Most current models of word representations(e.g.,GloVe) have successfullycaptured fine-grained semantics. However, semantic similarity exhibited inthese word embeddings is not suitable for resolving bridging anaphora, whichrequires the knowledge of associative similarity (i.e., relatedness) instead ofsemantic similarity information between synonyms or hypernyms. We create wordembeddings (embeddings_PP) to capture such relatedness by exploring thesyntactic structure of noun phrases. We demonstrate that using embeddings_PPalone achieves around 30% of accuracy for bridging anaphora resolution on theISNotes corpus. Furthermore, we achieve a substantial gain over thestate-of-the-art system (Hou et al., 2013) for bridging antecedent selection.