New York University 2014 Knowledge Base Population Systems

New York University (NYU) participated in three tracks of the 2014 TAC-KBP evaluation: English Slot Filling, Cold Start and Entity Discovery and Linking. While this year is the first time and second time we participated in entity discovery and linking (EDL) and cold start respectively, we have been working on the slot filling task for several years. With additional development time this year, our cold start system has borrowed more technologies from slot filling and gained significant improvement. In both slot filling and cold start systems, we find that besides incorporating new modules to expand the coverage, the systems greatly benefit from our effort to renovate some of the basic modules. Regarding entity discovery and linking, we develop a pipelined system with several components in which we introduce a variation of the PageRank algorithm to improve collaborative candidate ranking. 1 Slot Filling and Cold Start: Overview The NYU KBP slot filling and cold start systems for 2014 inherit the structure of our systems from the last several years (Sun et al., 2011; Min et al., 2012; Grishman, 2013). We improved several of our basic components: the name tagger, the relation extraction components based on hand-coded patterns, and distant supervision. These are the common components for both the slot filling and cold start systems. Note that the distant supervision module is new to the cold start system; because of time constraints, we only used patternbased relation extraction modules in the system last year. We introduce new inference modules to both cold start and slot filling systems this year. Although the inference components differ, they share the same background of manual rules. In addition, for slot filling only, we added one more distant supervision module employing the new guided multi-instance multi-label distant supervision framework we have developed recently (Pershina et al., 2014). Finally, a named entity disambiguation module is adopted to handle ambiguous names. An overview of our slot filling and cold start systems can be found in figures 1 and 2.