Interactive Construction of Complex Query Models - 2017 Annual Report
Activities:
We have deeply explored simulated interactive processes for retrieving sets of related names from text. This required first establishing an evaluation dataset which we did by using the TREC "list QA" task wherein a set of entities on a particular topic (e.g., drugs banned by the FDA) are to be found -- the idea is to use sets of entities that are similar to those that would not be trivially identified in advance without SearchIE technology.
Significant Results:
We have demonstrated (Foley et al., CIKM 2016) that target entities can be retrieved accurately and rapidly in response to a query without the use of natural language processing and relations between entities. This is an important verification step that retrieval can use a smaller set of features and still be effective.
We have determined that not all classes of features are useful for an interactive setting. This has resulted in our focusing mostly on features related to the semantic meaning of words (e.g., word embedding clusters).