1. Generate subtask nodes V from unstructured text
2. Infer temporal dependency edges E between nodes form unstructured text
3. Given a task t, input it into a search engine (like WikiHow) as "how to t" and store the k most relevant responses in collection Dk(t)
4. Graph induction becomes inducing the optimal graph Gs(t) = (V, E) using the data in Dk(t)

This is the same process mentioned before, but now with math notation.

University of Rochester

Decomposing and organizing a complex task becomes inducing a graph  Gs(t) = (V, E) from a complex task input t. We assume that the input is complex because distinguishing complex tasks from simple tasks is not the focus of this exercise.

Modeling Overview (Learning to Decompose and Organize Complex Tasks)

Zhang, Y., Jauhar, S. K., Kiseleva, J., White, R., &amp; Roth, D. (2021). Learning to Decompose and Organize Complex Tasks. ACL Anthology. https://www.aclweb.org/anthology/2021.naacl-main.217/. 

Learning to Decompose and Organize Complex Tasks

Key Steps of Modeling (Learning to Decompose and Organize Complex Tasks)

Generate subtasks ST(t) given t and Dk(t). The text fragments (the subtasks) must satisfy relevance, abstraction, and consensus.

The model used to generate the subtasks is built on BART, a sequence-to-sequence model for text generation, injected with relevance, abstraction, and consensus. The model they made is capable of synthesizing multi-source input, includes a relevance-aware cross attention layer, and cluster encoding techniques were added to guide the generation process.

Subtask Generation (Learning to Decompose and Organize Complex Tasks)

Finding the temporal dependencies: Given the set of subtasks V = ST(t), find E. This becomes a binary classification problem: Predict whether there is a dependency edge ∀(si , sj) ∈ V

Learn Before

Related