On "Redundancy'' in Selecting Attributes for Generating Referring Expressions
Philipp Spanger, Takehiro Kurosawa, Takenobu Tokunaga · Institutional Repositories DataBase (IRDB) · 2008
We seek to develop an efficient algorithm selecting attributes that approximates human selection. In contrast to previous work we sought to combine the strengths of cognitive theories and simple learning algorithms. We then developed a new algorithm for attribute selection based on observations from a corpus, which outperformed a simple base algorithm by a significant margin. We then carried out a detailed comparison between our algorithm and Reiter & Dale’s “Incremental Algorithm”. In terms of achieving a human-like attribute selection, the overall performance of both algorithms is fundamentally equivalent, while differing in the handling of redundancy in selected attributes. We further investigated this phenomenon and draw some conclusions for further improvement of attribute-selection algorithms.