MIT Press

Table B.1

Features for determining whether c fills iarg_n of predicate p. For each mention f (denoting a filler) in the coreference chain c′, p_f, and arg_f are the predicate and argument position of f. Unless otherwise noted, all argument positions (e.g., arg_n and iarg_n) should be interpreted as the integer label n instead of the underlying word content of the argument. The & symbol denotes concatenation; for example, a feature value of “p&iarg_n” for the iarg₀ position of sale would be “sale-0.” Features marked with an asterisk (*) are explained in Section 4.2. Features marked with a dagger (†) require external text corpora that have been automatically processed by existing NLP components (e.g., SRL systems). The final column gives a heuristic ranking score for the features across all evaluation folds (see Section 6.2 for discussion).

#	Feature value description	Importance score
1*	For every f, p_f&arg_f&p&iarg_n.	8.2
2*	Sentence distance from c to p.	4.0
3*	For every f, the head word of f& the verbal form of p&iarg_n.	3.6
4*	Same as 1 except generalizing p_f and p to their WordNet synsets.	3.3
5*	Same as 3 except generalizing f to its WordNet synset.	1.0
6	Whether or not c and p are themselves arguments to the same predicate.	1.0
7	p& the semantic head word of p's right sibling.	0.7
8	Whether or not any arg_f and iarg_n have the same integer argument position.	0.7
9*	Frame element path between arg_f of p_f and iarg_n of p in FrameNet (Baker, Fillmore, and Lowe 1998).	0.6
10	Percentage of elements in c′ that are subjects of a copular for which p is the object.	0.6
11	Whether or not the verb forms of p_f and p are in the same VerbNet class and arg_f and iarg_n have the same thematic role.	0.6
12	p& the last word of p's right sibling.	0.6
13*†	Maximum targeted PMI between arg_f of p_f and iarg_n of p.	0.6
14	p& the number of p's right siblings.	0.5
15	Percentage of elements in c′ that are objects of a copular for which p is the subject.	0.5
16	Frequency of the verbal form of p within the document.	0.5
17	p& the stemmed content words in a one–word window around p.	0.5
18	Whether or not p's left sibling is a quantifier (many, most, all, etc.). Quantified predicates tend not to take implicit arguments.	0.4
19	Percentage of elements in c′ that are copular objects.	0.4
20	TF cosine similarity between words from arguments of all p_f and words from arguments of p.	0.4
21	Whether the path defined in 9 exists.	0.4
22	Percentage of elements in c′ that are copular subjects.	0.4
23*	For every f, the VerbNet class/role of p_f/arg_f& the class/role of p/iarg_n.	0.4
24	Percentage of elements in c′ that are indefinite noun phrases.	0.4
25*	p& the syntactic head word of p's right sibling.	0.3
26	p& the stemmed content words in a two-word window around p.	0.3
27*†	Minimum selectional preference between any f and iarg_n of p. Uses the method described by (Resnik 1996) computed over an SRL-parsed version of the Penn TreeBank and Gigaword (Graff 2003) corpora.	0.3
28	p&p's synset in WordNet.	0.3
29†	Same as 27 except using the maximum.	0.3
30	Average per–sentence frequency of the verbal form of p within the document.	0.3
31	p itself.	0.3
32	p& whether p is the head of its parent.	0.3
33*†	Minimum coreference probability between arg_f of p_f and iarg_n of p.	0.3
34	p& whether p is before a passive verb.	0.3
35	Percentage of elements in c′ that are definite noun phrases.	0.3
36	Percentage of elements in c′ that are arguments to other predicates.	0.3
37	Maximum absolute sentence distance from any f to p.	0.3
38	p&p's syntactic category.	0.2
39	TF cosine similarity between the role description of iarg_n and the concatenated role descriptions of all arg_f.	0.2
40	Average TF cosine similarity between each arg_n of each p_f and the corresponding arg_n of p, where ns are equal.	0.2
41	Same as 40 except using the maximum.	0.2
42	Same as 40 except using the minimum.	0.2
43	p& the head of the following prepositional phrase's object.	0.2
44	Whether any f is located between p and any of the arguments annotated by NomBank for p. When true, this feature rules out false positives because it implies that the NomBank annotators considered and ignored f as a local argument to p.	0.2
45	Number of elements in c′.	0.2
46	p& the first word of p's right sibling.	0.2
47	p& the grammar rule that expands p's parent.	0.2
48	Number of elements in c′ that are arguments to other predicates.	0.2
49	Nominal form of p&iarg_n.	0.2
50	p& the syntactic parse tree path from p to the nearest passive verb.	0.2
51	Same as 37 except using the minimum.	0.2
52†	Same as 33 except using the average.	0.2
53	Verbal form of p&iarg_n.	0.2
54	p& the first word of p's left sibling.	0.2
55	Average per-sentence frequency of the nominal form of p within the document.	0.2
56	p& the part of speech of p's parent's head word.	0.2
57†	Same as 33 except using the maximum.	0.2
58	Same as 37 except using the average.	0.1
59*	Minimum path length between arg_f of p_f and iarg_n of p within VerbNet (Kipper 2005).	0.1
60	Frequency of the nominal form of p within the document.	0.1
61	p& the number of p's left siblings.	0.1
62	p&p's parent’s head word.	0.1
63	p& the syntactic category of p's right sibling.	0.1
64	p&p's morphological suffix.	0.1
65	TF cosine similarity between words from all f and words from the role description of iarg_n.	0.1
66	Percentage of elements in c′ that are quantified noun phrases.	0.1
67*	Discourse relation whose two discourse units cover c (the primary filler) and p.	0.1
68	For any f, the minimum semantic similarity between p_f and p using the method described by (Wu and Palmer 1994) over WordNet (Fellbaum 1998).	0.1
69	p& whether or not p is followed by a prepositional phrase.	0.1
70	p& the syntactic head word of p’s left sibling.	0.1
71	p& the stemmed content words in a three-word window around p.	0.1
72	Syntactic category of c&iarg_n& the verbal form of p.	0.1
73	Nominal form of p& the sorted integer argument indexes (the ns) from all arg_n of p.	0.1
74	Percentage of elements in c′ that are sentential subjects.	0.1
75	Whether or not the integer position of any arg_f equals that of iarg_n.	0.1
76†	Same as 13 except using the average.	0.1
77†	Same as 27 except using the average.	0.1
78	p&p's parent’s syntactic category.	0.1
79	p& the part of speech of the head word of p’s right sibling.	0.1
80	p& the semantic head word of p’s left sibling.	0.1
81†	Maximum targeted coreference probability between arg_f of p_f and iarg_n of p. This is a hybrid feature that calculates the coreference probability of Feature 33 using the corpus tuning method of Feature 13.	0.1

#	Feature value description	Importance score
1*	For every f, p_f&arg_f&p&iarg_n.	8.2
2*	Sentence distance from c to p.	4.0
3*	For every f, the head word of f& the verbal form of p&iarg_n.	3.6
4*	Same as 1 except generalizing p_f and p to their WordNet synsets.	3.3
5*	Same as 3 except generalizing f to its WordNet synset.	1.0
6	Whether or not c and p are themselves arguments to the same predicate.	1.0
7	p& the semantic head word of p's right sibling.	0.7
8	Whether or not any arg_f and iarg_n have the same integer argument position.	0.7
9*	Frame element path between arg_f of p_f and iarg_n of p in FrameNet (Baker, Fillmore, and Lowe 1998).	0.6
10	Percentage of elements in c′ that are subjects of a copular for which p is the object.	0.6
11	Whether or not the verb forms of p_f and p are in the same VerbNet class and arg_f and iarg_n have the same thematic role.	0.6
12	p& the last word of p's right sibling.	0.6
13*†	Maximum targeted PMI between arg_f of p_f and iarg_n of p.	0.6
14	p& the number of p's right siblings.	0.5
15	Percentage of elements in c′ that are objects of a copular for which p is the subject.	0.5
16	Frequency of the verbal form of p within the document.	0.5
17	p& the stemmed content words in a one–word window around p.	0.5
18	Whether or not p's left sibling is a quantifier (many, most, all, etc.). Quantified predicates tend not to take implicit arguments.	0.4
19	Percentage of elements in c′ that are copular objects.	0.4
20	TF cosine similarity between words from arguments of all p_f and words from arguments of p.	0.4
21	Whether the path defined in 9 exists.	0.4
22	Percentage of elements in c′ that are copular subjects.	0.4
23*	For every f, the VerbNet class/role of p_f/arg_f& the class/role of p/iarg_n.	0.4
24	Percentage of elements in c′ that are indefinite noun phrases.	0.4
25*	p& the syntactic head word of p's right sibling.	0.3
26	p& the stemmed content words in a two-word window around p.	0.3
27*†	Minimum selectional preference between any f and iarg_n of p. Uses the method described by (Resnik 1996) computed over an SRL-parsed version of the Penn TreeBank and Gigaword (Graff 2003) corpora.	0.3
28	p&p's synset in WordNet.	0.3
29†	Same as 27 except using the maximum.	0.3
30	Average per–sentence frequency of the verbal form of p within the document.	0.3
31	p itself.	0.3
32	p& whether p is the head of its parent.	0.3
33*†	Minimum coreference probability between arg_f of p_f and iarg_n of p.	0.3
34	p& whether p is before a passive verb.	0.3
35	Percentage of elements in c′ that are definite noun phrases.	0.3
36	Percentage of elements in c′ that are arguments to other predicates.	0.3
37	Maximum absolute sentence distance from any f to p.	0.3
38	p&p's syntactic category.	0.2
39	TF cosine similarity between the role description of iarg_n and the concatenated role descriptions of all arg_f.	0.2
40	Average TF cosine similarity between each arg_n of each p_f and the corresponding arg_n of p, where ns are equal.	0.2
41	Same as 40 except using the maximum.	0.2
42	Same as 40 except using the minimum.	0.2
43	p& the head of the following prepositional phrase's object.	0.2
44	Whether any f is located between p and any of the arguments annotated by NomBank for p. When true, this feature rules out false positives because it implies that the NomBank annotators considered and ignored f as a local argument to p.	0.2
45	Number of elements in c′.	0.2
46	p& the first word of p's right sibling.	0.2
47	p& the grammar rule that expands p's parent.	0.2
48	Number of elements in c′ that are arguments to other predicates.	0.2
49	Nominal form of p&iarg_n.	0.2
50	p& the syntactic parse tree path from p to the nearest passive verb.	0.2
51	Same as 37 except using the minimum.	0.2
52†	Same as 33 except using the average.	0.2
53	Verbal form of p&iarg_n.	0.2
54	p& the first word of p's left sibling.	0.2
55	Average per-sentence frequency of the nominal form of p within the document.	0.2
56	p& the part of speech of p's parent's head word.	0.2
57†	Same as 33 except using the maximum.	0.2
58	Same as 37 except using the average.	0.1
59*	Minimum path length between arg_f of p_f and iarg_n of p within VerbNet (Kipper 2005).	0.1
60	Frequency of the nominal form of p within the document.	0.1
61	p& the number of p's left siblings.	0.1
62	p&p's parent’s head word.	0.1
63	p& the syntactic category of p's right sibling.	0.1
64	p&p's morphological suffix.	0.1
65	TF cosine similarity between words from all f and words from the role description of iarg_n.	0.1
66	Percentage of elements in c′ that are quantified noun phrases.	0.1
67*	Discourse relation whose two discourse units cover c (the primary filler) and p.	0.1
68	For any f, the minimum semantic similarity between p_f and p using the method described by (Wu and Palmer 1994) over WordNet (Fellbaum 1998).	0.1
69	p& whether or not p is followed by a prepositional phrase.	0.1
70	p& the syntactic head word of p’s left sibling.	0.1
71	p& the stemmed content words in a three-word window around p.	0.1
72	Syntactic category of c&iarg_n& the verbal form of p.	0.1
73	Nominal form of p& the sorted integer argument indexes (the ns) from all arg_n of p.	0.1
74	Percentage of elements in c′ that are sentential subjects.	0.1
75	Whether or not the integer position of any arg_f equals that of iarg_n.	0.1
76†	Same as 13 except using the average.	0.1
77†	Same as 27 except using the average.	0.1
78	p&p's parent’s syntactic category.	0.1
79	p& the part of speech of the head word of p’s right sibling.	0.1
80	p& the semantic head word of p’s left sibling.	0.1
81†	Maximum targeted coreference probability between arg_f of p_f and iarg_n of p. This is a hybrid feature that calculates the coreference probability of Feature 33 using the corpus tuning method of Feature 13.	0.1

View Large

This Feature Is Available To Subscribers Only