Survey and evaluation of web search engine hit counts as research tools in computational linguistics

In recent years, many studies on computational linguistics have employed the Web as source for research. Specifically, the distribution of textual data in the Web is used to drive linguistic analyses in tasks such as information extraction, knowledge acquisition or natural language processing. For these purposes, commercial Web search engines are commonly used as the low-entry-cost way to access Web data and, more specifically, to estimate the distribution of the entity(ies) of interest from the hit count the search engines provide when querying such entities. Even though several studies have evaluated the effectiveness of Web search engines as information retrieval tools from the perspective of the end users, few authors have assessed the suitability of hit counts as research tools in computational linguistics; moreover, studies so far have focused on the most well-known search engines (typically Google, Bing and Yahoo!), and neglected potentially interesting alternatives that have recently surfaced. To fill this gap, in this work, we first compile and survey the general-purpose search engines that are currently available. Then, we evaluate the suitability of the hit counts they provide under several perspectives that are relevant for computational linguistics: flexibility of the query language, linguistic coherence, mathematical coherence and temporal consistency. The results of our survey show that, even though the choice of a particular search engine has been generally ignored by researchers relying on Web data, there are significant quality differences between the hit counts of current search engines, and that the most well-known and widely-used search engines do not provide the best results. In this respect, we also identify the search engines whose hit counts are best suited for research.