Web QA: retrieve evidence, rank answers, extend to more languages

This is a reading map rather than one end-to-end system: WDRASS supplies data and evaluation, Question–Context Alignment improves candidate ranking, and Cross-Lingual Knowledge Distillation extends answer selection to lower-resource languages.

Train and evaluate at web scale

WDRASS: A Web-scale Dataset for Document Retrieval and Answer Sentence Selection

WDRASS connects document retrieval, passage reranking, and answer sentence selection with roughly 64,000 questions and more than 800,000 labeled passages and sentences drawn from 30 million documents.

In its Open retrieval setting, the paper reports H@50 of 0.903 for WDRASS-trained DPR versus 0.777 for the NQ-trained comparison.

Web index retrieves candidates, annotators label answer sentences, then those labels propagate back to passages and docs. — Figure 1: ODQA pipeline for generating question-candidate pairs, which are labelled by expert annotators (upper/orange pipeline). The annotation is propagated back to the embedding passages and documents (lower/blue pipeline). Taken from the paper.

Read the paper page PDF Citation + BibTeX

Use question–context dependencies when ranking answers

Question-Context Alignment and Answer-Context Dependencies for Effective Answer Sentence Selection

The method aligns question and surrounding-context words with optimal transport, represents dependencies as a sentence graph, and derives candidate representations with a graph convolutional network.

On non-fine-tuned WikiQA, the paper reports P@1 of 74.16 versus 68.09 for the comparison.

RoBERTa encodes question plus candidate and neighbor sentences; optimal transport aligns words; a GCN scores the answer. — Figure 1: Illustration for our proposed model. Taken from the paper.

Read the paper page PDF Citation + BibTeX

Transfer answer selection across languages

Cross-Lingual Knowledge Distillation for Answer Sentence Selection in Low-Resource Languages

A strong English teacher supplies soft scores so target-language models can be trained without target-language labels.

On Xtr-WikiQA Arabic in the named configuration, the paper reports P@1 of 76.3 versus 64.2.

CLKD pipeline: an English teacher scores machine-translated QA pairs; the low-resource student learns to copy those scores. — Figure 1: Cross-Lingual Knowledge Distillation (CLKD) in two different scenarios: (Top) using unlabeled English AS2 dataset for target low-resource language lacking any data and (Bottom) using unlabeled original low-resource language AS2 dataset. CLKD enables student AS2 models to learn from English teacher AS2 models without human-annotated datasets. Taken from the paper.

Read the paper page PDF Citation + BibTeX

Opportunities for future work

Future work can extend this line with stronger multilingual evaluation, language-specific analysis, and practical deployment studies.