AUTOMATED DETERMINATION OF SUBLANGUAGE SYNTACTIC USAGE Ralph Grbhman and Ngo Thanh Nhan Courant Institute of Mathematical Sciences New York University New York, NY 10012 Elalne Marsh Navy Center for Applied 1~se, arch in ~ Intel~ Naval ~ Laboratory Wx,~hinm~, DC 20375 Lynel~ Hirxehnum Research and Development Division System Development Corpmation / A Burroughs Company Paofi, PA 19301 Abstract Sublanguages _differ from each other, and from the "stan- dard Ian~age, in their syntactic, semantic, and discourse vrolx:rties. Understanding these differences is important'if -we are to improve our ability to process these sublanguages. We have developed a sen~.'- automatic ~ure for identifying sublangnage syntact/c usage from a sample of text in the sublanguage We describe the results of applying this procedure to taree text samples: two sets of medical documents and a set of equipment failure me~ages. Introduction b A sub~age.is th.e f.oan.of natron." ~a~ y a oommumty ot s~ts m atm~mg a resmctea domain. Sublanguages differ from each other, and tron}. the "standard language, in their syntactic, ~antic, anti discourse properties. We describe ~ some rec~.t work on (-senii-)automatically determining the.syntactic_ properties of several sublangnages. This work m part ot a larger effort aimed at improving the techniques for parsing sublanguages. If we esamine a variety of scientific and technical sublanguages, we will encounter most of the constructs of the standard language, plus a number of syntactic exten- sions. For example, report" sublantgnag ~, such as are used in medical s||mmarles and eqmpment failure sum- maries, include both full sentences and a number of ~ag- merit forms [Marsh 1983]. Specific sublanguages differ in their usage of these syntactic constructs [Kittredge 1982, Lehrberger 1982]. Identifying these differences is important in under- standing how sublanguages differ from the Language as a whole. It also has immediate practical benefits, since it allows us to trim our grammar tO fit the specific sub- language we are processing. This can significantly speed up the analysis process and bl~.k some spurious parses which wouldbe obtained with a grammar of Overly broad coverage. Determining Syntaai¢ Usage Unf .ort~natcly, a~l uirin~ the data .about ,yn~'c usage can De very te~ous, masmuca ~ st reqmres .me analysis of hundreds (or even thousands) of s~. fence., for each new sublangnage to.be proces____~i. We nave mere- fore chosen to automate this process. We are fortunate to have available to us a very broad coverage English grammar, the Linguistic.St~ing Grammar [S~gor 1981], which hp been ex~. d~ include the sentence fragn~n_ ts of certain medical aria cquilnnent failure rcixn'm [Marsh 1983]. The gram, ," consmts of a context-~r=, component a.ugmehtc~l .by pr~ural restrictions which capture v_.anous synt.t.t ~ and sublanguage _semantic cons_tt'aints. "l]~e con~- . component is stated in terms ot lgra.mmatical camgones such as noun, tensed verb, and ad~:tive. To be. gin .the analysis proceSS, a sample .mrpus is usmg this gr~,-=-,: .The me of generanm par~s_ m reviewed manually to eliminate incorrect ~. x ne remalningparses are then fed to a program which .cc~ts for each parse tree and .cumulatively for ~ entb'e me the number of times that each production m me context-free component of the grammar was applied in building the tr¢~. This yields a "trimmed" context-fr¢~ grammar for. the sublangua!~e (consLsting ~. ~osc pro- ductions usea one or more tunes), atong w~m zrequency information on the various productions. This process was initially applied to text. sampl~ from two Sublanguages. The .fi~s. t is a set o.x s~ pauent documents (including patient his.tm'y., eTam,n.ation, .and plan of treatment). The second m a set ot electrical equipment failure relxals called "CASREPs', a class of operational report used by the U. S. Navy [Froscher 1983]. The parse file for the patient documents had correct parses for 236 sentences (and sentence frag- ments); the file for the CASREPS had correct parses tor 123 sentences. We have recently applied the process, to a third text sample, drawn from a subIanguage very stmflar to the first: a set of five hospital discharge summaries , Which include patient histories, e~nmlnnt[ous, and sum- maries of the murse of treatment in the hospital. This last sample included correct parses for 310 sentences. 96 Results The trimmed grarnrtl~l~ ~du~ from thc three sublanguage text samples were of comparable size. The grammar produced from the first set of patient docu- menU; col~tained 129 non-termlnal symbols and 248 pro- ductions; the grnmmar from the second set (the "discharge summaries") Was Slightly ]~trger, with 134 non-termin~ds and 282 productions. The grammar for the CASREP sublanguage was slightly smaller, with 124 non-terminal~ and 220 productions (this is probably a reflection of the smaller size of the CASR text sam- ple). These figures compare with 255 non-termlnal sym- bols and 744 productions in the "medical records" gram- mar used by the New York University Linguistic String Pro~=t (the "medical records" grammar iS the Lingttistic String Project English Grammar with extensions for sen- tencc fragments and other, sublanguagc specific, con- structs, and with a few options deleted). Figures 1 and 2 show the cumulative growth in the size of the I~"immed grammars for the three sublanguages as a function of the number of sentences in the sample. In Ftgure 1 we plot the number of non-term/hal symbols in the grammar as a function of sample size; in Figure 2, the number of productions in the ~ as a function of sample size. Note that the curves for the two medical sublanguages (curves A and B) have pretty much fiat- tcned out toward the end, indicating that, by that point, the trimmed grnmm~tr COVe'S a V~"y lar~ fra~on of the sentences in the sublanguage. (Some of the jumps in the growth curves for the medical grAmmarS refleet the ~vi- sion of the patient documents into sections (history, pl3y- sical exam, lab tests, etc.) with different syntactic charac- teristics. For the first few documents, wl3en a new see- tion bedim, constructs are encountered which did not appear m prior sections, thus producing a jump in the c11rve.) The sublanguage gramma~ arc substantially smaller than the full English grammar, reflecting the more lim- itcd range of modifiers and complements in these sub- languages. While the full grammar has 67 options for sentence object, the sublanguage grammars have substan- tially restricted mages: each of the three sublanguage grammars has only 14 object options. Further, the gram- mars greatly overlap, so that the three grammars com- bined contain only 20 different object options. While sentential complements of nouns are available in the full grammar, there arc no i~tanc~ of such a:~[lstrllcfions in either medical sublanguage, aad only one instance in the CASREP sublanguage. The range of modifiers iS also much restricted ia the sublangu=age grammars as com- pared to the full grammar. 15 options for sentential modifiers are available in the full grammar. These are restricted to 9 in the first medical sample, 11 in the second, and 8 in the equipment failure sublangua~e. Similarly, the full English gr~mmnr has 21 options tor right modifiers of nouns; the sublanguage gr~mma_~S had fewer, 11 in the first medical sumple, I0 m" the second, and 7 in the CASREP sublanguage. Here the sub- language grammars overlap almost completely: only 12 different right modifiers of noun are represented in the three grammars combined. Among the options occurring in all the sublanguage grammars, their relative frequency varies ao~o~ding to the domain of the text. For example, the frequency of prepositional phrases as right modifiers of nouns (meas; urea as instances per sentence or sentence fragment) was 0.36 and 0.46 for the two medical samples, as compared to 0.77 for the CASREPs. More striking was the fre- quency of noun phrases with nouns as modifiers of other nouns: 0.20 and 0.32 for the two medical ~mples, versus 0.80 for the CASREPs. We reparsed some of the sentences from the first set of medical documents with the trimmed grammar and, as ~, o.bserved a considerable " speed-up. The t.mgumuc ~mng rarser uses a p.op-uown pa.~mg algo- rithm with., .ba~track~" g. A,~Ldingly , for short, simple sentences which require little backtr~.king there was only a small gain in processing speed (about 25%). For long, complex sentences, however, which require extensive backtracking, the speed-up (by roughly a factor of 3) was approximately proportional to the reduction in the number of productions. In addition, the ~fyequcncy of bad parses decreased slightly (by <3%) with the l~mmed y.mm.r (because some of the bad parses involved syntactic constructs which did not appear m any o~,,~ect parse in the sublanguage sample). Discussion As natural .lan ~,uage interfaces become more mature, their portability the ability to move an inter- face to a new domain and sublenguage is becoming increasingly important. At 8 minimllm, portability requires us to isolate the domain dependent information in a natural ]aDgua.~.e system [C~OSZ 1983, Gri~hman 1983]. A more ambitious goal m to provide a discovery procedure for this information a procedure Wl~eh can determine the domain dependent information from sam- ple texts in the sublanguage. The tcchnklUeS described above provide a partial, semi-automatic discovery pro- cedure for the syntactic usages of a sublangua~.* By applying .these .t~gues to a small sublan~ sample, we ~ adapt a broad-coverage grammar tO the syntax of a particular sublanguage. Sub~.quont text from this sub- language caa then be i~xessed more efficiently. We are currently extending this work in two direc- tions. For sentences with two or more parses which ~ atisfy .both the syntactic and the sublanguage selectional semanu.'c) constraints, we intena to try using the/re- Cency information ga~ered for productions to select, a invol "ving the more frequent syntactic constructs.** Second, we are using a s~milAr approach to develop a discovery procedure for sublanguage selectional patterns. We are collecting, from the same sublanguage samples, statistics on the frequency of co-occurrence of particular sublan .guage (semantic) classes in subjeet.vedy.ob~:ct and host-adjunct relations, and are using this data as input to * Partial, because it cannot identify new extensions to the base gramme; semi-automatic, because the parses produced with the broad-coverage grammar • must be manually reviewed. * Some small experiments of this type have been one with a Japanese ~ [Naga 0 1982] with 1|mired success. Becat~ of the v~_ differ~t na- ture of the grammar, however, it is not dear whether this lass any implications for our experi- ments. 97 the grammar's sublanguage selectional restrictions. Acknowledgemeat This material is based upon work supported by the Nalional Science Foundation under Grants No. MCS-82- 02373 and MCS-82-02397. Referenem [Frmcher 1983] Froscher, J.; Grishmau, R.; Bachenko, J.; Marsh, E. "A linguistically motivated approach to automated analysis of military messages." To appear in Proc. 1983 Conf. on Artificial Intelligence, Rochester, MI, April 1983. [Grlslnnan 1983] Gfishman, R.; ~, L.; Fried. man, C. "Isolating domain dependencies in natural language interface__. Proc. Conf. Applied Natural l~nguage Processing, 46-53, Assn. for Computational Linguistics, 1983. [Greu 1963] Grosz, B. "TEAM: a transportable natural-language interface system," Proc. Conf. Applied Natural Language Processing, 39-45, Assn. for Comlmta- fional IAnguhflm, 1983. [Kittredge 1982] Kim-edge, 11. "Variation and homo- geneity of sublauguages3 In Sublanguage: Jmdies of language in reslricted semantic domains, ed. R. Kittredge and J. Lehrberger. Berlin & New York: Walter de Gruyter; 1982. on and the concept of sublanguage. In $ublan~a&e: sl~lies of language in restricted semantic domains, ed. R. Kittredge and J. Lehrberger. Berlin & New York: Walter de Gruyter; 1982. [Marsh 1983] Marsh, E "Utilizing domain-specific information for processing compact text." Proc. Conf. ied Namra[ Lansuage Processing, 99-103, Assn. for putational Linguistics, 1983. [Nape 1982] Nagao, M.; Nakamura, J. "A parser which learns the application order of rewriting rules." Proc. COLING 82, 253-258. [Sager 1981] Sager, N. Natural Lansuage lnform~on Pro- ceasing. Reading, MA: Addlson-Wesley; 1981. 98 130 120 110 100 80 80 90 60 50 40 30 0 SENTENCES VS. NJ~N-TERMINRL SYHBBLS • ' • ' " ' ' , ' , " , • , • , • , • I • v "r 2- Y A , i . , . . . , I / , i . i , i , i , ) , i . z° ~lo 80 oo I oo 12o 14o 18o 18o zoo zzo z4o x Figure 1. Growth in thc size of the gr~mm.r as a function of the size of the text sample. X = the number of sentences (and sentence frag- ments) in the text samplc; ~" = the number of non-terminal symbols m the context-free com- ponent of thc ~'ammar. Graph A: first set of patient documents Graph B: second set of pat/cnt documcnts ("discharge s-~-,-,'ics") Graph C: e~, uipment failure messages 140 130 1:)0 110 100 gO 8O 90 30 SENTENCES VS. NON-TERMINRL 5YHBBLS f / B SO , , • , , . . l , . . . . . . , . . . , . , . , . , . , . , . 0 ZO 40 60 80 100 IZO 140 130 180 ZOO ZZO 240 Z60 ZSO 300 3ZO X 1so 12o 11o SENTENCES VS. N~N-TERMINRL SYMBOLS • e • , , l • , • l , , • , , , , , , , , J / J . / ' / , , v , lOO 80 ) 80 70 80 3o C 4O • * , , • I s I , i , : * f , i , i • * , , * , • 30 0 10 ZO 30 40 30 60 70 30 ~0 100 110 120 1~0 X 99 30O 200 ZSO SENTENCES VS. PR°IDUCTI°JNS • , . [ • , . , • . . , . , . , . , • , , . . , _/7 A J ,,, , ~, ~0 40 6 100 12Q 140 1150 180 ZOO ZZO Z~O X Figure 2. Growth in the size of the grammar as a fuaction of the size of thc text sample. X = the number of sentences (and sentence frag- ments) in the text sample; Y = the number of productions in the context-free component of the grammar. Graph A: first set of patient documents Graph B: second set of pati_e~.t documents ("discharge s.~,-,,~cs ) Graph C: e~,. ,uipment failure messages (cAs~,Ps-) 220 20O 180 2~ 220 2(30 =,- 100 180 Z 40 SENTENCES VS. PRODUCTI°'INS ", 1 , i • i • , • a , i • J , , , i , i , J . i • J . , • i , 260 240 220 200 180 16G 140 120 lOG 80 80 40 J t2Q 80 60 , * , J . i • i , i , i . i . i . , , . , i , , , B , . . . . O ZO 40 60 OO 100 120 1"i0 150 150 ZOO 220 Z~O ZSO ZSO 30O 32O X SENTENCES VS. PRgDUCTI°INS 160 140 100 O0 / C 6O ZOo 10 ZO 30 40 O0 ~0 tO0 ;10 IZO X i00 . number of productions in the context-free component of the grammar. Graph A: first set of patient documents Graph B: second set of pati_e~.t documents. number of non-terminal symbols m the context-free com- ponent of thc ~'ammar. Graph A: first set of patient documents Graph B: second set of pat/cnt