conclusion.tex (2738B)
1 \section{Conclusion} 2 \label{sec:relation extraction:conclusion} 3 In this chapter, we introduced the relation extraction tasks (Section~\ref{sec:relation extraction:definition}) and the different supervision schema with which we can tackle them (Section~\ref{sec:relation extraction:supervision}). 4 As we showed, the development of supervised relation extraction models closely followed the evolution of \textsc{nlp} models introduced in Section~\ref{sec:context:sentence}. 5 This is particularly visible in Section~\ref{sec:relation extraction:sentential}, which follows the progress of sentential relation extraction approaches. 6 Furthermore, the expansion of the scale at which problems are tackled is visible both on the \textsc{nlp} side with the word-level to sentence-level evolution and on the information extraction side with the sentential to aggregate extraction evolution. 7 The aggregate models, which are more aligned with the information extraction field, are presented in Section~\ref{sec:relation extraction:aggregate}. 8 Within these models, we also see the evolution from the simple max-pooling of \textsc{miml} (Section~\ref{sec:relation extraction:miml}) toward more sophisticated approaches which model the topology of the dataset more finely (Section~\ref{sec:relation extraction:epgnn}). 9 10 We limited our presentation of supervised models to those critical to the development of unsupervised models. 11 Several recent approaches propose to reframe supervised relation extraction---and other tasks---as language modeling \parencitex{t5}[-9mm] or question answering \parencitex{span_prediction}[-2mm] tasks. 12 Since these approaches were not explored in the unsupervised setup yet, we omit them from our related work. 13 14 Finally, Section~\ref{sec:relation extraction:unsupervised} focused on the specific setup of interest to this thesis: unsupervised relation extraction. 15 This setup is particularly complex due to the discrepancy between the expressiveness of our supervised models and the weakness of the semantic signal we are seeking to extract. 16 As we saw, modeling hypotheses are central to tackling this problem. 17 Early models, including supervised ones, relied on strong hypotheses to facilitate training. 18 However, while supervised models can now use deep neural networks without any hypothesis other than the unbiasedness of their data, unsupervised models still need to rely on strong assumptions. 19 20 In the next section, we focus on unsupervised discriminative models, in particular the \textsc{vae} model presented in Section~\ref{sec:relation extraction:vae}. 21 In particular, we propose better losses for enforcing \hypothesis{uniform}, which avoid problematic degenerate solutions of the clustering relation extraction task.