Table 1.
Features used by support vector machine
| Feature | Description | |
| Baseline | TokenDistance | Number of tokens between markables |
| SentenceDistance | Number of sentences between markables | |
| ExactMatch | Markables are exact string matches | |
| StartMatch | Markables match at the start | |
| EndMatch | Markables match at the end | |
| SoonStr | Markables match besides determiners | |
| Pronoun1 | Proposed antecedent is pronoun | |
| Pronoun2 | Proposed anaphor is pronoun | |
| Definite1 (A) | Proposed antecedent is definite | |
| Definite2 | Proposed anaphor is definite | |
| Demonstrative2 | Proposed anaphor is demonstrative | |
| NumberMatch (A) | Markables have same number | |
| WnClass (A) | Markables have same named entity semantic category | |
| Alias | Markables have UMLS Concept Unique Identifier (CUI) overlap | |
| ProStr (A) | Markables are same and are pronouns | |
| SoonStrNonpro | Markables are same and are not pronouns | |
| WordOverlap | Markables share at least one word | |
| WordSubstr | With stopwords removed, one is substring of the other | |
| BothDefinites | Both markables are definite | |
| BothPronouns | Both markables are pronouns | |
| Indefinite (A) | Antecedent is indefinite | |
| Pronoun | Antecedent is pronoun and anaphor is not | |
| ClosestComp | Antecedent is closest semantically compatible markable | |
| NPHead (A) | Antecedent span ends noun phrase (NP) span | |
| Anaph | Probability output by anaphoricity classifier (NP markables only) | |
| PermStrDist | String similarity under various permutations | |
| Manually selected syntactic features | PathLength | Length of path between markables in syntax tree |
| NPunderVP1 | Antecedent node is child of VP (verb phrase) node | |
| NPunderVP2 | Anaphor node is child of VP node | |
| NPunderS1 | Antecedent node is child of S (sentence) node | |
| NPunderS2 | Anaphor node is child of S node | |
| NPunderPP1 | Antecedent node is child of PP (prepositional phrase) node | |
| NPunderPP2 | Anaphor node is child of PP node | |
| NPSubj1 | Antecedent node has SBJ (subject) function tag | |
| NPSubj2 | Anaphor node has SBJ function tag | |
| NPSubjBoth | Both nodes have SBJ function tag | |
| Automatically selected syntactic features | Path n-grams | See text in the ‘Features’ section |
Italicized features indicate those directly taken from Ng and Cardie.22 ‘(A)’ indicates the feature is used in the anaphoricity classifier.