Hybridly Explainable Verified Language Processing

The volume and diversity of digital information has led to a growing reliance on Machine Learning techniques, such as Natural Language Processing, for interpreting and accessing appropriate data. Among these techniques, vector and graph embeddings are used for representing data points, ranging from individual words to entire documents across multiple corpora. To retrieve this data, we need accurate similarity pipelines, to ensure we get relevant information from a given queried full-text. Current state-of-the-art does not guarantee this, as explainability is not certain. We demonstrate that our pipeline can achieve hybrid explainability, through combining graphs and logic to produce First-Order Logic representations, that are machine and human-readable via Montague Grammar. Preliminary results remark the effectiveness of the proposed approach in accurately capture full-text similarity by comparing our results with the cosine similarity derivable from sentence embedding generated by HuggingFace transformers.