DualTKB: A Dual Learning Bridge between Text and Knowledge Base

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 8605–8616, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 8605 DualTKB: A Dual Learning Bridge between Text and Knowledge Base Pierre L. Dognin∗ IBM Research pdognin@us.ibm.com Igor Melnyk* IBM Research igor.melnyk@ibm.com Inkit Padhi* IBM Research inkpad@ibm.com Cicero Nogueira dos Santos AWS AI cicnog@amazon.com Payel Das IBM Research daspa@us.ibm.com Abstract In this work, we present a dual learning ap- proach for unsupervised text to path and path to text transfers in Commonsense Knowledge Bases (KBs). We investigate the impact of weak supervision by creating a weakly su- pervised dataset and show that even a slight amount of supervision can significantly im- prove the model performance and enable better-quality transfers. We examine different model architectures, and evaluation metrics, proposing a novel Commonsense KB comple- tion metric tailored for generative models. Ex- tensive experimental results show that the pro- posed method compares very favorably to the existing baselines. This approach is a viable step towards a more advanced system for au- tomatic KB construction/expansion and the re- verse operation of KB conversion to coherent textual descriptions. 1 Introduction The automatic construction of Knowledge Bases (KBs) from text and the reverse operation of sen- tence generation from KBs are dual tasks that are both active research topics. The first task of automatic KB construction re- mains a significant challenge due to the difficulty of detecting parts of text representing meaningful facts and summarizing them in a systematic form. A simpler sub-task of KB completion, i.e., extend- ing or filling-in missing nodes or edges, has also attracted the attention of the research community. For both tasks, the system needs to generate new or complete existing graph entities coherently, possi- bly matching to the already existing graph structure. The dual task of decoding the information from KB back to text is a valuable functionality. This enables knowledge transfer from potentially large complex ∗* Equal contribution voice [1] capable of captain [2] capable of taste food cross street thank another water plant wash cloth eat think close eyes your eyes rest snores [3] desires [4] used for [6] capable of [7] not desires [8] desires [10] desires [15] causes [14] motivated by goal [12] has last subevent [9] has prerequisite food person sleep [11] capable of [13] causes [5] capable of [1] a person can voice an opinion [2] a person can captain a ship [3] a person can taste food [4] person can cross the street [5] person can thank another person [6] a person can water a plant [7] person can wash cloths [8] person can eat [9] you eat food [10] a person can think [11] a person can sleep [12] if you want to sleep then you should close eyes [13] sleep would make you want to snore [14] if you want to sleep then you should close your eyes [15] you sleep to rest. Figure 1: Text to Path. Part of a larger graph generated from test sentences from …

Content truncated to 500 words