[Paper Review] Building Dynamic Knowledge Graphs from Text using Machine Reading Comprehension

Rajarshi Das, Tsendsuren Munkhdalai+3 more|arXiv (Cornell University)|Oct 12, 2018

Topic Modeling19 references44 citations

TL;DR

Kg-Mrc builds dynamic knowledge graphs from procedural text and uses an MRC-based querying mechanism to track entity states, achieving state-of-the-art results on ProPara tasks and competitive results on Recipes.

ABSTRACT

We propose a neural machine-reading model that constructs dynamic knowledge graphs from procedural text. It builds these graphs recurrently for each step of the described procedure, and uses them to track the evolving states of participant entities. We harness and extend a recently proposed machine reading comprehension (MRC) model to query for entity states, since these states are generally communicated in spans of text and MRC models perform well in extracting entity-centric spans. The explicit, structured, and evolving knowledge graph representations that our model constructs can be used in downstream question answering tasks to improve machine comprehension of text, as we demonstrate empirically. On two comprehension tasks from the recently proposed PROPARA dataset (Dalvi et al., 2018), our model achieves state-of-the-art results. We further show that our model is competitive on the RECIPES dataset (Kiddon et al., 2015), suggesting it may be generally applicable. We present some evidence that the model's knowledge graphs help it to impose commonsense constraints on its predictions.

Motivation & Objective

• Motivate automatic construction of structured, evolving knowledge graphs from procedural text.
• Develop an entity-centric model that tracks state changes over time using recurrent graph updates.
• Leverage machine reading comprehension to extract entity-state spans conditioned on graphs.
• Demonstrate end-to-end trainability with supervision from MRC spans without graph-only supervision.

Proposed method

• Extend a DrQA-based MRC component to query current entity states at each time step.
• Represent entities and their locations as nodes in a bipartite dynamic graph that updates over time.
• Apply soft co-reference across and within time steps to handle aliasing and non-changing states.
• Condition the MRC queries on the current graph embeddings to obtain location spans as ψ_{i,t}.
• Update graph nodes via a recurrent graph module with residual connections to propagate state information.
• Train end-to-end by minimizing the MRC span loss, with paragraph prefix representations and pre-trained word embeddings.

Experimental results

Research questions

• RQ1 Can dynamic, graph-structured representations improve understanding of evolving entity states in procedural text?
• RQ2 Do MRC-based queries conditioned on evolving graphs yield better state-tracking than static or non-graph baselines?
• RQ3 To what extent can soft co-reference and recurrence help disambiguate entity mentions across time?
• RQ4 Do the learned knowledge graphs impose commonsense-like constraints to improve predictions?

Key findings

• Kg-Mrc achieves state-of-the-art results on ProPara sentence-level task (Task 1) and document-level task (Task 2).
• On Task 1, Kg-Mrc attains 46.62 micro-avg versus prior 45.37 (Pro-Global) and 47.64 micro-avg in ablations show effectiveness.
• On Task 2, Kg-Mrc reaches 56.77 F1, improving over ProStruct (53.75).
• On Recipes, Kg-Mrc slightly outperforms the baseline NPN with only 10k training examples.
• Ablations show that both cross-time and within-time soft co-reference contribute to performance, and the recurrent graph module is crucial.
• Kg-Mrc learns commonsense-like constraints, reducing violations without explicit constraints.