = Seq("Hello there 7l | real|y like Spark!").toDF("text")
val db = Seq(
  "Hello there ??! I really like Spark ??!", 
  "Can anyone suggest an efficient algorithm"
).toDF("text")

val model = pipeline.fit(db)

Transform both:

val dbHashed = model.transform(db)
val queryHashed = model.transform(query)

and join

model.stages.last.asInstanceOf[MinHashLSHModel]
  .approxSimilarityJoin(dbHashed, queryHashed, 0.75).show

+--------------------+--------------------+------------------+                  
|            datasetA|            datasetB|           distCol|
+--------------------+--------------------+------------------+
|[Hello there ??! ...|[Hello there 7l |...|0.5106382978723405|
+--------------------+--------------------+------------------+

The same approach can be used in Pyspark

from pyspark.ml import Pipeline
from pyspark.ml.feature import RegexTokenizer, NGram, HashingTF, MinHashLSH

query = spark.createDataFrame(
    ["Hello there 7l | real|y like Spark!"], "string"
).toDF("text")

db = spark.createDataFrame([
    "Hello there ??! I really like Spark ??!", 
    "Can anyone suggest an efficient algorithm"
], "string").toDF("text")


model = Pipeline(stages=[
    RegexTokenizer(
        pattern="", inputCol="text", outputCol="tokens", minTokenLength=1
    ),
    NGram(n=3, inputCol="tokens", outputCol="ngrams"),
    HashingTF(inputCol="ngrams", outputCol="vectors"),
    MinHashLSH(inputCol="vectors", outputCol="lsh")
]).fit(db)

db_hashed = model.transform(db)
query_hashed = model.transform(query)

model.stages[-1].approxSimilarityJoin(db_hashed, query_hashed, 0.75).show()
# +--------------------+--------------------+------------------+
# |            datasetA|            datasetB|           distCol|
# +--------------------+--------------------+------------------+
# |[Hello there ??! ...|[Hello there 7l |...|0.5106382978723405|
# +--------------------+--------------------+------------------+

Related

• Optimize Spark job that has to calculate each to each entry similarity and output top N similar items for each