BM25opt

faster BM25 search algorithms in Python

based on https://github.com/dorianbrown/rank_bm25 by Dorian Brown

Apache License, Version 2.0 http://www.apache.org/licenses/LICENSE-2.0

---

News:

• 1.1.0 supports updating the index with new add_documents(), delete_documents() and update_documents() functions, see Example 4

---

Usage:

Input:

• corpus is a list of strings, e.g. [ 'bla bla bla', 'this is document two', ... ]
• question is a string, e.g. 'which text contains the word two?'
• optional arguments:
  • algo : BM25 algorithm, the default is 'okapi'; 'l' and 'plus' available
  • tokenizer_function : the default is tokenizer_default which is split-on-whitespace, lowercase, remove common punctiations
  • idf_algo : default uses the same IDF as rank_bm25; values 'okapi', 'l' and 'plus' can override to fix dorianbrown/rank_bm25#35
  • k1, b, epsilon, delta : constants with standard default values, see https://en.wikipedia.org/wiki/Okapi_BM25

Example 1:

This example uses the default tokenizer and the default BM25Okapi algorithm and returns the top 5 highest scoring document ids and scores.

# creating the index
bm25opt_index = BM25opt( corpus )

# search
results = bm25opt_index.topk( question, 5 )
print( 'results[0] id', results[0][0], 'results[0] score', results[0][1], 'results[0] document', corpus[ results[0][0] ] )

Example 2:

This example returns the list of document scores (order is the same as the document order in corpus), shows algoritm selection and custom tokenizer function.

bm25opt_index = BM25opt( corpus, algo='plus', tokenizer_function=some_tokenizer_function )
doc_scores = bm25opt_index.get_scores( question )

Example 3: comparison with rank_bm25

This example shows the score list and the similarity with rank_bm25, but NOTE: BM25opt input is not tokenized beforehand.

corpus = [ ... ]
question = '...'
tokenized_corpus = [ tokenizer_default(document) for document in corpus ]
tokenized_question = tokenizer_default( question )

rank_bm25_index = BM25Okapi( tokenized_corpus )
bm25opt_index = BM25opt( corpus, algo='okapi' )

rank_bm25_scores = rank_bm25_index.get_scores( tokenizedquestion )
bm25opt_scores = bm25opt_index.get_scores( question )

Example 4: updating the index

# creating the index
bm25opt_index = BM25opt( corpus )

# add new documents
bm25opt_index.add_documents( corpus2 ) 

# delete from the index
delete_ids = [ 1, 3, 5 ] # list of document ids (indices in corpus) to delete from the index
bm25opt_index.delete_documents( delete_ids )

# in-place update changed documents in the index
update_ids = [ 1, 3, 5 ] # list of document ids (indices in corpus) to change
updated_documents = [ 'first changed document', 'second changed document', ... ]
bm25opt_index.update_documents( update_ids, updated_documents )

---

Notes:

This is an optimized variant of rank_bm25 where the key insight is that we can calculate almost everything at index creation time in __init__() , resulting a words * documents-score dict, e.g.

wsmap = {
  'word1': [ word1_doc1_score, word1_doc2_score, ... ],
  'word2': [ word2_doc1_score, word2_doc2_score, ... ],
  ...
}

then the query function is just adding the score lists for each word in the question, e.g.

question = 'word1 word2'
doc_scores = [ wsmap['word1'][0] + wsmap['word2'][0], wsmap['word1'][1] + wsmap['word2'][1], ... ]

Another important change is the un-tokenized inputs and registration of the tokenizer function, which is important to avoid situations where the corpus would be tokenized with a different function than the queries later. A simple tokenizer_default() function is provided as a default.