Natural Language Processing (Almost) from Scratch

6. The Temptation

1. What happened when Collobert et al. added standard task-specific features?
   Moderate gains, reaching or passing the benchmarks: two-character suffixes for POS (97.29), the CoNLL gazetteer for NER (89.59), POS tags for chunking (94.32), chunk tags for SRL (74.72). Ensembles of ten networks gave small further gains.
   • Collobert et al., 2011, Natural Language Processing (Almost) from Scratch (Journal of Machine Learning Research), 6. The Temptation, pp. 2518–2521
2. Why did a gazetteer help the NER network so much?
   Gazetteers contain multi-word entities; the network only had per-word representations. united and bicycle are probably non-entities alone, but United Bicycle may be one, which needs higher-level information.
   • Collobert et al., 2011, Natural Language Processing (Almost) from Scratch (Journal of Machine Learning Research), 6. The Temptation, pp. 2518–2521