Graph Embeddings for Frame Identification

Proceedings of Recent Advances in Natural Language Processing, pages 939–948, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_109 939 Graph Embeddings for Frame Identification Alexander Popov IICT, Bulgarian Academy of Sciences Sofia, Bulgaria alex.popov @bultreebank.org Jennifer Sikos IMS, University of Stuttgart Stuttgart, Germany jen.sikos @ims.uni-stuttgart.de Abstract Lexical resources such as WordNet (Miller, 1995) and FrameNet (Baker et al., 1998) are organized as graphs, where rela- tionships between words are made explicit via the structure of the resource. This work explores how structural information from these lexical resources can lead to gains in a downstream task, namely frame identification. While much of the current work in frame identification uses various neural architectures to predict frames, those neural architectures only use representations of frames based on an- notated corpus data. We demonstrate how incorporating knowledge directly from the FrameNet graph structure improves the performance of a neural network-based frame identification system. Specifically, we construct a bidirectional LSTM with a loss function that incorporates various graph- and corpus-based frame embed- dings for learning and ultimately achieves strong performance gains with the graph- based embeddings over corpus-based embeddings alone. 1 Introduction Frames are common scenarios, expressed by their typical participants and the predicates which evoke them. The INFECTING frame, where some- one transmits an illness, has participants IN- FECTED ENTITY and INFECTION CAUSE and is evoked by predicates infect and give. A single predicate can evoke one or more frames; give, for instance, can evoke INFECTING (give someone a cold) or GIVING (give someone a present). The disambiguation of which frame is evoked in con- text is the task of frame identification (FrameID), which is, in essence, a word sense disambiguation task where the senses are frames (Das et al., 2010). The FrameNet resource (Baker et al., 1998) pro- vides concrete definitions of frames, predicates, and participants (called frame elements) and is structured as a hierarchical graph. Frames to- wards the top of the hierarchy are more abstract (ex. INTENTIONALLY ACT), while frames lower in the hierarchy are more granular (ex., SUBMIT- TING DOCUMENTS). FrameNet’s graph structure captures the relationship between frames, their predicates, and their frame elements, such that a single frame can be connected to multiple frames via different relationships (Baker and Ellsworth, 2017). Given the structure of the resource, it is surprising that the graph is not used in cur- rent FrameID systems, as the relationships in the FrameNet graph have been previously shown as relevant to frame prediction, especially in cases of unseen predicates (Das et al., 2014). This work leverages FrameNet’s graph structure to boost the performance of neural architectures for FrameID. Specifically, we construct frame em- beddings from the FrameNet graph structure and use them as input to a neural network. Although frame embeddings are often used in neural ar- chitectures for FrameID, prior work only learns these embeddings from frame-annotated corpus data (Hartmann et al., 2017; Hermann et al., 2014). To our knowledge, this is the first work to incor- porate embeddings composed from the FrameNet graph structure itself, …

Content truncated to 500 words