conclusion.tex (1389B)
1 \section{Conclusion} 2 \label{sec:graph:conclusion} 3 In this chapter, we explore aggregate approaches to unsupervised relation extraction using graphs. 4 In Section~\ref{sec:graph:analysis}, we show that a large amount of information can be leveraged from the neighborhood of a sample. 5 This, together with the observation that previous unsupervised methods always ignored the neighborhood of a sample at inference, opens a new research direction for unsupervised methods. 6 In Section~\ref{sec:graph:approach}, we propose several models that make use of the neighborhood information. 7 In particular, we propose a novel unsupervised training loss in Section~\ref{sec:graph:refining}, which makes very few modeling assumptions while still being able to exploit the neighborhood information both at training and prediction time. 8 9 Our contributions lie in using a multigraph with arcs attributed with sentences (Sections~\ref{sec:graph:encoding}), our method to approximate the quantity of information extractible from this graph (Sections~\ref{sec:graph:analysis}) and our proposed approach to utilize this additional information (Section~\ref{sec:graph:approach}). 10 Despite encouraging early results showing the soundness of using the relation extraction graph, at the present time we only improved nonparametric models. 11 More experimentation is still needed to fully exploit topological information.