The file describes the main differences between the two versions of the data library.

1) We re-estimated our CBOW model. This re-estimation reflects improvements in the spell-checking of our on-line job ad text, which was an input into the CBOW model estimation. Previously, our on-line job ad text included words like 'eventsparticipatingin,' 'managementadaptabilitystress,' and 'servicemanagementretail.' We have spell-checked these words. Removal of words like this alters the output of the estimation of the CBOW model.  

2) Within the list of words associated with the Spitz-Oener nonroutine manual task, we removed the word "service."  We did so for two reasons. First, something we didn’t appreciate until recently, in many cases “service” referred not to a task the worker was expected to perform, but instead to the employer's line of business (e.g., “we are a financial service firm”). Second, the single word “service” represented a large fraction, slightly more than half, of the mentions of nonroutine manual words.

3) Within the list of technologies we searched for, we added eight: CNC, IBM 5520, POS (Point of Sale technologies), Sabre, Vydec, Xerox 630, Xerox 800, and Xerox 860.

4) We added a mapping of job titles to OCC (ACS Occupation) codes, and constructed a version of our database based on this occupation classification scheme. As a warning: While we have checked the accuracy of our job title to SOC mapping by comparing SOC occupations' O*NET measures to the analogous measures based on our newspaper data, it is impossible to perform the same comparison with the OCC-based occupation classification. Moreover, we have reason to believe that the job title to SOC mapping may be more accurate than the job title to OCC mapping. In particular, there are substantially fewer job titles within the Census 2000 Occupation Index (giving mappings between job titles and OCC codes) than there are in the union of the BLS Sample of Reported Titles or Alternate Sample of Reported Titles (giving mappings between job titles and SOC codes): 24 thousand versus 43 thousand. Since there are fewer job titles (for which we observe occupation codes) to match to with the OCC classification scheme, it is possible that this will lead to a less accurate final mapping between job titles that we observe in our newspaper data to occupation codes.