📄 Joint Parsing and Named Entity Recognition

Summary

Large NLP systems often run a named entity recognizer and a parser independently. Their outputs can then disagree: an entity span may cross constituent brackets, which makes the two annotations hard to combine. Finkel and Manning address this with a single model that does constituency parsing and NER together. They extend the discriminative, feature-rich CRF-CFG parser of Finkel and Manning (2008). Entity information is encoded inside the tree: each entity is forced to sit under one phrasal node, and the labels below that node carry the entity type. To counter the sparsity this creates, the grammar and lexicon are smoothed with synthetic rules. The model is trained and tested on six broadcast-news portions of OntoNotes Release 2.0. Compared with separate models using the same features, the joint model's outputs are always consistent. It also improves NER on every subset and parsing on most of them, by up to 1.36 F1 for parsing and about 9 F1 for NER. This matters because earlier joint-modeling attempts, such as parsing plus SRL, usually failed to beat decoupled pipelines.

Key points

• Data: LDC2008T04 OntoNotes Release 2.0, evaluated separately on six sources (ABC, CNN, MNB, NBC, PRI, VOA). Parses were scored with evalB and entities with conlleval.
• Parsing results: joint-model F1 rose on CNN (77.03→77.71), NBC (61.63→62.93), PRI (76.35→77.41) and VOA (76.15→77.51). It was nearly flat on MNB and slightly lower on ABC (70.15→69.99).
• NER results: F1 improved on all six subsets, for example VOA 79.23→88.18, MNB 62.21→66.49 and CNN 75.78→78.70.
• Significance: on the pooled data, the improvements in both precision and recall reached significance at the p < 10⁻⁴ level, using sentence-level stratified shuffling (as in Bikel's comparator).
• Features:
  • Word shapes.
  • Clark (2000) distributional clusters (200 clusters, trained on BNC plus English Gigaword).
  • Standard NER features in the style of Finkel et al. (2005).
  • The parser features of Finkel and Manning (2008), computed with entity marks stripped.
  • Joint features over entity-augmented rules and labels.

Interesting ideas

• I1. Encode one layer inside the other's structure. When an entity covers only some contiguous children of a constituent, the authors insert a dedicated NamedEntity node, so every entity becomes exactly one constituent. Consistency between layers is then built into the representation instead of being checked afterwards. This trick could carry over to other span-level annotations layered on trees.
• I2. Reuse sequence-model features inside a tree model. Because binarized trees join every pair of adjacent words through some binary rule, a linear-chain CRF's pairwise label features can be computed at those rules. Per-word features are computed at the leaves. This lets mature tagger feature sets be ported into a parser.
• I3. Smooth an augmented grammar with synthetic rules. For each entity-annotated rule, the authors add copies with the entity mark removed and copies relabeled with every other entity type. They also add entity-typed variants of rules with NP children. This is a cheap remedy whenever label refinement fragments the grammar.
• I4. Prune the tag space with distributional clusters. A word may take the POS tags seen with it in training, plus tags seen with other words in its distributional cluster. This keeps decoding tractable without being too strict for rare words.
• I5. Nesting falls out of the representation. Rules such as organization → GPE organization can encode nested entities, so the model could learn how they decompose if the training data annotated them, as GENIA does.