Summarising LinkedIn : generating human-like texts with variation

Abstract

With the rise to prominence of social networking over the last 10 years or so, the presence of online profiles has offered a large amount of data readily available on the internet. Using Linkedln as an example, they have recently reached a milestone of 300 million registered users and the website has been labelled as one of the fastest growing social networks. Having seen the emergence of such online profiles, and being fascinated by the issue of variation and probability within natural language generation, the idea of coupling these two things together seemed intriguing to say the least. This dissertation focuses on an NLG system called SLIP (Summarising Linkedln Profiles), which is able to generate summaries based on the data found on a person's Linked In profile. In order to build a system that produces texts of a similar quality to those produced by humans a study was undertaken where participants were asked to summarise excerpts of Linkedln profiles. The data collected was then analysed and the results were replicated in the SLIP system, with a particular focus on variation at both word and sentence level. The system was implemented using Java, making use of the SimpleNLG library to aid with the realisation process. A human based evaluation was then undertaken, where participants were asked to identify the computer generated text when presented with a both a text generated by the SLIP system, as well as a text written by a human. In just over 59% of cases, participants correctly identified one of the texts as being computer generated and one of the texts as being written by a human. However, 41% of participants did not correctly identify the two texts, and such results are analysed in some detail, along with the wide array of interesting comments left by participants.