Professional Documents
Culture Documents
Poster On Explanation Regeneration Via Multi-Hop ILP Inference Over Knowledge Base
Poster On Explanation Regeneration Via Multi-Hop ILP Inference Over Knowledge Base
Srinivasaraghavan
International Institute of Information Technology Bangalore
• Generate a baseline ranking of individual facts in KB by fine-tuning BERT model[2] and get top-K ranked
facts with respect to each question and its correct answer
Task Description • Use the fact chain of top-K ranked facts and create an Integer Linear Program (ILP) with variables and Conclusion
constraints consisting of a graph of nodes and edges created from constituents of question, answer and
A multi-hop inferencing task of regenerating the fact chain. The ILP maximizes the graph[3] that has maximum alignment between the edges for re- Considering the multi-hop nature of the task,
explanations wherein the gold explanation is scoring each fact in the chain in order to gather the relevant subgraph of correct chain of facts that can instead of ranking each fact from the KB
formed by chaining together individual facts explain the answer: individually, we devised an ILP that can infer
from a Knowledge Base (KB) ordered such that links between question, answer and KB facts to
they form correct reasoning behind the score and rank a chain of explanation facts with
answer to a question. • Different Wordnet- based alignment scores used for computing phrasal entailment (edge) scores between the regression model.
constituent nodes of question, answer and fact chain include: Question Fact Chain Alignments (QFA), Fact
Chain Answer Alignments (FAA), Intra-Fact Chain Alignments (IFA), Inter-Fact Chain Alignments (IFA2). The addition of Inter-Fact Chain Alignment
• Generate combined scores from a regression model that takes alignment scores from the ILP model for feature did not improve the MAP score
each QA pair as input features and produce a final sorted list of scores to re-rank facts in a chain leading probably because of the semantic drift among
to generation of the final explanation for each question: facts in an explanation indicating requirement
of better ILP parameter tuning and phrasal
entailment scoring methods.