How to Make Manual Conjunctive Normal Form Queries Work in Patents Search
This year we focused on the Technology Survey (TS) task: Given a natural language description of the topic, look for related patents about that topic. The task is close to an ad hoc retrieval task, except for the additional information of the specific chemicals or chemical reactions that the user cares about. Since there are only 6 topics for the TS task, this notebook paper is more of a case study report, than the ordinary TREC report with significance tests. We found that on average, with the infAP measure, manually created conjunctive normal form queries performed similarly as automatic keyword search with some tuning of term weights. Manual queries do not seem to always help, especially when initial keyword performance is high, but can give large improvements on difficult queries. We also used the same querying strategy in the Patent Olympics 2011 ChemAthlon task, and also include some of the ChemAthlon cases in this report. Since CNF queries are strictly more expressive than keyword queries, we try to identify problems that may have caused the manual CNF queries to be seen sometimes performing worse than the automatic keyword queries.