This book documents the SUSANNE corpus, [NOTE 1] a full syntactic annotation of the Gothenburg corpus (Ellegård 1978), comprising 64 of the 500 texts (approximately 130,000 words) in the Brown corpus (Francis & Kucera 1989 [1964]), 16 in each of the four Brown genre categories: press reportage, belles lettres, technical prose, and adventure fiction. It is a small corpus by contemporary standards (cf. Oostdijk 1991, Marcus et al. 1993), requiring only 5.3 megabytes of storage, but the detail of its annotation is unrivalled.
Each word and punctuation mark in the corpus is tagged with one of nearly 400 distinct 'wordtags' (105-20, 447-8) based on, but greatly extending, the 1985 version of the Lancaster tagset, with 166 members (Garside et al. 1987: 165-78). Provision is also made for the further tagging of certain words with a reference to appropriate entries in Hornby et al. 1973, more precisely to the computer-usable dictionary based on that dictionary's typesetting tape (Mitton 1986). [NOTE 2] Then, for each paragraph in the corpus (44), a tree diagram is constructed, in which each phrase, clause, and other supraword constituent is tagged with one of several hundred 'tagmatags' (168-70). [NOTE 3] Moreover, all wordtags and tagmatags may be further elaborated by various suffixes (170-1). In addition, one of 23 'functiontags' (362-3) is assigned to each of certain constituents of clauses 'to represent logical properties of the respective constituents' (354), and '[e]xtra nodes dominating no wording ("ghost nodes") are added to the parsetrees to represent the logical position of elements that have been moved or deleted in surface structure' (353). Finally, '[n]umbers, called "indices", mark the relationship between nodes marked grammatically as counterparts, such as a ghost and the corresponding full surface constituent' (354).
For example, 1 gives the full annotation of example G06:0950 (42); and 2 the supraword annotation of example N04:0560 (356). [NOTE 4] In the latter, :G is a functiontag indicating a 'guest node', a constituent of a tagma that corresponds to a ghost node in another tagma; 135 and 136 are indices; and 136 (occurring by itself) and o135 are ghost nodes.
(1) [O [S? [Rq:q [RRQq Where ] [P [II in ] [Nns [NP1g Europe ] ] ] ] [Vosb [VBDZ was ] ] [Nas:s [PPHS1m he ] ] [Vrg [VVGi going ] ] [Rw:t [RTo now ] ] ] [YQ ? ] ]
(2) [S It was [Ns:e135 a [J terrible 136 ] thing [Ti:G136 to do o135 ] ] ].
As these examples show, constituent structures in the SUSANNE corpus are relatively flat. They posit neither the intermediate phrase nodes of the X-bar Theory, nor the functional projections of the Minimalist Program, [NOTE 5]and certain apparently discontinuous constituents such as was Y going in 1 are not tagged as such. [NOTE 6] In addition, phrase nodes corresponding to certain lexical nodes are also not posited (175-6, 302-4). [NOTE 7] S's structural conservatism is based on his observation that there is little agreement about the correct assignment of constituent structures to naturally occurring sentences of English (4-5). He contends that with the exception of the treatment of coordination, the higher level structures marked in the SUSANNE corpus are relatively uncontroversial. [NOTE 8]
S's one structural innovation is to assimilate coordination to subordination: 'the second and any subsequent conjunctions of a co-ordinate structure are treated as subordinate to the first, "main" conjunct' (310-1). For example, 3 shows the annotation of the phrase N13:0490.
(3) [N the heat [Ns+ and the dust ] ]
Thus, 'within the co-ordination, the heat is not a tagma' (311), nor is the dust. Whether or not S's innovation is well motivated structurally, it is easy to recode his analysis into a more standard one, such as 4, in which both the heat and the dust are constituents of the heat and the dust.
(4) [N [Ns the heat ] [Ns+ and [Ns the dust ] ] ]
Another area of innovation in the SUSANNE tagging scheme is its treatment of names. S observes that the category of proper name is appropriately assigned at the phrase, rather than at the word, level. Accordingly, the wordtag that is assigned to an entry in Hornby et al. which occurs as (part of) a proper name is the one appropriate for that entry; for example the word Flagstaff, when occurring as a proper name, is tagged both as a singular common count noun, NN1c, and as a proper noun phrase, NnY (87). SUSANNE encoders have, in fact, exercised considerable ingenuity in the tagging of proper names, as illustrated by how they have tagged the names of the American states (146-8).
One aspect of the SUSANNE parsing scheme which 'has proved (unexpectedly, from the author's point of view) to be quite controversial' (80) is that 'a SUSANNE wordtag is a string of characters which is intended as a single atomic symbol, not as shorthand for a set of grammatical features' (79). Although 'sets of tags for words having some common grammatical property do often have a character in common[,] Y this is a matter merely of practical convenience; it does not imply that there is some specific set of binary- or multiple-valued grammatical features underlying the tagset such that any individual wordtag can be completely translated into a distribution of values over the various features' (80). As a matter of fact, however, the SUSANNE wordtag set can be translated into a set of feature structures, each of which is a bundle of features, whose values are binary- or multiple-valued features, or other feature structures. [NOTE 9] The latter is significant, since it overcomes S's major objection to the use of sets of features to encode SUSANNE wordtags, namely that a common grammatical feature may have different interpretations in different contexts; for example, that 'for a possessive pronoun to be plural (e.g. our v. my) has nothing at all to do with verb agreement' (81). Moreover the overall design of the SUSANNE tagging scheme incorporates something like feature structures, since the class of tagmatags is defined to allow certain characters, which can be understood as feature values, to be combined (see note 3); and other symbols, also interpretable as feature values, can be suffixed to both wordtags and tagmatags.
S points out that the SUSANNE corpus includes a parsing scheme, not a parsing system, since SUSANNE tagging is too complex to be done automatically (4). Moreover, S does not provide a grammar for defining the class of well-formed SUSANNE annotation structures. The latter is, perhaps, achievable, and it would be a useful exercise to attempt to come up with one, if only to discover hidden inconsistencies in the present scheme, and to provide validation for the tagging of other corpora using the scheme. As it stands, however, the SUSANNE corpus is a very useful resource for the linguistic research community, and the SUSANNE team and the funding agencies which have supported it richly deserve our thanks.
Ellegård, Alvar 1978. The syntactic structure of English texts (Gothenburg Studies in English 43). Gothenburg, Sweden: University of Gothenburg.
Francis, W. Nelson, and Henry Kucera 1989. Manual of information to accompany a standard corpus of present-day edited American English, for use with digital computers, corrected and revised edition. Providence, RI: Brown University (1st edition, 1964).
Garside, Roger, Geoffrey Leach, and Geoffrey Sampson (eds.) 1987. The computational analysis of English: A corpus-based approach. London: Longman.
Hornby, A. S., E. V. Gatenby, and H. Wakefield 1973. Oxford Advanced Learner's Dictionary, 3rd ed. Oxford: Oxford University Press.
Marcus, Mitchell P., Beatrice Santorini, and Mary Ann Marcinkiewicz 1993. Building a large annotated corpus of English: the Penn treebank. Computational Linguistics 19.313-330.
Langendoen, D. Terence and Gary F. Simons 1995. A rationale for the Text Encoding Initiative recommendations for feature-structure markup. Computers and the Humanities 29.191-205.
Mitton, Roger 1986. A partial dictionary of English in computer-usable form. Literary and Linguistic Computing 1.214-215.
Oostdijk, Nelleke 1991. Corpus linguistics and the automatic analysis of English. Amsterdam: Rodopi.
Sperberg-McQueen, C. Michael, and Lou Burnard (eds.) 1994. Guidelines for electronic text encoding and interchange (TEI P3). Chicago, IL and Oxford: Text Encoding Initiative of the Association for Computers and the Humanities, Association for Computational Linguistics, and Association for Literary and Linguistic Computing.
1. SUSANNE stands for 'Surface and underlying structural analyses of natural English'. Work on the corpus was began in 1988, and is continuing. The current version of the corpus and further documentation are available from the Oxford Text Archive.
2. The scheme for providing sense codes tied to the electronic version of Hornby et al. 1973 (67-73) is not included in current SUSANNE releases.
3. Since certain of the characters used to define the 79 listed tagmatags can be combined to create additional ones (e.g., given the tag Ns designating a noun phrase marked as singular, and Nn designating a proper name, one can form the tagmatag Nns, a proper name marked as singular), there are at least several hundred legal tagmatags.
4. Example numbering in the SUSANNE corpus corresponds to that of the tagged Brown corpus (Francis & Kucera 1989). The notation in examples 1 and 2 is not precisely that of the SUSANNE corpus file, but a shorthand that S uses for text displays. I also omit labels on right brackets.
5. S does not consider constituents based on X-bar Theory or the Minimalist Program, but does discuss the highly articulated constituent structures of Generative Semantics (352), and concludes that they are not appropriate for a syntactic encoding standard such as SUSANNE.
6. Though the separate phrase-level tagging of was and going as Vosb and Vrg respectively could be used to reconstruct them as a unit by a program which uses SUSANNE tagging as input.
7. Some of S's rules for determining whether a single word constitutes a phrase of its own are rather delicate and seemingly arbitrary. For example, a single noun is analyzed as a noun phrase if it is an immediate constituent of a clause, but not if it is an immediate constituent of a prepositional phrase (176).
8. 'The treatment of co-ordination is probably the most nonstandard aspect of the SUSANNE parsing scheme' (310).
9. See Sperberg-McQueen and Burnard 1994:515-9 for a feature-structure re-encoding of part of the wordtag set currently being used for the encoding of the British National Corpus. For objections to the use of sets of atomic tags for the encoding of grammatical structure, see Langendoen and Simons 1995.