Xiaoshi Zhong - Time Expression and Named Entity Recognition: 10 (Socio-Affective Computing, 10)

Volume 10

Series Editors

Xiaoshi Zhong and Erik Cambria

1st ed. 2021

ISSN 2509-5706 e-ISSN 2509-5714

ISBN 978-3-030-78960-2 e-ISBN 978-3-030-78961-9

https://doi.org/10.1007/978-3-030-78961-9

The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021

This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.

The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Preface

To better illustrate my research results and how they come, I use the diagram shown in Fig. to briefly introduce my research style. In most scientific research, the general paradigm is the hypothesis + verification procedure: for a problem, we first of all have a rough idea or guess or conjuncture about this problem, and then formulate this conjuncture into a hypothesis; after that, we collect numerous data or concrete evidence to verify whether the hypothesis is true or not. When we apply this paradigm to the research on the fields of data analytics, computational linguistics, natural language processing, and other data-science areas, there involve two important components: data characteristics and model assumptions.

In current natural language processing, many works mainly propose models and then use data to evaluate the quality of these models according to some criteria such as precision, recall, and F1 score. Almost all these works however are explicitly or implicitly verifying whether the assumptions of proposed models fit the characteristics of data. Fig. , if model assumptions fit data characteristics, then proposed models will achieve good results; otherwise, we need to design a new model that fits the characteristics of analyzed data. The current research in natural language processing tends to concentrate more attention on proposing new models while ignoring to analyze data characteristics. My research style instead is to first analyze data characteristics and then design appropriate models according to these characteristics for specific tasks. Such a style can not only enhance my understanding into a task and data, but also enhance the probability that the assumptions of my designed model fit data characteristics and therefore lead my model to achieve good results. Moreover, since I have a better understanding about model assumptions and data characteristics, I can also interpret these results and explain the advantage and disadvantage of my model.

In this book, I focus on analyzing common characteristics of time expressions and named entities from diverse datasets, and then according to these characteristics design effective and efficient algorithms for recognizing time expressions and named entities from unstructured text.

When analyzing time expressions, I believe that there must be some common characteristics in diverse datasets, because we humans share some common habits to express time information, no matter what platforms we are expressing on and no matter what topics we are talking about. In the moment when I figured out that there are only about 70 distinct time tokens in an individual dataset and only about 125 distinct time tokens across four diverse datasets, I knew that my conjecture is verified. Regarding why I define token types that separate rules and specific tokens, and then design general heuristic rules that works on token types and that are independent of specific tokens, at that time I am reading books about western philosophy (e.g.,