Measuring scientific buzz

TL;DR

This paper analyzes the lifecycle and burstiness of keywords in AI literature, revealing how quickly new terms fade and identifying thematic trends over time using keyword data from a large corpus.

Contribution

It provides the first large-scale analysis of keyword lifecycle and thematic bursts in AI, highlighting differences between journals and conferences.

Findings

80% of keywords fade within a year

Keywords last longer in journals than conferences

Identified two thematic bursts in AI research

Abstract

Keywords are important for information retrieval. They are used to classify and sort papers. However, these terms can also be used to study trends within and across fields. We want to explore the lifecycle of new keywords. How often do new terms come into existence and how long till they fade out? In this paper, we present our preliminary analysis where we measure the burstiness of keywords within the field of AI. We examine 150k keywords in approximately 100k journal and conference papers. We find that nearly 80\% of the keywords die off before year one for both journals and conferences but that terms last longer in journals versus conferences. We also observe time periods of thematic bursts in AI -- one where the terms are more neuroscience inspired and one more oriented to computational optimization. This work shows promise of using author keywords to better understand dynamics of buzz within science.

Tables

Table 1. Table 1: Descriptive Statistics

	Num	Papers w/	Num	Keywords/
Paper Type	Papers	Keywords	Keywords	Paper
Journal Papers	43516	39760	84683	2.13
Conference Papers	51639	29997	48691	1.62
Misc Papers	23854	11963	23492	1.96
All Papers	119009	81720	156866	1.92

Table 2. Table 2: Log-Rank Test

Group	Number	Observed(O)	Expected(E)	(O-E)2/E	Chi Sq Test	p value
Group 1 (Journ)	38245	36432	33765	211	1306	<2e-16
Group 2 (Conf)	46252	39540	42207	169	1306	-

Full text

1 Introduction

Keywords can do more than just classify papers for information retrieval. They represent unique concepts associated with a paper and can be used as a proxy for knowledge creation. They can provide clues to the movement of ideas and trends within and across disciplines. For example, we can track hot terms like ’big data’ to see where they originate, what disciplines they spread to and how long they last within the literature. Evaluating how terms change over time can give us insights into how disciplines evolve and respond to new trends in technology and methods. This kind of information could be useful to researchers trying to capture the pulse of a field or help them avoid ’buzzy’ terms and instead focus on growing topics. Funding agencies would also find this useful for allocating funds to topics on the rise.

In this preliminary poster, we bring together methods, not brought together before, for measuring the burstiness of keywords within the field of Artificial Intelligence (AI)111We plan to extend this to other fields and over longer time periods. and lay groundwork for doing this more broadly. AI is known for its booms and busts. In fact, researchers often point to AI ”winters”. This boom and bust cycle make it ideal for studying the trendiness of jargon in the literature. We talk about the lessons learned and how these methods can be applied to other fields outside AI.

2 Methods

2.1 Data

Abstract data for this work comes from the Web of Science(WoS). A local copy of all WoS paper metadata resides in a MySQL database managed by the DataLab of the Information School at the University of Washington. From the data we filtered out all papers with subject_traditional as ’Computer Science, Artificial Intelligence’. The reason for doing this is to restrict our dataset to only computer science research oriented AI papers since we want to focus on knowledge creation and exclude papers on applications of AI. We performed some simple keyword data cleaning procedure such as combining similar terms like ’neural network’, ’neural-network’, ’NEURAL NETWORK’ into ’neural networks’. However, we did not want to extensively clean the keywords since a keyword could have different meanings depending on the context.222We plan to use novel techniques to cluster similar keywords together in future work. Additionally, each paper also contained a document type entry that indicated the type of paper. We used ’Proceedings Paper; Meeting’ as conference papers and ’Article; Article’ as journal papers. Conference papers include conferences such as IEEE Conference on Computer Vision and Pattern Recognition (CVPR) and journal papers include journals such as Neurocomputing. Miscellaneous papers include books, biographies, and editorial letters. Due to missing historical data, we restrict our analysis to papers published between 1990 and 2016.

2.2 Identifying Keyword Bursts

To measure burstiness of keywords we use Kleinberg’s bursty algorithm [1]. This is a fairly well vetted popular algorithm used in the scientometrics community to map fading and emerging themes [2]. The resultant burst weight of a keyword from the algorithm takes into account the proportion of papers containing that keyword and provides a metric for strength of influence of that keyword in that ’bursty’ time frame. For this initial work, we only considered keywords that appeared in atleast 20 papers. The distribution of terms in this threshold gives us a good representation of data to investigate the bursts. For the bursty analysis333Code can be found at www.github.com/kishorevasan/measuring-scientific-buzz, we created a year-by-keyword matrix using 74232 papers and 2770 keywords. The severe drop in keywords is due to the 20 paper requirement. With this matrix, we perform the burst detection algorithm, enumerate the keywords and present the timeline of top bursts in Fig 3.

2.3 Survival of Keywords

One of our main questions was to investigate how long terms last once they are newly introduced. To conduct this analysis, we restrict our dataset to keywords introduced between 2003 and 2014, thus allowing two years (2015 and 2016) for subsequent observation to see if they resurfaced. We chose this because it had good representation of terms and no major bursts in that time period. In this time frame, 38245(78.55% of overall) new keywords were introduced by conference papers and 46252(54.61% of overall) new keywords were introduced by journal papers. Two separate curves were fit for conference and journal papers as displayed in Fig 2. We used a non-parametric log rank test [4] to test if the survival curve of conference and journal keywords are identical.

3 Preliminary Results

Keyword bursts describe thematic bursts. The major theme for keywords in the early bursts seem to be related to neuroscience (’visual cortex’,’neurons’,’cortex’, ’neural networks’) and the representation side of machine learning (’knowledge representation’,’logic’,’learning’), whereas recent bursts focus more on computational concepts(’extreme learning machine’,’sparse representation’,’particle swarm optimization’). We notice this thematic shift over two decades that went from more neuroscience terminology, among the bursting terms, to the current ’bursty’ focus on novel computational concepts.

Another finding is that, only 8.85% of keywords made it to more than 4 papers. This indicates that very few keywords resurface in multiple papers after introduction, at least in AI. From the log-rank test results we observe that with 0.01 level of significance, conference and journal keyword don’t survive at the same rate. Keywords by journals tend to stick around longer than keywords by conferences. Publication venues are represented differently by different disciplines. We plan to look at this in other fields as well and speculate whether this is likely the case in other fields as well.

4 Conclusion

Our analysis on AI keywords reveal three preliminary findings. One, most terms die out before year one. Nearly 80% of the keywords don’t make it past year zero. Two, conferences seem to have shorter-lived keywords than journals. And, three, we notice two major thematic bursts in AI. The first burst was dominated by terms that were neuronally inspired (e.g., neural network, visual cortex), while the second major burst contained computationally oriented terms (e.g., particle swarm optimization, sparse representation). We plan to extend this work to other fields in order to test how well this model identifies these major changes.