transformer_rankers.negative_samplers.negative_sampling.BM25NegativeSamplerPyserini

class transformer_rankers.negative_samplers.negative_sampling.BM25NegativeSamplerPyserini(candidates, num_candidates_samples, path_index, sample_data, anserini_folder, set_rm3=False, seed=42)[source]

Bases: object

Sample candidates from a list of candidates using BM25.

The class uses anserini and pyserini which requires JAVA and a installation of anserini. It first generates the candidates, saving then to files, then creates the index via anserini IndexCollection.

Parameters

  • candidates – list of str containing the candidates

  • num_candidates_samples – int containing the number of negative samples for each query.

  • path_index – str containing the path to create/load the anserini index.

  • sample_data – int indicating amount of candidates in the index (-1 if all)

  • anserini_folder – str containing the bin <anserini_folder>/target/appassembler/bin/IndexCollection

  • set_rm3 – boolean indicating whether to use rm3 or not.

  • seed – int with the random seed

__init__(candidates, num_candidates_samples, path_index, sample_data, anserini_folder, set_rm3=False, seed=42)[source]

Initialize self. See help(type(self)) for accurate signature.

Methods

__init__(candidates, num_candidates_samples, …)

Initialize self.

sample(query_str, relevant_docs[, max_query_len])

Samples from a list of candidates using BM25.
sample(query_str, relevant_docs, max_query_len=512)[source]

Samples from a list of candidates using BM25.

If the samples match the relevant doc, then removes it and re-samples randomly.

Parameters

  • query_str – the str of the query to be used for BM25

  • relevant_docs – list with the str of the relevant documents, to avoid sampling them as negative sample.

  • max_query_len – int containing the maximum number of characters to use as input. (Very long queries will raise a maxClauseCount from anserini.)

Returns

First the sampled_documents, their respective scores and then indicators if the NS retrieved the relevant document, and if so at which position.