KECA

Implementation of "Knowledge Enhanced Latent Relevance Mining For Question Answering" (ICASSP 2020). Some code in this repository is based on the excellent open-source project https://github.com/wjbianjason/Dynamic-Clip-Attention.

Requirements

• python 3.x
• Keras 2.2.4
• Tensorflow 1.12.0

Data

• WikiQA
• TrecQA

Running

usage: main.py [-h] [-t Task] [-m MODEL] [-d HIDDEN_DIM] [-e PATIENCE] [-l LR] [-b BATCH_SIZE] [--no_entity_cmp] [--no_cxt_weight_mixed] [--no_ent_weight_mixed]

WikiQA

• KECA

python main.py -t wikiqa

• w/o knowledge-aware attention

python main.py -t wikiqa --no_ent_weight_mixed --no_cxt_weight_mixed

• w/o knowledge-based features

python main.py -t wikiqa --no_entity_cmp

• w/o external knowledge (KGs)

python main.py -t wikiqa --no_entity_cmp --no_cxt_weight_mixed

TrecQA

It's similar to WikiQA, change task from 'wikiqa' to 'trecqa'. For example:

python main.py -t trecqa