Semi-supervised Deep Embedded Clustering with Anomaly Detection for Semantic Frame Induction

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 3509–3519 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 3509 Semi-supervised Deep Embedded Clustering with Anomaly Detection for Semantic Frame Induction Zheng-Xin Yong1, Tiago Timponi Torrent2 1Minerva Schools at Keck Graduate Institute, San Francisco, CA - United States 2Federal University of Juiz de Fora - FrameNet Brasil, Juiz de Fora, MG - Brazil zhengxin.yong@minerva.kgi.edu, tiago.torrent@ufjf.edu.br Abstract Although FrameNet is recognized as one of the most fine-grained lexical databases, its coverage of lexical units is still limited. To tackle this issue, we propose a two-step frame induction process: for a set of lexical units not yet present in Berkeley FrameNet data release 1.7, first remove those that cannot fit into any existing semantic frame in FrameNet; then, assign the remaining lexical units to their correct frames. We also present the Semi-supervised Deep Embedded Clustering with Anomaly Detection (SDEC-AD) model—an algorithm that maps high-dimensional contextualized vector representations of lexical units to a low-dimensional latent space for better frame prediction and uses reconstruction error to identify lexical units that cannot evoke frames in FrameNet. SDEC-AD outperforms the state-of-the-art methods in both steps of the frame induction process. Empirical results also show that definitions provide contextual information for representing and characterizing the frame membership of lexical units. Keywords: Semantic Frame Induction, Deep Embedded Clustering, Semi-supervised Learning, Anomaly Detection 1. Introduction A semantic frame is a conceptual structure that models a type of situation, entity, or event (Ruppenhofer et al., 2006). Frame semantics is useful for inference-based nat- ural language processing tasks such as question answering (Shen and Lapata, 2007), text summarization (Han et al., 2016), and information extraction (Moschitti et al., 2003; Barzdins, 2014). A widely-used frame semantic resource is Berkeley FrameNet (Baker et al., 1998), whose current release 1.7 (BFN 1.7) contains 1224 hierarchically-related semantic frames and 13,669 lexical units (LUs). An LU is a pair of word lemma and the semantic frame it evokes. For example, the LU abandon.v falls under the Abandonment frame, which describes an agent leaving behind an object and rendering such object no longer within one’s control. The lexical coverage of BFN 1.7 is low compared to other semantic lexical resources such as WordNet (Miller, 1995), which contains 210,000 entries, as FrameNet is built en- tirely manually by linguistic experts. Expanding FrameNet automatically is challenging because of the high num- ber and uneven granularity of semantic frames (Rastogi and Van Durme, 2014), and polysemous lemmas such as shoot.v, which can be assigned to multiple frames such as Hit target and Ingest substance. The NOTR-LU (lexical unit with no training data) coverage gap (Palmer and Sporleder, 2010), where 24% of LUs in BFN 1.7 lack example sentences, further complicates the challenge as the state-of-the-art frame induction methods (Anwar et al., 2019; Arefyev et al., 2019; Ribeiro et al., 2019) require example sentences featuring the LUs to create vector repre- sentations in the semantic space and induce frames. At the same time, …

Content truncated to 500 words