Re-ranking for Machine Reading

For the Machine Learning for Language Processing course at Cambridge, I explored some extensions to our (then under-review) ACL 2020 paper “Machine Reading of Historical Events” (Honovich et al., 2020). In that paper, we explore how to order historical events chronologically based on short textual descriptions of events and, optionally, some contextual information extracted from Wikipedia.

In my report, I conducted two sets of experiments. The first consisted of adding attention to our architecture. The second was exploring a ranking formulating of this task using RankNet (Burges et al., 2005). Both of these gave mild performance improvements at best in the setting we cared about, however, I did find that if we only had year annotations for some datapoints, but a full chronological ordering of them, this approach netted more substantial improvements.

Happy to share the report and code upon request.