This repository reimplements some key components of staged query graph framework for KBQA, which is proposed in Semantic Parsing via Staged Query Graph Generation: Question Answering with Knowledge Base. The original paper uses FreeBase and doesn't open source code. For simplicity, this repository only implements core inferential chain and ranking models based on the data provided here, and doesn't involve any knowledge graph operation.
The goal of idenfitying core inferential chain is to map the natural utterance of the question to the correct predicate sequence. As the original paper does, I use a SiameseCNN for this purpose. Train the inferential chain model to calculate the PatChain score:
python3 train_infer_chain.py patchain ./data/webquestions.examples.train.e2e.top10.filter.patrel.sid.tsv /path/to/save/modelTrain the inferential chain model to calculate the QuesEP score:
python3 train_infer_chain.py patchain ./data/webquestions.examples.train.e2e.top10.filter.qep.sid.tsv /path/to/save/modelNow you've got PatChain and QuesEP models. If you want to use pretrained word embedding such as GLoVe and word2vec, please set --word_vec with your word embdding file.
Before training the ranker, you need to extract features of PatChain and QuesEP. For example, if you want to extract PatChain features from ./data/webquestions.examples.train.e2e.top10.filter.patrel.sid.tsv:
python3 extract_features.py patchain /path/of/patchain/model /path/of/patchain/vocab ./data/webquestions.examples.train.e2e.top10.filter.patrel.sid.tsvYou'll get a patchain_features.txt file containing PatChain features in current directory. Entity linking features can be loaded from ./data/webquestions.examples.train.e2e.top10.filter.sid.tsv directly. Other features such as ClueWeb, ConstraintEntityWord, ConstraintEntityInQ depending on the structure of knowledge graph are not involved in this repository.
Given a query graph, once features of entity linking, PatChain and QuesEP are extracted, we can use them to train a LambdaRank model to assign a score for the query graph. Train the rank model:
python3 train_ranking.py ./data/webquestions.examples.train.e2e.top10.filter.sid.tsv data/webquestions.examples.train.e2e.top10.filter.q_ep.sid.tsv data/matching_scores.txt /path/to/save/modelBecause of absence of knowledge graph, I can't provide a complete inference for KBQA. However, once the ranker model is trained, you've owned the core component of this KBQA method. Given a query, you can load the ranker model to get score for each retrieved document:
import torch
model = torch.load('rank_best_model.pth')
model.eval()
x = torch.rand(2, 3) # assume the query retrieves two documents
y_score = model(x)
print(y_score)