Domain-Specific Entity Linking via Fake Named Entity Detection

A Two-Stage Joint Model for Domain-Specific Entity Detection and Linking Leveraging an Unlabeled Corpus

“…A re-ranking-based model suitable for longer documents is introduced in [10], in which joint EDL is preceded by a pipeline EDL model to generate high-quality candidate mentions and entity links. Zhang et al [7] propose a joint model for DSEDL, where features of entity linking and NER models are updated iteratively until convergence, and an iterative graph-based algorithm is further introduced in [31] to capture entity-entity and entity-mention dependency.…”

“…org/) and Amazon (https://www.amazon.com/) products' KB, and related domain-specific tasks (e.g., customer service QA, in-context advertising) have put forward demands for effective DSEDL techniques. On the other hand, methods for the general EDL task may not be suitable for DSEDL, because they cannot handle problems like the increase of fake named entities [7] in texts and the sparsity of links in DSKBs [8]. In a word, DSEDL is a meaningful and challenging task that deserves further research.…”

“…As a result, in some recent works [7,9,10], entity detection and entity linking are jointly modeled to reinforce each other by leveraging their interdependency. A general idea of these joint models is that, for every n-gram in the input text, if it is the anchor phrase for at least one target entity, it will be selected as a candidate mention.…”

“…A general idea of these joint models is that, for every n-gram in the input text, if it is the anchor phrase for at least one target entity, it will be selected as a candidate mention. This idea is designed to improve the recall rate of EDL, but it also generates a large number of fake mentions, resulting in much more negative candidate samples than positive ones in either the binary classification model [7] or the structure learning model [9]. Too many candidates means high computational complexity, while data imbalance also impairs the final performance.…”

“…In the pipeline architecture, textual context is used in entity detection models to determine mention boundaries and infer mention types. However, the labeled data are often insufficient to train such entity detection models in specific domains, so the bag of words representation, which neglects the syntax and word orders, is adopted in some joint models [7,9,10]. (2) The existing evaluation methods of entity coherence are not applicable for DSEDL.…”