When collecting Nepali text corpus, we usually collect it from various online sources such as Wikipedia, News portals, and other websites. The online sources introduce a lot of errors due to imperfect online tools such as translators, font convertors, spelling checker, etc. Some of these errors include typos, spelling mistakes, foreign words, incorrect symbols. Dealing with these errors poses a challenging task. In this post, we will look at a simple heuristic-based algorithm to filter Non-Devanagari words from a Nepali corpus.

The Algorithm

In our algorithm, we will check the individual characters of a token whether they belong to the Devanagari Unicode range or not. For this will use this regular expression [\u0900-\u097F\\].

Then, we will keep count of the Devanagari characters.

The algorithm uses a simple heuristic: does majority of characters in a token belong to the Devanagari range or not. If it does, then token is a Devanagari word, else not.

Pseudo Code

for token in sentence
    if majority of characters in token belong to Non-Devanagari  
        token is not in Devanagari script
    else
	    token is in Devanagari script

This algorithm classifies words with punctuations to Devanagari as oppose to an algorithm that checks if the token consists of Non-Devanagari characters.

Limitations

One of the major limitations of the procedure is that it removes the Non-Devanagari words rendering the sentence incorrect/incomplete.

Implementation

import re

devanagari_range = r'[\u0900-\u097F\\]'

def getDevanagariCharCount(token):
    return len(list(filter(lambda char: re.match(devanagari_range, char), (char for char in token))))

def isDevanagari(token):
    return True if getDevanagariCharCount(token) >= len(token)/2 else False 

def filterTokens(line):
    return list(filter(lambda t: isDevanagari(t), line.split(" ")))

sentence = "मलाई उपन्यास पढ्न, trekking जान र फूतball खेल्न मन लाग्छ।"
filtered = filterTokens(sentence)

print(filtered)

# Output
# मलाई उपन्यास पढ्न, जान र खेल्न मन लाग्छ।"