On the Objectivity and Reliability of Corpus Data for Linguistic Research

随着语料库的规模越来越大，以及“用户友好”环境的发展和传播，语料库在语言学研究中得到了广泛的应用，但这些“用户友好”环境在发挥其作用的同时，也使语料库和语料库研究成为一种“黑箱”；因此，用户往往只关注软件的输出，而忽略了输入和过程，而不检查输出是否可以解释为适合他们研究的数据。为了改进基于语料库的研究，我进行了以下调查：I)我首先检查了几篇文章中显示的基本数据，当发现错误时，我推断错误的过程和原因，然后进行分类；ii)从将语法视为说话者内部状态的一部分的角度来看，我研究了现在可以从公开可用的语料库中获得关于语法哪些方面的信息；iii)除其他问题外，我还研究了与语料库和语料库数据的代表性、统计分数（如t-score和MI）的可靠性和有效性以及术语“搭配”的模糊性有关的问题。

With the sizes of corpora increasingly large and with the development and spread of “user-friendly” environments, corpora are now widely used in linguistic research, but these “user-friendly” environments, in spite of their usefulness, also make corpora and corpus studies a kind of “black box”; as a result, users often pay attention only to the output of software while disregarding the input and the process and not examining whether the output can be interpreted as appropriate data for their studies. For the improvement of corpus-based research, I conducted the following surveys: i) I first examined basic data shown in several articles-when errors were detected, I inferred the processes and the causes of the errors, then classified them; ii) from the standpoint of viewing a grammar as a part of the internal state of a speaker, I examined what kinds of information about what aspects of the grammar can be obtained from corpora publicly available now; and iii) I also examined, among other issues, problems concerning representativeness of corpora and corpus data, the reliability and validity of statistical scores such as t-score and MI, and the ambiguity of the term “collocation.”