QA on Differentiable Knowledge Bases- Reified KB

This is an unofficial adaption of the work presented in Scalable Neural Methods for Reasoning With a Symbolic Knowledge Base- Paper. Utilizes a transformer based encoder, similar to the work here, as opposed to word2vec which is how it was implemented in the Symbolic KB paper

The purpose of this implementation is to provide insights into the implementation level information of these papers and the ideas asscociated with it - such as enabling a differentiable KG through the "follow" operation.

Implemenation

Implements a QA model to retrieve questions from a KB. This consists of an encoder, MiniLM-6/MiniLM-12, which encodes a question. And then, it performs the differentiable query operation on the knowledge base as mentioned in the reified KB paper above. This encode archicture is different from the paper in that it's not based on a word2vec model, but a tranformer. We fine-tune a MiniLM-6 Sentence Transformer - primarily because of the memory and compute efficiency. Can replace this with any other model from the huggingface library

How to run:

• Refer to kgqa/models

Benchmarks

Bencharked on the MetaQA dataset, similar to the other symbolic kb paper linked above

One expectation was that 3 hop performance to be worse than 2-hop, which was the case in reified kb paper as well. The reified KB paper does report higher 2-hop performance, but this 1-hop and 3-hop outperforms it.

User Experimentation

The purpose of this repository is to provide a baseline to users to implement improvements in this space of question answering with reified differential KBs

• Update encoder:
  • Utilize different transformer model - Pass in any model from Hugginface library
  • To change encoder to word2vec, update the torch dataset (train_model_metaqa) and forward pass in net (models_nhop)
• Implement updated architectures- for eg. to not require subject_id as part of model input - perform entity recongition and train end to end

To add

• To add training and inference on custom datasets
• Method to integrate 1hop, 2hop, 3hop into the same model architecture

Installation

Requires pytorch-lightning >= 1.5.0, pytorch >= 1.7, tqdm