Training-Free Score Calibration for Complex Query Decomposition

Answering complex queries on incomplete knowledge graphs poses significant challenges, as models must infer their answers despite gaps in the available data. Previous research has addressed this problem by developing end-to-end architectures specifically designed for complex query answering. These models are difficult to interpret and require extensive data and computational resources for training. Alternatively, some approaches have focused on leveraging existing neural link predictors, which have been designed for simple queries, to handle complex queries. This approach reduces the amount of training examples needed and offers more transparent reasoning. However, the output scores of the neural link predictors may require calibration for effective interaction during the reasoning process and a special adaption function has to be learned to achieve this. In this work, (i) we show that depending on the query type, standard normalization methods are equally as effective as learning an adaption function. (ii) Furthermore, we replace the neural link predictor with a rule-based approach that does not require any score calibration. With such an approach we achieve new state-of-the-art results and increase the mean reciprocal ranks from 35.1% to 37.1% averaged across datasets and query types. (iii) We conduct comprehensive empirical analysis to support our claims (The code and data for all our experiments can be accessed here: https://figshare.com/s/4f1fbd5f5d2c4aca7c2e).