Mitchell P. Marcus, Beatrice Santorini, Mary Ann Marcinkiewicz
Department of Computer Department of Linguistics and Information Sciences Northwestern University University of Pennsylvania Evanston, IL 60208 Philadelphia, PA 19104
There is a growing consensus that significant, rapid progress can be made in both text understanding and spoken language understanding by investigating those phenomena that occur most centrally in naturally occurring unconstrained materials and by attempting to automatically extract information about language from very large corpora. Such corpora are beginning to serve as an important research tool for investigators in natural language processing, speech recognition, and integrated spoken language systems, as well as in theoretical linguistics. Annotated corpora promise to be valuable for enterprises as diverse as the automatic construction of statistical models for the grammar of the written and the colloquial spoken language, the development of explicit formal theories of the differing grammars of writing and speech, the investigation of prosodic phenomena in speech, and the evaluation and comparison of the adequacy of parsing models.
In this paper, we review our experience with constructing one such large annotated corpus---the Penn Treebank, a corpus consisting of over 4.5 million words of American English. During the first three-year phase of the Penn Treebank Project (1989--1992), this corpus has been annotated for part-of-speech (POS) information. In addition, over half of it has been annotated for skeletal syntactic structure. These materials are available to members of the Linguistic Data Consortium; for details, see section 5.1 .
The paper is organized as follows. Section 2 discusses the POS tagging task. After outlining the considerations that informed the design of our POS tagset and presenting the tagset itself, we describe our two-stage tagging process, in which text is first assigned POS tags automatically and then corrected by human annotators. Section 3 briefly presents the results of a comparison between entirely manual and semi-automated tagging, with the latter being shown to be superior on three counts: speed, consistency, and accuracy. In section 4, we turn to the bracketing task. Just as with the tagging task, we have partially automated the bracketing task: the output of the POS tagging phase is automatically parsed and simplified to yield a skeletal syntactic representation, which is then corrected by human annotators. After presenting the set of syntactic tags that we use, we illustrate and discuss the bracketing process. In particular, we will outline various factors that affect the speed with which annotators are able to correct bracketed structures, a task which---not surprisingly---is considerably more difficult than correcting POS-tagged text. Finally, section 5 describes the composition and size of the current Treebank corpus, briefly reviews some of the research projects that have relied on it to date, and indicates the directions that the project is likely to take in the future.
2.1 A simplified POS tagset for English
The POS tagsets used to annotate large corpora in the past have traditionally been fairly extensive. The pioneering Brown Corpus distinguishes 87 simple tags ([Francis 1964]), ([Francis and Kucera 1982]) and allows the formation of compound tags; thus, the contraction I'm is tagged as PPSS+BEM (PPSS for "non-3rd person nominative personal pronoun" and BEM for "am, 'm". Subsequent projects have tended to elaborate the Brown Corpus tagset. For instance, the Lancaster-Oslo/Bergen (LOB) Corpus uses about 135 tags, the Lancaster UCREL group about 165 tags, and the London-Lund Corpus of Spoken English 197 tags. A useful overview of the relation of these and other tagsets to each other and to the Brown Corpus tagset is given in Appendix B of [Garside et al 1987]. The rationale behind developing such large, richly articulated tagsets is to approach ``the ideal of providing distinct codings for all classes of words having distinct grammatical behaviour'' ([Garside et al 1987]).
2.1.1 Recoverability
Like the tagsets just mentioned, the Penn Treebank tagset is based on that of the Brown Corpus. However, the stochastic orientation of the Penn Treebank and the resulting concern with sparse data led us to modify the Brown Corpus tagset by paring it down considerably. A key strategy in reducing the tagset was to eliminate redundancy by taking into account both lexical and syntactic information. Thus, whereas many POS tags in the Brown Corpus tagset are unique to a particular lexical item, the Penn Treebank tagset strives to eliminate such instances of lexical redundancy. For instance, the Brown Corpus distinguishes five different forms for main verbs: the base form is tagged VB, and forms with overt endings are indicated by appending D for past tense, G for present participle/gerund, N for past participle and Z for third person singular present. Exactly the same paradigm is recognized for the have, but have (regardless of whether it is used as an auxiliary or a main verb) is assigned its own base tag HV. The Brown Corpus further distinguishes three forms of do---the base form (DO), the past tense (DOD), and the third person singular present (DOZ),The gerund and the participle of do are tagged VBG and VBN in the Brown Corpus, respectively---presumably because they are never used as auxiliary verbs. and eight forms of be---the five forms distinguished for regular verbs as well as the irregular forms am (BEM), are (BER) and was (BEDZ). By contrast, since the distinctions between the forms of VB on the one hand and the forms of BE, DO and HV on the other are lexically recoverable, they are eliminated in the Penn Treebank, as shown in Table 1.The irregular present tense forms am and are are tagged as VBP in the Penn Treebank (see section ), just like any other non-third person singular present tense form.
Table 1: Elimination of lexically recoverable distinctions sing/VB be/VB do/VB have/VB sings/VBZ is/VBZ does/VBZ has/VBZ sang/VBD was/VBD did/VBD had/VBD singing/VBG being/VBG doing/VBG having/VBG sung/VBN been/VBN done/VBN had/VBNA second example of lexical recoverability concerns those words that can precede articles in noun phrases. The Brown Corpus assigns a separate tag to pre-qualifiers (quite, rather, such), pre-quantifiers (all, half, many, nary) and both. The Penn Treebank, on the other hand, assigns all of these words to a single category PDT (predeterminer). Further examples of lexically recoverable categories are the Brown Corpus categories PPL (singular reflexive pronoun) and PPLS (plural reflexive pronoun), which we collapse with PRP (personal pronoun), and the Brown Corpus category RN (nominal adverb), which we collapse with RB (adverb).
Beyond reducing lexically recoverable distinctions, we also eliminated certain POS distinctions that are recoverable with reference to syntactic structure. For instance, the Penn Treebank tagset does not distinguish subject pronouns from object pronouns even in cases where the distinction is not recoverable from the pronoun's form, as with you, since the distinction is recoverable on the basis of the pronoun's position in the parse tree in the parsed version of the corpus. Similarly, the Penn Treebank tagset conflates subordinating conjunctions with prepositions, tagging both categories as IN. The distinction between the two categories is not lost, however, since subordinating conjunctions can be recovered as those instances of IN that precede clauses, whereas prepositions are those instances of IN that precede noun phrases or prepositional phrases. We would like to emphasize that the lexical and syntactic recoverability inherent in the POS-tagged version of the Penn Treebank corpus allows end users to employ a much richer tagset than the small one described in section 2.2 if the need arises.
2.1.2 Consistency
As noted above, one reason for eliminating a POS tag such as RN (nominal adverb) is its lexical recoverability. Another important reason for doing so is consistency. For instance, in the Brown Corpus, the deictic adverbs there and now are always tagged RB (adverb), whereas their counterparts here and then are inconsistently tagged as RB (adverb) or RN (nominal adverb)---even in identical syntactic contexts, such as after a preposition. It is clear that reducing the size of the tagset reduces the chances of such tagging inconsistencies.
2.1.3 Syntactic function
A further difference between the Penn Treebank and the Brown Corpus concerns the significance accorded to syntactic context. In the Brown Corpus, words tend to be tagged independently of their syntactic function. For instance, in the phrase the one, one is always tagged as CD (cardinal number), whereas in the corresponding plural phrase the ones, ones is always tagged as NNS (plural common noun), despite the parallel function of one and ones as heads of their noun phrase. By contrast, since one of the main roles of the tagged version of the Penn Treebank corpus is to serve as the basis for a bracketed version of the corpus, we encode a word's syntactic function in its POS tag whenever possible. Thus, one is tagged as NN (singular common noun) rather than as CD (cardinal number) when it is the head of a noun phrase. Similarly, while the Brown Corpus tags both as ABX (pre-quantifier, double conjunction), regardless of whether it functions as a prenominal modifier (both the boys), a postnominal modifier (the boys both), the head of a noun phrase (both of the boys) or part of a complex coordinating conjunction (both boys and girls), the Penn Treebank tags both differently in each of these syntactic contexts---as PDT (predeterminer), RB (adverb), NNS (plural common noun) and coordinating conjunction (CC), respectively.
There is one case in which our concern with tagging by syntactic function has led us to bifurcate Brown Corpus categories rather than to collapse them: namely, in the case of the uninflected form of verbs. Whereas the Brown Corpus tags the bare form of a verb as VB regardless of whether it occurs in a tensed clause, the Penn Treebank tagset distinguishes VB (infinitive or imperative) from VBP (non-third person singular present tense).
2.1.4 Indeterminacy
A final difference between the Penn Treebank tagset and all other tagsets we are aware of concerns the issue of indeterminacy: both POS ambiguity in the text and annotator uncertainty. In many cases, POS ambiguity can be resolved with reference to the linguistic context. So, for instance, in Katherine Hepburn's witty line Grant can be outspoken---but not by anyone I know, the presence of the by-phrase forces us to consider outspoken as the past participle of a transitive derivative of speak---outspeak---rather than as the adjective outspoken. However, even given explicit criteria for assigning POS tags to potentially ambiguous words, it is not always possible to assign a unique tag to a word with confidence. Since a major concern of the Treebank is avoid requiring annotators to make arbitrary decisions, we allow words to be associated with more than one POS tag. Such multiple tagging indicates either that the word's part of speech simply cannot be decided or that the annotator is unsure which of the alternative tags is the correct one. In principle, annotators can tag a word with any number of tags, but in practice, multiple tags are restricted to a small number of recurring two-tag combinations: JJNN (adjective or noun as prenominal modifier), JJVBG (adjective or gerund/present participle), JJVBN (adjective or past participle), NNVBG (noun or gerund), and RBRP (adverb or particle).
2.2 The POS tagset
The Penn Treebank tagset is given in Table 2. It contains 36 POS tags and 12 other tags (for punctuation and currency symbols). A detailed description of the guidelines governing the use of the tagset is available in [Satorini 1990].
Table 2:
The Penn Treebank POS tagset
1. CC Coordinating conjunction 25.TO to
2. CD Cardinal number 26.UH Interjection
3. DT Determiner 27.VB Verb, base form
4. EX Existential there 28.VBD Verb, past tense
5. FW Foreign word 29.VBG Verb, gerund/present participle
6. IN Preposition/subord. 30.VBN Verb, past participle
218z conjunction
7. JJ Adjective 31.VBP Verb, non-3rd ps. sing. present
8. JJR Adjective, comparative 32.VBZ Verb, 3rd ps. sing. present
9. JJS Adjective, superlative 33.WDT wh-determiner
10.LS List item marker 34.WP wh-pronoun
11.MD Modal 35.WP Possessive wh-pronoun
12.NN Noun, singular or mass 36.WRB wh-adverb
13.NNS Noun, plural 37. # Pound sign
14.NNP Proper noun, singular 38. $ Dollar sign
15.NNPS Proper noun, plural 39. . Sentence-final punctuation
16.PDT Predeterminer 40. , Comma
17.POS Possessive ending 41. : Colon, semi-colon
18.PRP Personal pronoun 42. ( Left bracket character
19.PP Possessive pronoun 43. ) Right bracket character
20.RB Adverb 44. " Straight double quote
21.RBR Adverb, comparative 45. ` Left open single quote
22.RBS Adverb, superlative 46. " Left open double quote
23.RP Particle 47. ' Right close single quote
24.SYM Symbol 48. " Right close double quote
(mathematical or scientific)
__________________
Some symbols cannot be displayed in HTML format
2.3 The POS tagging process
The tagged version of the Penn Treebank corpus is produced in two stages,
using a combination of automatic POS assignment and manual correction.
2.3.1 Automated stage
During the early stages of the Penn Treebank project, the initial automatic POS assignment was provided by PARTS (), a stochastic algorithm developed at ATT Bell Labs. PARTS uses a modified version of the Brown Corpus tagset close to our own and assigns POS tags with an error rate of 3-5%. The output of PARTS was automatically tokenized and the tags assigned by PARTS were automatically mapped onto the Penn Treebank tagset. This mapping introduces about 4 error, since the Penn Treebank tagset makes certain distinctions that the PARTS tagset does not. A sample of the resulting tagged text, which has an error rate of 7-9%, is shown in Figure 1.
Battle-tested/NNP Japanese/NNP industrial/JJ managers/NNS here/RB always/RB buck/VB up/IN nervous/JJ newcomers/NNS with/IN the/DT tale/NN of/IN the/DT first/JJ of/IN their/PP countrymen/NNS to/TO visit/VB Mexico/NNP ,/, a/DT boatload/NN of/IN samurai/NNS warriors/NNS blown/VBN ashore/RB 375/CD years/NNS ago/RB ./.
"/" From/IN the/DT beginning/NN ,/, it/PRP took/VBD a/DT man/NN with/IN extraordinary/JJ qualities/NNS to/TO succeed/VB in/IN Mexico/NNP ,/, "/" says/VBZ Kimihide/NNP Takimura/NNP ,/, president/NN of/IN Mitsui/NNS group/NN 's/POS Kensetsu/NNP Engineering/NNP Inc./NNP unit/NN ./.
2.3.2 Manual correction stage
The result of the first, automated stage of POS tagging is given to annotators to correct. The annotators use a mouse-based package written in GNU Emacs Lisp, which is embedded within the GNU Emacs editor ([Lewis et al 1990]). The package allows annotators to correct POS assignment errors by positioning the cursor on an incorrectly tagged word and then entering the desired correct tag (or sequence of multiple tags). The annotators' input is automatically checked against the list of legal tags in Table 2 and, if valid, appended to the original word-tag pair separated by an asterisk. Appending the new tag rather than replacing the old tag allows us to easily identify recurring errors at the automatic POS assignment stage. We believe that the confusion matrices that can be extracted from this information should also prove useful in designing better automatic taggers in the future. The result of this second stage of POS tagging is shown in Figure 2. Finally, in the distribution version of the tagged corpus, any incorrect tags assigned at the first, automatic stage are removed.
Battle-tested/NNP*/JJ Japanese/NNP*/JJ industrial/JJ managers/NNS here/RB always/RB buck/VB*/VBP up/IN*/RP nervous/JJ newcomers/NNS with/IN the/DT tale/NN of/IN the/DT first/JJ of/IN their/PP countrymen/NNS to/TO visit/VB Mexico/NNP ,/, a/DT boatload/NN of/IN samurai/NNS*/FW warriors/NNS blown/VBN ashore/RB 375/CD years/NNS ago/RB ./.
"/" From/IN the/DT beginning/NN ,/, it/PRP took/VBD a/DT man/NN with/IN extraordinary/JJ qualities/NNS to/TO succeed/VB in/IN Mexico/NNP ,/, "/" says/VBZ Kimihide/NNP Takimura/NNP ,/, president/NN of/IN Mitsui/NNS*/NNP group/NN 's/POS Kensetsu/NNP Engineering/NNP Inc./NNP unit/NN ./.
To determine how to maximize the speed, inter-annotator consistency and accuracy of POS tagging, we performed an experiment at the very beginning of the project to compare two alternative modes of annotation. In the first annotation mode (``tagging''), annotators tagged unannotated text entirely by hand; in the second mode (``correcting''), they verified and corrected the output of PARTS, modified as described above. This experiment showed that manual tagging took about twice as long as correcting, with about twice the inter-annotator disagreement rate and an error rate that was about 50% higher.
Four annotators, all with graduate training in linguistics, participated in the experiment. All completed a training sequence consisting of fifteen hours of correcting, followed by six hours of tagging. The training material was selected from a variety of nonfiction genres in the Brown Corpus. All the annotators were familiar with GNU Emacs at the outset of the experiment. Eight 2,000 word samples were selected from the Brown Corpus, two each from four different genres (two fiction, two nonfiction), none of which the annotators had encountered in training. The texts for the correction task were automatically tagged as described in section 2.3. Each annotator first manually tagged four texts and then corrected four automatically tagged texts. Each annotator completed the four genres in a different permutation.
A repeated measures analysis of annotation speed with annotator identity, genre and annotation mode (tagging vs. correcting) as classification variables showed a significant annotation mode effect ( p = .05). No other effects or interactions were significant. The average speed for correcting was more than twice as fast as the average speed for tagging: 20 minutes vs. 44 minutes per 1,000 words. (Median speeds per 1,000 words were 22 vs. 42 minutes.)
A simple measure of tagging consistency is inter-annotator disagreement rate, the rate at which annotators disagree with one another over the tagging of lexical tokens, expressed as a percentage of the raw number of such disagreements over the number of words in a given text sample.
For a given text and annotators, there are C(n,2) {Chose n,2} disagreement ratios (one for each possible pair of annotators). Mean inter-annotator disagreement was 7.2% for the tagging task and 4.1% for the correcting task (with medians 7.2% and 3.6%, respectively). Upon examination, a disproportionate amount of disagreement in the correcting case was found to be caused by one text that contained many instances of a cover symbol for chemical and other formulas. In the absence of an explicit guideline for tagging this case, the annotators had made different decisions on what part of speech this cover symbol represented. When this text is excluded from consideration, mean inter-annotator disagreement for the correcting task drops to 3.5%, with the median unchanged at 3.6%.
Consistency, while desirable, tells us nothing about the validity of the annotators' corrections. We therefore compared each annotator's output not only with the output of each of the others, but also with a benchmark version of the eight texts. This benchmark version was derived from the tagged Brown Corpus by (1) mapping the original Brown Corpus tags onto the Penn Treebank tagset and (2) carefully hand-correcting the revised version in accordance with the tagging conventions in force at the time of the experiment. Accuracy was then computed as the rate of disagreement between each annotator's results and the benchmark version. The mean accuracy was 5.4% for the tagging task (median 5.7%) and 4.0% for the correcting task (median 3.4%). Excluding the same text as above gives a revised mean accuracy for the correcting task of 3.4%, with the median unchanged.
We obtained a further measure of the annotators' accuracy by comparing their error rates to the rates at which the raw output of Church's PARTS program---appropriately modified to conform to the Penn Treebank tagset---disagreed with the benchmark version. The mean disagreement rate between PARTS and the benchmark version was 9.6%, while the corrected version had a mean disagreement rate of 5.4%, as noted above. The annotators were thus reducing the error rate by about 4.2%.
4.1 Basic methodology
The methodology for bracketing the corpus is completely parallel to that for tagging---hand correction of the output of an errorful automatic process. Fidditch, a deterministic parser developed by Donald Hindle first at the University of Pennsylvania and subsequently at ATT Bell Labs ([Hindle 1983], [Hindle 1989]), is used to provide an initial parse of the material. Annotators then hand correct the parser's output using a mouse-based interface implemented in GNU Emacs Lisp. Fidditch has three properties that make it ideally suited to serve as a preprocessor to hand correction:
4.2 The syntactic tagset
Table 3 shows the set of syntactic tags and null elements that we use in our skeletal bracketing. More detailed information on the syntactic tagset and guidelines concerning its use are to be found in [Santorini & Marcinkiewicz 1991].
Table 3:
The Penn Treebank syntactic tagset
Tags
1. ADJP Adjective phrase
2. ADVP Adverb phrase
3. NP Noun phrase
4. PP Prepositional phrase
5. S Simple declarative clause
6. SBAR Clause introduced by subordinating conjunction or
0 (see below)
7. SBARQ Direct question introduced by wh-word or wh-phrase
8. SINV Declarative sentence with subject-aux inversion
9. SQ Subconstituent of SBARQ excluding wh-word or wh-phrase
10. VP Verb phrase
11. WHADVP Wh-adverb phrase
12. WHNP Wh-noun phrase
13. WHPP Wh-prepositional phrase
14. X Constituent of unknown or uncertain category
Null elements
1. * ``Understood'' subject of infinitive or imperative
2. 0 Zero variant of that in subordinate clauses
3. T Trace---marks position where moved wh-constituent
is interpreted
4. NIL Marks position where preposition is interpreted in
pied-piping contexts
Although different in detail, our tagset is similar in delicacy to that
used by the Lancaster Treebank Project, except that we allow null elements in the syntactic annotation. Because of
the need to achieve a fairly high output per hour, it was decided not to
require annotators to create distinctions beyond those provided by the
parser. Our approach to developing the syntactic tagset was highly
pragmatic and strongly influenced by the need to create a large body of
annotated material given limited human resources. Despite the skeletal
nature of the bracketing, however, it is possible to make quite delicate
distinctions when using the corpus by searching for combinations of
structures. For example, an SBAR containing the word to immediately
before the VP will necessarily be infinitival, while an SBAR containing a
verb or auxiliary with a tense feature will necessarily be tensed. To take
another example, so-called that-clauses can be identified easily by
searching for SBARs containing the word that or the null element
0 in initial position.
As can be seen from Table 3, the syntactic tagset used by the Penn Treebank includes a variety of null elements, a subset of the null elements introduced by Fidditch. While it would be expensive to insert null elements entirely by hand, it has not proved overly onerous to maintain and correct those that are automatically provided. We have chosen to retain these null elements because we believe that they can be exploited in many cases to establish a sentence's predicate-argument structure; at least one recipient of the parsed corpus has used it to bootstrap the development of lexicons for particular NLP projects and has found the presence of null elements to be a considerable aid in determining verb transitivity (Robert Ingria, personal communication). While these null elements correspond more directly to entities in some grammatical theories than in others, it is not our intention to lean toward one or another theoretical view in producing our corpus. Rather, since the representational framework for grammatical structure in the Treebank is a relatively impoverished flat context-free notation, the easiest mechanism to include information about predicate-argument structure, although indirectly, is by allowing the parse tr to contain explicit null items.
4.3 Sample bracketing output
Below, we illustrate the bracketing process for the first sentence of our sample text. Figure 3 shows the output of Fidditch (modified slightly to include our POS tags).
( (S
(NP (NBAR (ADJP (ADJ "Battle-tested/JJ")
(ADJ "industrial/JJ"))
(NPL "managers/NNS")))
(? (ADV "here/RB"))
(? (ADV "always/RB"))
(AUX (TNS *))
(VP (VPRES "buck/VBP")))
(? (PP (PREP "up/RP")
(NP (NBAR (ADJ "nervous/JJ")
(NPL "newcomers/NNS")))))
(? (PP (PREP "with/IN")
(NP (DART "the/DT")
(NBAR (N "tale/NN"))
(PP of/PREP
(NP (DART "the/DT")
(NBAR (ADJP
(ADJ "first/JJ"))))))))
(? (PP of/PREP
(NP (PROS "their/PP$")
(NBAR (NPL "countrymen/NNS"))))
(? (S (NP (PRO *))
(AUX to/TNS)
(VP (V "visit/VB")
(NP (PNP "Mexico/NNP")))))
(? (MID ",/,"))
(? (NP (IART "a/DT")
(NBAR (N "boatload/NN"))
(PP of/PREP
(NP (NBAR
(NPL "warriors/NNS"))))
(VP (VPPRT "blown/VBN")
(? (ADV "ashore/RB"))
(NP (NBAR (CARD "375/CD")
(NPL "years/NNS"))))))
(? (ADV "ago/RB"))
(? (FIN "./.")))
The original design of the Treebank called for a level of syntactic analysis comparable to the skeletal analysis used by the Lancaster Treebank, but a limited experiment was performed early in the project to investigate the feasibility of providing greater levels of structural detail. While the results were somewhat unclear, there was evidence that annotators could maintain a much faster rate of hand correction if the parser output was simplified in various ways, reducing the visual complexity of the tree representations and eliminating a range of minor decisions. The key results of this experiment were:
It is worth noting that experienced annotators can proofread previously corrected material at very high speeds. A parsed subcorpus of over one million words was recently proofread at an average speed of approx. 4,000 words per annotator per hour. At this rate of productivity, annotators are able to find and correct gross errors in parsing, but do not have time to check, for example, whether they agree with all prepositional phrase attachments.
The process that creates the skeletal representations to be corrected by the annotators simplifies and flattens the structures shown in Figure 3 by removing POS tags, non-branching lexical nodes and certain phrasal nodes, notably NBAR. The output of the first automated stage of the bracketing task is shown in Figure 4.
( (S
(NP (ADJP Battle-tested industrial)
managers)
(? here)
(? always)
(VP buck))
(? (PP up
(NP nervous newcomers)))
(? (PP with
(NP the tale
(PP of
(NP the
(ADJP first))))))
(? (PP of
(NP their countrymen)))
(? (S (NP *)
to
(VP visit
(NP Mexico))))
(? ,)
(? (NP a boatload
(PP of
(NP warriors))
(VP blown
(? ashore)
(NP 375 years))))
(? ago)
(? .))
( (S
(NP Battle-tested industrial managers
here)
always
(VP buck
up
(NP nervous newcomers)
(PP with
(NP the tale
(PP of
(NP (NP the
(ADJP first
(PP of
(NP their countrymen)))
(S (NP *)
to
(VP visit
(NP Mexico))))
,
(NP (NP a boatload
(PP of
(NP (NP warriors)
(VP-1 blown
ashore
(ADVP (NP 375 years)
ago)))))
(VP-1 *pseudo-attach*))))))))
.)
The reader may have noticed that the ADJP brackets in Figure 4 have vanished in Figure 5. For the sake of the overall efficiency of the annotation task, we leave all ADJP brackets in the simplified structure, with the annotators expected to remove many of them during annotation. The reason for this is somewhat complex, but provides a good example of the considerations that come into play in designing the details of annotation methods. The first relevant fact is that Fidditch only outputs ADJP brackets within NPs for adjective phrases containing more than one lexical item. To be consistent, the final structure must contain ADJP nodes for all adjective phrases within NPs or for none; we have chosen to delete all such nodes within NPs under normal circumstances. (This does not affect the use of the ADJP tag for predicative adjective phrases outside of NPs.) In a seemingly unrelated guideline, all coordinate structures are annotated in the Treebank; such coordinate structures are represented by Chomsky-adjunction when the two conjoined constituents bear the same label. This means that if an NP contains coordinated adjective phrases, then an ADJP tag will be used to tag that coordination even though simple ADJPs within NPs will not be bear an APJP tag. Experience has shown that annotators can delete pairs of brackets extremely quickly using the mouse-based tools, whereas creating brackets is a much slower operation. Because the coordination of adjectives is quite common, it is more efficient to leave in ADJP labels, and delete them if they are not part of a coordinate structure, than to reintroduce them if necessary.
5.1 Composition and size of corpus
Table 4 shows the output of the Penn Treebank project at the end of its first phase.
Table 4:
Penn Treebank
(as of 11/92)
Tagged for Skeletal
Description Part-of-Speech Parsing
(Tokens) (Tokens)
Dept. of Energy abstract 231,404 231,404
Dow Jones Newswire stories 3,065,776 1,061,166
Dept. of Agriculture bulletins 78,555 78,555
Library of America texts 105,652 105,652
MUC-3 messages 111,828 111,828
IBM Manual sentences 89,121 89,121
WBUR radio transcripts 11,589 11,589
ATIS sentences 19,832 19,832
Brown Corpus, retagged 1,172,041 1,172,041
Total: 4,885,798 2,881,188
All the materials listed above are available on CD-ROM to members of the
Linguistic Data Consortium .
About 3 million words of POS-tagged material and a small sampling of
skeletally parsed text are available as part of the first Association for
Computational Linguistics/Data Collection Initiative CD-ROM, and a somewhat
larger subset of materials is available on cartridge tape directly from the
Penn Treebank Project. For information, contact the first author of
this paper or send email to treebank@unagi.cis.upenn.edu.
Some comments on the materials included:
5.2 Future directions
A large number of research efforts, both at the University of Pennsylvania and elsewhere, have relied on the output of the Penn Treebank Project to date. A few examples already in print: A number of projects investigating stochastic parsing have used either the POS-tagged materials ([Magerman & Marcus 1990], [Brill et al 1990], [Brill 1991]) or the skeletally parsed corpus ([Weischedel et al 1991], [pereira & Schabes 1992]). The POS-tagged corpus has also been used to train a number of different POS taggers including [Meteer et al 1991], and the skeletally parsed corpus has been used in connection with the development of new methods to exploit intonational cues in disambiguating the parsing of spoken sentences (Veilleux & Ostendorf 1992]). The Penn Treebank has been used to bootstrap the development of lexicons for particular applications (Robert Ingria, personal communication) and is being used as a source of examples for linguistic theory and psychological modelling (e.g. [Niv 1991]). To aid in the search for specific examples of grammatical phenomena using the Treebank, Richard Pito has developed tgrep, a tool for very fast context-free pattern matching against the skeletally parsed corpus, which is available through the Linguistic Data Consortium.
While the Treebank is being widely used, the annotation scheme employed has a variety of limitations. Many otherwise clear argument/adjunct relations in the corpus are not indicated due to the current Treebank's essentially context-free representation. For example, there is at present no satisfactory representation for sentences in which complement noun phrases or clauses occur after a sentential level adverb. Either the adverb is trapped within the VP, so that the complement can occur within the VP where it belongs, or else the adverb is attached to the S, closing off the VP and forcing the complement to attach to the S. This "trapping" problem serves as a limitation for groups that currently use Treebank material to semiautomatically derive lexicons for particular applications. For most of these problems, however, solutions are possible on the basis of mechanisms already used by the Treebank Project. For example, the pseudo-attachment notation can be extended to indicate a variety of crossing dependencies. We have recently begun to use this mechanism to represent various kinds of dislocations, and the Treebank annotators themselves have developed a detailed proposal to extend pseudo-attachment to a wide range of similar phenomena.
A variety of inconsistencies in the annotation scheme used within the Treebank have also become apparent with time. The annotation schemes for some syntactic categories should be unified to allow a consistent approach to determining predicate-argument structure. To take a very simple example, sentential adverbs attach under VP when they occur between auxiliaries and predicative ADJPs, but attach under S when they occur between auxiliaries and VPs. These structures need to be regularized.
As the current Treebank has been exploited by a variety of users, a significant number have expressed a need for forms of annotation richer than provided by the project's first phase. Some users would like a less skeletal form of annotation of surface grammatical structure, expanding the essentially context-free analysis of the current Penn Treebank to indicate a wide variety of non-contiguous structures and dependencies. A wide range of Treebank users now strongly desire a level of annotation which makes explicit some form of predicate-argument structure. The desired level of representation would make explicit the logical subject and logical object of the verb, and would indicate, at least in clear cases, which subconstituents serve as arguments of the underlying predicates and which serve as modifiers.
During the next phase of the Treebank project, we expect to provide both a richer analysis of the existing corpus and to provide a parallel corpus of predicate-argument structures. This will be done by first enriching the annotation of the current corpus, and then automatically extracting predicate-argument structure, at the level of distinguishing logical subjects and objects, and distinguishing arguments from adjuncts for clear cases. Enrichment will be achieved by automatically transforming the current Penn Treebank into a level of structure close to the intended target, and then completing the conversion by hand.
Brill, Eric, 1991.
Discovering the lexical features of a language.
In Proceedings of the 29th Annual Meeting of the Association for
Computational Linguistics, Berkeley, CA.
Brill, Eric; Magerman, David; Marcus, Mitchell P.; and Santorini, Beatrice,
1990.
Deducing linguistic structure from the statistics of large corpora.
In Proceedings of the DARPA Speech and Natural Language
Workshop, June 1990, pages 275--282.
Church, Kenneth W., 1980.
Memory limitations in natural language processing, MIT LCS
Technical Report 245.
Master's thesis, Massachusetts Institute of Technology.
Church, Kenneth W., 1988.
A stochastic parts program and noun phrase parser for unrestricted
text.
In Proceedings of the Second Conference on Applied Natural
Language Processing. 26th Annual Meeting of the Association for Computational
Linguistics, pages 136--143.
Francis, W. Nelson, 1964.
A standard sample of present-day English for use with digital
computers. Report to the U.S Office of Education on Cooperative
Research Project No. E--007.
Brown University, Providence.
Francis, W. Nelson and Kucera, Henry, 1982.
Frequency analysis of English usage. Lexicon and grammar.
Houghton Mifflin, Boston.
Garside, Roger; Leech, Geoffrey; and Sampson, Geoffrey, 1987.
The computational analysis of English. A corpus-based approach.
Longman, London.
Hindle, Donald, 1983.
User manual for Fidditch.
Technical memorandum 7590--142, Naval Research Laboratory.
Hindle, Donald, 1989.
Acquiring disambiguation rules from text.
In Proceedings of the 27th Annual Meeting of the Association for
Computational Linguistics.
Lewis, Bil; LaLiberte, Dan; and the GNU Manual Group, 1990.
The GNU Emacs Lisp reference manual.
Free Software Foundation, Cambridge.
Magerman, David and Marcus, Mitchell P., 1990.
Parsing a natural language using mutual information statistics.
In Proceedings of AAAI-90.
Meteer, Marie; Schwartz, Richard; and Weischedel, Ralph, 1991.
Studies in part of speech labelling.
In Proceedings of the Fourth DARPA Speech and Natural Language
Workshop, February 1991, Draft version.
Niv, Michael, 1991.
Syntactic disambiguation.
In The Penn Review of Linguistics 14, pages 120--126.
Pereira, Fernando and Schabes, Yves, 1992.
Inside-outside reestimation from partially bracketed corpora.
In Proceedings of the 30th Annual Meeting of the Association for
Computational Linguistics.
Santorini, Beatrice, 1990.
Part-of-speech tagging guidelines for the Penn Treebank
Project.
Technical report MS-CIS-90-47, Department of Computer and
Information Science, University of Pennsylvania.
Santorini, Beatrice and Marcinkiewicz, Mary Ann, 1991.
Bracketing guidelines for the Penn Treebank Project.
Ms., Department of Computer and Information Science,
University of Pennsylvania.
Veilleux, N. M. and Ostendorf, Mari, 1992.
Probabilistic parse scoring based on prosodic features.
In Proceedings of the Fifth DARPA Speech and Natural
Language Workshop, February 1992.
Weischedel, Ralph; Ayuso, Damaris; Bobrow, R.; Boisen, Sean; Ingria, Robert;
and Palmucci, Jeff, 1991.
Partial parsing: a report of work in progress.
In Proceedings of the Fourth DARPA Speech and Natural
Language Workshop, February 1991, Draft version.