As scienti c communities grow and evolve, there is emergence of new techniques and decline of old ones. The tremendous amount of research publications available online aims to solve a lot of interesting problems. With time, some of the elds have been studied well and research problems solved to a great extent. However, there are few di cult research problems which are yet not solved completely and interests a lot of researchers. In this paper, we aim to nd research elds which are saturated and research elds which need to be explored yet. We rst extract research problems in a semi supervised manner using a proven bootstrap framework from scienti c literature of the last fty years. We show how a simple statistics based model on top of the research problems extracted can nd the saturated elds and grand challenges in any domain of computer science.
A consistently thriving global research community has over decades produced a colossal amount of research papers that are published online, which makes it crucial to organize this huge bulk of information systematically so that upcoming researchers can navigate through e ciently and continue to push boundaries of scienti c research. Such an organization over intellectual information will not only boost the rate of further research work but also augment researchers with a better holistic view of development in research and the directions in which it is evolving into. One of the rst elementary steps we take as researchers is to gure out which problems to focus on solving, and structured analysis on present research status will help researchers identify critical problems and also give insight about how they developed across time. Due to this it will be easier to realize if particular problem has got no recent improvement in the recent past and has moved into a thriving application and so on. Analysis is the foundation to organization of cumulative knowledge garnered by the research community in decades, and this paper deals with this rst step in direction.
[
A recent challenge on Scienti c Information Extraction (ScienceIE) [
As the rst step of time based analysis, we aim to nd saturated elds and grand challenges. We de ne saturated elds as those research problems which have been studied to a great extent and nothing much is left to achieve in them. On the other hand grand challenges are de ned as those problems which have been tried to solve over a large period of time and are still worked upon extensively. 2
Saturated Problems: Problems which were very actively studied in the yesteryears and are now solved to a great extent. Example, parts of speech tagging in NLP.
Grand Challenges: Problems which were de ned in yesteryears and are still worked upon extensively. Example, machine translation in NLP. Research during the 1980s typically relied on translation through some variety of intermediary linguistic representation involving morphological, syntactic, and semantic analysis. In current times, research has focused on moving from domain speci c systems to domain independent translation systems.
Our approach is based on a proven method followed by [
to the conference in which they were published. Then 8 papers in the same
group, we cluster their extracted phrases by running DBSCAN [
From the rst step, we have research problems which have been studied as \AIM" for the last fty years. We also have techniques\METHOD" used to solve these problems over these years. We rst extract data for each research eld, p, and nd the number of times paper published on them for each of the years in the range 1971 to 2013.
{ Count vs year plot for such problems should show a steep decline in the current years. { Based on exploratory data analysis we came up with the following rules for nding saturated problems from the data collected above { We list a problem p as a saturated problem if:
T1 is the rst year when the problem appeared in the literature. T2 is the last time when the problem appeared in the literature.
Count of p appearing as aim in T2 should be less than the count of p appearing as aim in T1 Peak of count vs year plot should have occured much before 2013. Suppose problem p1 has peak at time t1 and problem p2 has peak at time t2. P1 is a better candidate for saturated problem than p2 if the di erence between T2 of p1 and t1 is more than the di erence between T2 of p2 and t2.
{ Count vs year plot for such problems should start from yester years and be consistent over the time. Peaks should be current years as well as yester years. { Based on exploratory data analysis we came up with the following rules for nding grand challenges from the data collected above
We list a problem p as a grand challenge if: ∗ T1 is the rst year when the problem appeared in the literature. T2 is the last time when the problem appeared in the literature. ∗ T1 for problem p to be classi ed as a grand challenge should be before 2000 and T2 after 2010. Time span between T1 and T2 should be more than 10 years. ∗ Count of p appearing as an aim in T2 should be more than some threshold. This is to rule out the edge cases where there is occurrence of few counts in current years.
We rank these problems based on the following formula:
∗ To capture the fact that more the span of the problem over the years,
more likely it is a grand challenge; we propose rank to be directly
proportional to the number of years it spans to.
∗ To capture the fact the count needs to be consistent over the years;
we propose rank to be inversely proportional to Pin=1(count[i]
count[i 1]) where i iterates over all the years in which a problem p
occurs.
All experiments were done on DBLP citation network version 7. We chose DBLP
dataset to get a wide variety of research papers from di erent domains over a
large time period. It has 2,244,021 papers and 4,354,534 citation relationships.
After pruning out some papers and data cleaning we came up with 332,793
papers having 1,508,560 citation links. These papers range from 1936 to 2013.
However for the period 1936- 1971, the number of papers available were relatively
very less for time based analysis. So we pruned the data further and worked on
papers from 1971 to 2013.
results, we extracted top 100 problems in both the categories. We represent our
results as word clouds [
Fig. 2. Word Cloud for Saturated Problems 5
In this paper, we show the temporal analysis of scienti c literature by extracting saturated problems and grand challenges. We propose this as the rst step towards time based analysis. We plan to further do time based analysis by nding transition time for problems where transition time is de ned as the time period where a problem starts occurring as method instead of aim.