Target Type Identification¶
A characteristic property of entities is that they are typed. Naturally, entity-bearing queries may be complemented with target types (types of its relevant entities). Given a search query, Target Type Identification (TTI) is the task of returning a ranked list of types from an underlying type taxonomy. Entity retrieval performance can be significantly improved when explicit target type information is identified for a query.
The following methods for target type identification are implemented in Nordlys:
- EC: the Entity Centric (EC) method, as described in [Balog and Neumayer, 2012]. Both BM25 and LM models can be used as a retrieval model. This method fits the late fusion design pattern in [Zhang and Balog, 2017].
- TC: the Type Centric (TC) method based on [Balog and Neumayer, 2012]. Both BM25 and LM models can be used as a retrieval model. This method fits the early fusion design pattern in [Zhang and Balog, 2017].
- LTR: the Learing-To-Rank (LTR) method, as proposed in [Garigliotti et al., 2017]. This method establishes the state-of-the-art performance in TTI.
Usage¶
Benchmark results¶
Below, we present retrieval results on the target type identification in [Garigliotti et al., 2017].
Method | NDCG@1 | NDCG@5 |
---|---|---|
EC, BM25 (K = 20) | 0.1490 | 0.3223 |
EC, LM (K = 20) | 0.1417 | 0.3161 |
TC, BM25 | 0.2015 | 0.3109 |
TC, LM | 0.2341 | 0.3780 |
LTR | 0.4842 | 0.6355 |
The corresponding files with rankings can be found on Github, specifically under output directory.
References¶
- Darío Garigliotti, Faegheh Hasibi, and Krisztian Balog. Target Type Identification for Entity-Bearing Queries. In: 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’17). [BIB] [PDF]