Skip to content

Latest commit

 

History

History
95 lines (89 loc) · 5.73 KB

README.md

File metadata and controls

95 lines (89 loc) · 5.73 KB

Exploring Gender Biases in Information Retrieval Relevance Judgement Datasets

This repository contains the code and resources for detecting the gender of queries (Female, Male, Neutral) along with psychological characteristics of their relevance judgement documents.

Query Gender Identification and Labeling

In this work, we proposed a Query Gender classifier. As the first step and in order to be able to label queries based on their gender at scale, we employed the gender-annotated dataset released by Navid Rekabsaz and Markus Schedl to train relevant classifiers. This dataset consists of 742 female, 1,202 male and 1,765 neutral queries. We trained various types of classifiers on this dataset and in order to evaluate the performance of the classifiers, we adopt a 5-fold cross-validation strategy.

Category Classifier Accuracy F1-Score

Dynamic Embeddings
 BERT (base uncased) 0.856 0.816 0.872 0.862
DistilBERT (base uncased) 0.847 0.815 0.861 0.853
RoBERTa 0.810 0.733 0.820 0.836
DistilBERT (base cased) 0.800 0.730 0.823 0.833
BERT (base cased) 0.797 0.710 0.805 0.827
XLNet (base cased) 0.795 0.710 0.805 0.826
Static Embeddings Word2Vec 0.757 0.626 0.756 0.809
fastText 0.750 0.615 0.759 0.792

In the table above the performance of each of the developed classifiers is reported. As shown the uncased fine-tuned BERT model shows the best performance for query gender identification. Finally, for the purpose of measuring bias in relevance judgements, we used our best-performed model to identify the gender of queries in MS MARCO Dev set that had at least one related human-judged relevance judgement document - equivalent to 51,827 queries. Note that, the queries of gender-annotated dataset were removed from this dataset to avoid unintended leakage.

The following table illustrates a few queries labeled using our fine-tuned BERT classifier. Here is all the the 1405 female, 1405 male, and 1405 neutral labeled queries.

QID Query Predicted Gender
80095 can you take naproxen during pregnancy Female
14757 aimee osbourne net worth Female
189154 foods that can prevent prostate cancer Male
11251 adam devine net worth Male
40234 average percentage of accepted scholarships Neutral

Code

  • Training - code/train.py: The code for fine-tuning BERT on queries_gender_annotated dataset or any other dataset.
  • Predicting - codes/predict.py: In any case that you do not want to train the model, you can download our fine-tuned model and use predict.py for predicting the gender of queries.

Psychological Characteristics Quantification

Our approach for quantifying bias is based on measuring different psychological characteristics of the relevance judgement documents associated with each query. To investigate this, we employ Linguistic Inquiry and Word Count (LIWC) text analytics toolkit to compute the degree to which different psychological characteristics are observed in relevance judgement documents. These Psychological characteristics related to the queries of each group can be found in results/psychological analysis folder.