Interesting paper (it's being presented as I type this):
The Trend is the Cycle: Job Polarization and Jobless Recoveries, by Nir Jaimovich and Henry E. Siu, NBER: Abstract Job polarization refers to the recent disappearance of employment in occupations in the middle of the skill distribution. Jobless recoveries refers to the slow rebound in aggregate employment following recent recessions, despite recoveries in aggregate output. We show how these two phenomena are related. First, job polarization is not a gradual process; essentially all of the job loss in middle-skill occupations occurs in economic downturns. Second, jobless recoveries in the aggregate are accounted for by jobless recoveries in the middle-skill occupations that are disappearing.
1 Introduction In the past 30 years, the US labor market has seen the emergence of two new phenomena: "job polarization" and "jobless recoveries." Job polarization refers to the increasing concentration of employment in the highest- and lowest-wage occupations, as job opportunities in middle-skill occupations disappear. Jobless recoveries refer to periods following recessions in which rebounds in aggregate output are accompanied by much slower recoveries in aggregate employment. We argue that these two phenomena are related.
Consider first the phenomenon of job polarization. Acemoglu (1999), Autor et al. (2006), Goos and Manning (2007), and Goos et al. (2009) (among others) document that, since the 1980s, employment is becoming increasingly concentrated at the tails of the occupational skill distribution. This hollowing out of the middle has been linked to the disappearance of jobs focused on "routine" tasks -- those activities that can be performed by following a well-defined set of procedures. Autor et al. (2003) and the subsequent literature demonstrates that job polarization is due to progress in technologies that substitute for labor in routine tasks.1
In this same time period, Gordon and Baily (1993), Groshen and Potter (2003), Bernanke (2003), and Bernanke (2009) (among others) discuss the emergence of jobless recoveries. In the past three recessions, aggregate employment continues to decline for years following the turning point in aggregate income and output. No consensus has yet emerged regarding the source of these jobless recoveries.
In this paper, we demonstrate that the two phenomena are connected to each other. We make two related claims. First, job polarization is not simply a gradual phenomenon: the loss of middle-skill, routine jobs is concentrated in economic downturns. Specifically, 92% of the job loss in these occupations since the mid-1980s occurs within a 12 month window of NBER dated recessions (that have all been characterized by jobless recoveries). In this sense, the job polarization "trend" is a business "cycle" phenomenon. This contrasts to the existing literature, in which job polarization is oftentimes depicted as a gradual phenomenon, though a number of researchers have noted that this process has been accelerated by the Great Recession (see Autor (2010); and Brynjolfsson and McAfee (2011)). Our First point is that routine employment loss happens almost entirely in recessions.
Our second point is that job polarization accounts for jobless recoveries. This argument is based on three facts. First, employment in the routine occupations identified by Autor et al. (2003) and others account for a significant fraction of aggregate employment; averaged over the jobless recovery era, these jobs account for more than 50% of total employment. Second, essentially all of the contraction in aggregate employment during NBER dated recessions can be attributed to recessions in these middle-skill, routine occupations. Third, jobless recoveries are observed only in these disappearing, middle-skill jobs. The high- and low-skill occupations to which employment is polarizing either do not experience contractions, or if they do, rebound soon after the turning point in aggregate output. Hence, jobless recoveries can be traced to the disappearance of routine occupations in recessions. Finally, it is important to note that jobless recoveries were not observed in routine occupations (nor in aggregate employment) prior to the era of job polarization. ...
 See also Firpo et al. (2011), Goos et al. (2011), and the references therein regarding the role of outsourcing and offshoring in job polarization
Here are few graphs showing routine versus non-routine employment changes. The point is that in the period of increased job polarization, most of the job losses have occurred during recessions (the first graph is non-routine cognitive, the second is non-routine manual, and the third is routine -- the third graph shows the best, after 1990 job losses are no recovered after the recession ends as they were in earlier years
Interestingly, the paper argues the explanation for jobless recoveries is not an education story:
The share of low educated workers in the labor force (i.e., those with high school diplomas or less) has declined in the last three decades, and these workers exhibit greater business cycle sensitivity than those with higher education. It is thus reasonable to conjecture that the terms "routine" and "low education" are interchangeable. In what follows, we show that this is not the case.
And it's not a manufacturing story:
we first demonstrate that job loss in manufacturing accounts for only a fraction of job polarization. Secondly, we show that the jobless recoveries experienced in the past 30 years cannot be explained by jobless recoveries in the manufacturing sector.
So what story is it? The answer is in the conclusion to the paper:
In the last 30 years the US labor market has been characterized by job polarization and jobless recoveries. In this paper we demonstrate how these are related. We first show that the loss of middle-skill, routine jobs is concentrated in economic downturns. In this sense, the job polarization trend is a business cycle phenomenon. Second, we show that job polarization accounts for jobless recoveries. This argument is based on the fact that almost all of the contraction in aggregate employment during recessions can be attributed to job losses in middle-skill, routine occupations (that account for a large fraction of total employment), and that jobless recoveries are observed only in these disappearing routine jobs since job polarization began. We then propose a simple search-and-matching model of the labor market with occupational choice to rationalize these facts. We show how a trend in routine-biased technological change can lead to job polarization that is concentrated in downturns, and recoveries from these recessions that are jobless.
That is, in recessions, the job separation rate isn't much different than in the past. But the job finding rate is much lower. Thus, the story is that recessions generate job separations, and "In the recession, job separations were concentrated among the middle-skill, routine workers. The recovery in aggregate employment then depends on the post-recession job finding rate of these workers now searching," and this finding rate is low (and when jobs are found, the outcome is polarizing).