Subdeletion in Verb Phrase Ellipsis 
Paul G. Donecker 
Villanova University 
800 Lancaster Avenue 
Villanova, PA 19085 
donecker@monet.vill.edu 
Abstract 
This paper stems from an ongoing research 
project ~ on verb phrase ellipsis. The project's 
goals are to implement a verb phrase ellipsis 
resolution algorithm, automatically test the 
algorithm on corpus data, then automatically 
evaluate the algorithm against human-generated 
answers. The paper will establish the current 
status of the algorithm based on this automatic 
evaluation, categorizing current problem 
situations. An algorithm to handle one of these 
problems, the ease of subdeletion, will be 
described and evaluated. The algorithm attempts 
to detect and solve subdeletion by locating 
adjuncts of similar types in a verb phrase ellipsis 
and corresponding antecedent. 
1. Introduction 
A verb phrase ellipsis (VPE) exists when a 
sentence has an auxiliary verb but no verb phrase 
(VP). For example, in the sentence "Gather ye 
rosebuds while ye may," "may" is the beginning of a 
VPE. Its antecedent is "gather ye rosebuds." The 
research described in this paper is part of a project to 
automate the resolution of VPE occurrences, and also 
to automate the evaluation of the success of the VPE 
resolution (Hardt 1995). 
Based on these evaluations of the algorithm, 
several distinct categories of error situations have been 
determined. We have focused on errors in which the 
program selects the correct head verb as antecedent. 
These cases can be divided into the following 
categories: 1) too much material included from the 
antecedent, 2) not enough much material included 
from the antecedent, 3) discontinuous antecedents, and 
4) miscellaneous. 
For a subset of case 1, subdeletion, an algorithm 
derived from (Lappin and McCord, 1990) is evaluated 
1 This research was supported in part by NSF 
Career Grant, no. IRI-9502257. 
in regard to the Brown Corpus. 
2. Background 
Previous studies on evaluating discourse 
processing (e.g., Walker, 1989; Hobbs, 1978) have 
involved subjectively examining test cases to 
determine correctness. With the development of 
resources such as the Penn Treebank (Marcus, 
Santorini, and Marcinkiewicz, 1993), it has become 
possible to automate empirical tests of discourse 
processing systems to obtain a more objective 
measure of their success. Towards this end, an 
algorithm was implemented in a Common Lisp 
program called VPEAL (Verb Phrase Ellipsis 
Antecedent Locator) (Hardt, 1995), drawing on the 
Penn Treebank as input. The portion of the Penn 
Treebank examined--the Brown Corpus, about a 
million words--contains about 400 VPEs. 
Furthermore, to automatically evaluate the 
algorithm, utilities were developed to automatically 
test the output of VPEAL for correctness. The most 
recent version of VPEAL contained 18 sub-parts for 
ranking and choosing antecedents. Testing the 
program's performance involved finding the 
percentage of correct antecedents found by any or all 
of these algorithms. This was achieved by having 
human coders read plain text versions of the parsed 
passages, marking what they felt to be the antecedent. 
Antecedents selected by VPEAL were considered 
correct if they matched the antecedents selected by the 
coders. 
The remainder of this paper will describe the 
categories of errors observed, then describe an 
approach to reducing one category of errors. 
3. Categories of Errors 
The most recent version of VPEAL correctly 
selects 257 out of 380 antecedents from the Brown 
Corpus. We have divided the categories into the 
following categories: 
A. Incorrect verb: 90 cases. In these cases, 
VPEAL selected an incorrect head verb for the 
348 
antecedent. The causes of these errors are being 
evaluated. 
B. Incorrect antecedent but correct verb: 33 cases. 
VPEAL selected the correct verb to head the 
antecedent, but the selected antecedent was either 
incomplete or included incorrect information. These 
cases can be further divided into: 1) too much material 
included from the antecedent, 2) not enough much 
material included from the antecedent, 3) 
discontinuous antecedents, and 4) miscellaneous. 
These subcategories are described below. 
1. Too much material is included from the 
antecedent: 11 cases. 
Example (excerpt from Penn Treebank): 
produce humorous effects in his novels and tales 
as they did in the writing of Longstreet and 
Hooper and Harris 
VPE: did 
VPEAL's antecedent: produce humorous effects in his 
novels and tales 
Coder's antecedent: produce humorous effects 
Normally, an entire verb phrase is selected as the 
antecedent. In these cases, though, part of the selected 
antecedent was not required by the VPE. The most 
common situation (6 cases), as in the above example, 
was subdeletion--when the VPE structure contains a 
noun phrase or prepositional phrase which substitutes 
for a corresponding structure in the antecedent verb 
phrase. 
2. Not enough material is included from the 
antecedent: 10 cases. 
Example (excerpt from Penn Treebank): 
But even if we can not see the repulsive 
characteristics in this new image of America, 
foreigners can 
VPE: can 
VPEAL's antecedent: see the repulsive characteristics 
Coder's antecedent: see the repulsive characteristics in 
this new image of America 
By default, only text contained by the selected 
verb phrase is included in the antecedent. In these 
cases, however, human coders have selected text that 
is adjacent to but not parsed as contained by the verb 
phrase as part of the antecedent. It can be argued that 
these errors are not the fault of the VPEAL 
algorithm--that if text is parsed as not being a part of 
the verb phrase then it should still not be included 
when the verb phrase is chosen as the antecedent. If 
the above prepositional phrase "in this new image of 
America" were parsed as part of the verb phrase-- as 
indeed it should have been--then the algorithm would 
have derived the correct antecedent. 
3. Discontinuous antecedents--the correct 
antecedent is split into two parts: 5 cases. 
Example (excerpt from Penn Treebank): 
representing as I do today my wife 
VPE: do 
VPEAL's antecedent: representing 
Coder's antecedent: representing my wife 
This situation is similar to B2 in that the 
antecedent is incorrect because text not contained by 
the selected verb phrase should be included in the 
antecedent. In these cases, however, the reason the 
omitted text is not contained by the antecedent verb 
phrase is that an interposing phrase (in the example 
above, the VPE itself) occurs in the middle of the 
antecedent. 
4. Miscellaneous: 7 cases. 
4. Improving Performance in the Case of 
Subdeletion 
In this section an algorithm is described to reduce 
the errors in error category B 1 caused by subdeletion. 
Subdeletion is probably the most straightforward of 
the error categories. The problem category occurred 
when prepositional phrases and noun phrases in the 
antecedent verb phrases were unnecessary because of 
analogous phrases adjacent to the VPE. The proposed 
solution was to check whether the VPE has a sister 
node that is a prepositional phrase or noun phrase. If 
it does, and a phrase of the same type exists as a sister 
node to the head verb in the antecedent, then the 
phrase in the antecedent is removed. This is 
essentially the strategy outlined by Lappin and 
McCord (1990). Following are the specific steps to 
implementing the algorithm: 
1. Check if there are any prepositional phrases or 
noun phrases that are sister nodes to the antecedent 
head verb. 
2. Check if there are any prepositional phrases or 
noun phrases that are sister nodes to the VPE head 
verb. 
3. If a prepositional phrase or noun phrase is 
found in step 1, and a phrase of the same type is found 
349 
in step 2, then remove the phrase found in step 1 from 
the antecedent. 
For example, refer to the example from error case 
B. 1. Step 1 would locate the noun phrase "humorous 
effects" and the prepositional phrase "in his novels and 
tales" as sister nodes to the antecedent head verb 
"produce." 
Step 2 would locate the prepositional phrase "in 
the writing of Longstreet and Hooper and Harris" as a 
sister node to the VPE head verb "did." 
Step 3 would determine that a prepositional 
phrase exists after both the antecedent's head verb and 
the VPE and therefore would delete "in his novels and 
tales" from the antecedent, resulting in the correct 
antecedent, "produce humorous effects." 
This algorithm will correctly handle the 6 cases of 
subdeletion in the Brown Corpus. However, examples 
can be constructed for which this algorithm does not 
account. In the sentence "Julie drove to school on 
Friday, and Laura did on Saturday," for example, the 
VPE is "did" and the correct antecedent is "drove to 
school." In this example, two prepositional 
phrases--"to school" and "on Friday"--follow the 
anteeedent's head verb "drove." A prepositional 
phrase, "on Saturday," also exists following the VPE's 
head verb. Following the above algorithm, both 
prepositional phrases "to school" and "on Friday" 
would be deleted, resulting in an incorrect antecedent. 
The algorithm makes no provisions for cases 
containing multiple prepositional phrases and noun 
phrases. Fortunately, such situations seem rare, as 
none were found in the Brown Corpus. 
More significantly, the algorithm also assumes 
that analogous phrases following the antecedent and 
VPE always implies subdeletion. That is, it assumes 
that prepositional phrases or noun phrases following 
the VPE always implies that like phrases should be 
deleted from the antecedent. Again, it is possible to 
imagine a counterexample, for example, "Dad stayed 
in the Hilton like Morn did in Pittsburgh." Here, the 
above algorithm would incorrectly remove the 
prepositional phrase "in the Hilton." 
The expectation was that these counter examples 
would be less frequent than the cases in which the 
algorithm would correctly remove unwanted text. A 
manual sampling of VPEs in the Brown Corpus 
showed this to be true. When the algorithm was 
implemented, however, the number of correct answers 
improved to 258, an increase of 1. In addition to 
solving the 6 cases of subdeletion, the algorithm 
350 
inlxoduced 5 errors; each of these new errors involved 
a noun phrase or prepositional phrase in the VPE that 
did not require the deletion of a counterpart in the 
antecedent. For example, one of the newly introduced 
errors occurred in the fragment "...creaking in the fog 
as it had for thirty years." The prepositional phrase 
"for thirty years" in the VPE caused the removal of 
the phrase "in the fog" from the antecedent, even 
though the phrases are not parallel in meaning. 
These results imply that the structure of a 
sentence alone is insufficient to detect subdeletion. It 
is possible, however, that a larger sample of relevant 
examples would suggest the best choice (to delete or 
not to delete) in the absence of additional information. 
Towards these ends, other corpora in the Penn 
Treebank will be examined with VPEAL. Also, newer 
versions of the Treebank include semantic tags to 
adjunct phrases which will aid in preventing the 
misidentification of subdeletion described above. 
5. Conclusion 
Improving the results of the VPEAL program is 
an iterative process. We have categorized the errors 
occurring in VPEAL. An algorithm for solving the 
error category of subdeletion was described and 
examined. Potential problem situations for the 
algorithm were also presented. Empirical evaluation 
of the algorithm indicates that a purely syntactic 
approach to detecting subdeletion is probably 
insufficient. Additional approaches to the problem of 
subdeletion were suggested. Other cases of errors will 
be likewise evaluated. 

Bibliography 
Hardt, Daniel. 1995. An empirical approach to VP 
ellipsis, 
Hobbs, Jerry. 1978. Resolving pronoun references. 
Lingua, 44:311-388. 
Lappin, Shalom and Michael McCord. 1990. 
Anaphora Resolution in Slot Grammar. 
Computational Linguistics, 16(4). 
Marcus, Mitchell P., Beatrice Santorini, and Mary 
Ann Marcinkiewicz. 1993. Building a large 
annotated corpus of english: The penn treebank. 
Computational Linguistics, 19(2). 
Walker, Marilyn. 1989. Evaluating discourse 
processing algorithms. In Proceedings, 27th 
Annual Meeting of the ACL, Vancouver, Canada. 
