Clause Aggregation Using Linguistic Knowledge 
James Shaw 
Dept. of Computer Science 
Columbia University 
New York, NY 10027, USA 
shaw@cs.columbia.edu 
Abstract 
By combining multiple clauses into one single sentence, a text generation system can express 
the same amount of information in fewer words and at the same time, produce a great variety 
of complex constructions. In this paper, we describe hypotactic and paratactic operators for 
generating complex sentences from clause-sized semantic representations. These two types of 
operators are portable and reusable because they are based on general resources such as the 
lexicon and the grammar. 
1 Introduction 
An expression is more concise than another expression if it conveys the same amount of informa- 
tion in fewer words. Complex sentences generated by combining clauses are more concise than 
corresponding simple sentences because multiple references to the recurring entities are removed. 
For example, clauses like "Jones is a patient" and "Jones has hypertension" can be combined into 
a more concise sentence "Jones is a hypertensive patient. '~ To illustrate the common occurrence 
of such repeated entities in generation, let us take a shipping company's database as an example. 
Each database tuple being conveyed is transformed into one or multiple propositions or clauses 
(we use these terms interchangeably throughout the paper). Each proposition refers to a piece of 
information which usually corresponds to a simple sentence. The database might Contain multiple 
shipments to the same location possibly on the same day. Generating a sentence for each tuple sep- 
arately would containrepetitive and potentially redundant references to the same location Or date. 
Though we used a relational database as the example, the observation about recurring entities in 
the input is also valid for other types of input, such as execution traces from expert systems. 
CASPER (Clause Aggregation in Sentence PlannER) is a sentence planner which focuses on 
generating concise sentences. Clause aggregation can happen at three levels: inferential, rhetori- 
cal, and linguistic. At the inferential level, user modeling, domain knowledge, and common sense 
reasoning are used to reduce the number of concepts to convey. Such operations are implemented 
in the content planner and clauses are combined without consulting lexical resources. Text sum- 
marization is an application which uses inferential operators extensively. For example, the two 
sentences "John hit Mary" and "Mary kicked John" might imply that "John and Mary fought." 
To define a set of inferential operators for unrestricted text is beyond the state-of-art. Because 
it is unlikely that the inferential operators for our domains (medical briefings and telephone net- 
work plan descriptions ) will be reusable for other applications, we have directed our effort into 
aggregation operations at other levels. At the rhetorical level, clauses are combined based on their 
rhetorical relationships \[Mann and Thompson, 1986\], such as CONTRAST and CONDITION. We 
will take advantage of such information in future aggregation work. At the linguistic level, lexical 
and Syntactic information are used to combine clauses. In this paper, we concentrate on two types 
138 
I 
I 
I 
I 
! 
I 
I 
!1 
ii 
il 
I 
!1 
I 
! 
!1 
il 
I 
I 
The patient's past medical history is significant for bladder carcinoma1, status post cystectomy with 
a urostomy tube insertion2, left nephrolithiasis~, status post surgery4, recurrent syncopes, questionable 
vagovagal6, a neurological workup was negativer, and the EPS was negatives, abdominal aortic aneurysm 
approximately 5 cmg, high cholesterol10, exertional anginan, past tobacco smoker, quit about one year ago12. 
Figure 1: The sentence with maximum number of propositions in the corpus 
Of linguistic aggregation operators: hypotactic and paratactic. The term, hypotaxis, describes the 
relation between a dependent element and its dominant element. Hypotactic operators transform 
one clause into a modifier and attach the modifier to the dominant clause. In contrast, parat- 
actic aggregation operators combine clauses together using constructions of equal status, such as. 
coordination. 
CASPER is used in two separate projects, MAGIC (Multimedia Abstract Generation for 
Intensive Care) and PLANDoc, to increase the fluency of the generated text. MAGIC 
\[Dalal et al., 1996, MeKeown et al., 1997\] automatically generates multimedia briefings to describe 
the post-operative status of a patient after undergoing Coronary Artery Bypass Graft (CABG) 
surgery. It uses the existing computerized information infrastructure in the operating rooms at 
Columbia Presbyterian Medical Center. PLANDoc\[Kukich et al, 1994, McKeown et al., 1994\] 
generates English summaries based on somewhat cryptic traces of the interaction between planning 
engineers and LEIS-PLAN TM. It documents the timing, placement and cost of new facilities for 
routes in telephone networks. 
In Section 2, we present a corpus analysis to identify the complexity of the target output in 
MAGIC. Section 3 describes the semantic representation used in CASPER. Details of hypotactic 
operators are presented in Section 4. Paratactic operators are described in Section 5. Section 6 
describes related work. 
2 Corpus Analysis 
We conducted a corpus analysis to study various styles and types of aggregation. The corpus 
consists of the first few sentences in the discharge summaries for 54 patients in the medical domain. 
These sentences describe patients' demographics and medical conditions pertinent to patient care 
in the Intensive Care Unit. In our study, the first step was to find out how many propositions were 
combined in each sentence. A proposition is defined as a piece of information that the physician 
(the speaker) might choose to convey in a stand-alone sentence tothe nurses in the Intensive Care 
Unit (the hearer). For example; a sentence "The patient is a 40 year old female admitted for heart 
surgery:' contains three propositions: "The patient is a female.", "The patient is 40 years old.", 
and "The patient was admitted for heart surgery." 
The small corpus contained 121 sentences with 2262 words. From the 121 sentences, we obtained 
418 propositions after manual decomposition, with a maximum 12 propositions in a single sentence 
as shown in Figure 1. On average, there are 3.5 propositions per sentence. Out of 54 summary 
sentences (the first sentence in each discharge summary) for each patient, doctors prefer to use 
prepositional phrases (PPs) ('%vith aortic stenosis") rather than relative clauses ("who likely has 
endocarditis...';) to insert medical conditions into a sentence (35 occurrences vs. 4). In only two 
cases, both PPs and relative clauses were used; all others have neither. Our studies revealed the 
following: 
• Physicians produce very complex sentences. 
• Coordinate constructions are the most popular aggregation operations, followed by PPs, and 
then adjectives. Present and past participle clauses are less common; relative clauses are rare. 
• These aggregation operations result in long distance dependencies and non-constituent coor- 
dinations (conjoin'ing constituents with different syntactic types): 
139 
The analysis also indicates that people prefer using linguistic devices that are simpler (e.g., words 
over phrases over clauses) \[Scott and de Souza, 1990, Hovy, 1993\]. 
We encountered sentences from the corpus which could be formulated more concisely. The 
doctors did very little editing to the discharge summaries. In this respect, the summaries are 
somewhat similar to speech. As a result, doctors prefer to use more flexible linguistic constructions, 
such as PPs, instead of producing the most concise sentences. Concepts such as "hypertension" 
and "diabetes" have both noun and adjective forms. Even though the noun form is longer (it 
is always used together with other words as in "patient with hypertension", or "patient who has 
hypertension"), the shorter adjective form ("hypertensive patient") did not appear in the corpus. 
In only one case, an adjective "obese" is used instead of the PP "with obesity" to indicate medical 
conditions. Since many medical conditions have no adjective forms, such as "peptic ulcers", the 
speaker is more likely to use noun forms to group together all medical conditions. In addition, 
more information can be attached to nouns but not adjectives. In the noun form, the medical 
condition "diabetes" might be modified in the corpus, as in "type 1 diabetes with extensive end 
organ damage" and "borderline diabetes": Such flexibility with nouns explains the popularity of 
its usage over adjectives. 
In summary, our analysis shows that a high level of aggregation is typical in the domain. Judging 
from the number of the PPs in comparison to relative clauses used, clause aggregation using simpler 
syntactic constituents is preferred. DoCtors generate summaries in real-time without examining all 
the information right in front of them. As a result, they might not generate the most concise 
sentences. MAGIC, on the other hand, generates text off-line, with all the conveying information 
available. This would allow MAGIC to generate more concise text by taking advantage of linguistic 
opportunities. 
3 Semantic Representation 
CASPER uses a representation influenced by Lexical-Functional Grammar (LFG) 
\[Kaplan and Bresnan, 1982\] and Semantic Structures \[Jackendoff, 1990\]. An example of the se- 
mantic representation is provided inFigure 2. In our representation, the roles for each event or 
state are PRED, ARG1, ARG2, ARG3, and MOD. The slot PRED stores the verb concept. Depend- 
ing on the concept in PRED, ARG1, ARG2, and ARG3 can take on different thematic roles, such 
as Actor, Goal, and Beneficiary, respectively, as in "John gave a red book to Mary yesterday." 
The optional slot MOD stores modifiers of the PRED. It can have one or multiple circumstantial 
elements, including MANNER, PLACE, or TIME. Inside each argument slot, it too has a MOD slot 
to store information such as adjectives or PPs. 
4 Hypotactic Operators 
We will use an example from MAGIC to demonstrate how hypotactic operators work. The surface 
forms of the propositions from the content planner are shown in Figure 3. In addition to the 
propositions, the content planner also indicates that the focus of the discourse is "the patient", 
with an entity-id, ID1. CASPER picks the first proposition, la, as the dominant proposition because 
it contains the focus entity, and it has C-NAME entity. Since, the entity in focus should appear 
as early as possible to provide a context, the proposition la is transformed from "The patient 
has name - Jones" into the semantic representation for "Jones is a patient". The PRED of the 
proposition is changed from C-IttS-ATTRIBtrrE to C-IS-INSTANCE, in addition to swapping of ARG1 
and ARG2. Each proposition is represented similarly to the one shown in Figure 2. We use the 
concept C-HAS-ATTRIBUTE to denote that the entity in ARG1 has the attribute stored in ARG2. 
Depending on the lexical properties of the attribute in ARG2, the proposition le in Figure 3, can 
be realized as "the patient has diabetesnou~" or "the patient is diabeticaaj". 
140 
I 
I 
((pred ((pred c-has-attribute) (type EVENT) (tense present))) 
(argl ((pred c-doctor) (type THING) 
(mod ((pred c-patient) (type THING) 
(modify-type POSSESSOR) (entity-id IDI))))) 
(arg2 ((pred c-name) (type THING) 
(last-name "Smith")))) 
Figure 2: Semantic representation for lf: 'q'he patient's doctor is Smith." 
la. The 
lb. The 
lc. The 
Id. The 
le. The 
If. The 
Ig. The 
patient has name - Jones. 
patient has gender - female. 
patient has age - 80 year. 
patient has hypertension. 
patient has diabetes. 
patient's doctor has name - Smith, 
patient is undergoing CABG. 
Figure 3: input propositions for "Ms. Jones is an 80 year old hypertensive diabetic female patient of Doctor 
Smith undergoing CABG.'" 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
I 
To aggregate two propositions using hypotactic operators, the propositionsmust share some 
entities in common. When they do, hypotactic operators try to transform one of the clauses into 
a modifier. Since the goal is to generate concise text, CASPER prefers transforming a proposition 
into an adjective if possible, then a PP, a participle clause, and if il else fails, a relative clause. 
This preference of syntactically simple expressions over more complex ones was also proposed in 
\[Scott and de Souza, 1990\]. In the future, we plan to incorporate constraints from the corpus to 
determine which aggregation operators to apply and in what order. 
To transform a proposition into an adjective, a propositions must satisfy the following two 
preconditions. First, the slot PLIED of the proposition being transformed must be C-HAS-ATTRIBUTE 
(the patient has age - 80 years). The other requirement is that the ARG2 of the proposition (age 
80 years) can be mapped to an adjective, as permitted in the lexicon. Using the algorithm, 
propositions lb, lc, ld, le can all be transformed into adjectives and attached to proposition la 
resulting in "Jones is an 80 year old hypertensive diabetic female patient." There are two interesting 
things to note here. First, because of the PRED of the dominant proposition is C-IS-INSTANCE, the 
transformed modifiers (age, gender, etc) are attached to the ARG2 slot of the dominant proposition 
('% patient") instead of ARG1 ("Jones"). Second, the sequential order of the modifiers is not 
determined yet at this stage. The goal of CASPER is to produce a concise semantic representation 
for a set of propositions and to guarantee that there is at least one way to express the result in the 
later generation modules. To guarantee expressibility \[Meteer, 1991\], CASPER looks ahead into the 
lexicon, but it does not make detailed lexical decisions for efficiency reasons. The exact lexical and 
syntactic decisions, including the ordering between modifiers, are made later in the lexical chooser. 
Consider another proposition: "the patient has peptic ulcers". This proposition cannot be 
transformed into an adjective because there is no adjective form for C-PEPTIC-ULCER in the lexicon. 
A proposition can be transformed into a PP with a general preposition '%vith" if the PRED of the 
proposition is C-HAS-ATTRIBUTE and the concept in its ARG2 can be mapped into a noun phrase. If 
we apply the PP operator to the proposition, we would have "Jones is an 80 year old hypertensive 
diabetic female patient with peptic ulcers." CASPER currently uses an ontology which can identify 
that C-PEPTIC-ULCER, C-HYPERTENSION, and C=DIABETES are all medical disorders and group them 
together for cohesion. Since all these medical conditions can be mapped to nouns but not to 
adjectives, they will all be realized as PPs: "Jones is an 80 year old female patient with hypertension, 
diabetes and peptic ulcers\]' 
141 
I 
I 
((pred ((pred c-install) (type EVENT) (tense past))) 
(argl ((pred c-name) (TYPE THING) 
(first-name "Alice") )) 
(arg2 ((pred c-MS-0ffice) (type THING))) 
(rood "(((pred "on") (type TIME) 
(argl ((pred "Monday") (type TIME-THING)))) 
((pred "for") (type BENEFICIARY) 
(argl •((pred c-name) (type THING) 
(first-name "John"))))))) 
Figure 4: The attribute-value pair representation for "Alice installed MS Office for John on Monday. I' ~(). 
= a list. 
I 
I 
I 
I In If in Figure 3, "The patient's doctor has name - Smith", is transformed into a PP ("of 
Smith"). The POSSESSOR modifier in ARC1, as shown in Figure 2, can be transformed into 
a PP using of-genitive\[Quirk et al., 1985\]. This phenomenon holds for relationships similar to 
patient/doctor, such as advisor/advisee, and boss/employee. 
All propositions can be transformed into a relative clause of another as long as they: share a 
common entity. In the example, proposition lg does not satisfy the precondition s of the previous 
hypotactic operators. In this case, it is combined as a present participle clause because present 
participle clause is simpler and shorter. The result of the hypotactic operators is a semantic 
representation for "Jones is an 80 year old hypertensive diabetic female patient of Smith undergoing 
CABG." 
Similar to parsing long sentences, efficiency is an important issue in generating long and complex 
sentences. Search space grows exponentially in respect to the length in both cases. CASPER is able 
to generate complex sentences efficiently because it delays the difficult detailed lexical decisions until 
absolutely needed. At the sentence planning level, CASPER looks ahead into the lexicon and merges 
those propositions that satisfy the required lexical constraints. This prevents the lexical chooser 
from •trying to combine incompatible clauses later. By determining sentence boundaries before 
carrying out detailed lexical decisions, CASPER cuts down the search space of the lexical chooser 
drastically. In STREAK \[Robin, 1995\], a generation system which also implements hypotactic 
aggregation, detailed lexical decisions are made whenever a proposition is aggregated. This is 
costly because the best lexical decisions• for n propositions might not be useful or correct for 
n + ! propositions. The strategy generates impressive complex sentences, but for some complex 
sentences, STREAK took more than half an hour. Since •CASPER does not use detailed lexical 
information when it makes sentence boundary determination, it traded some optimal aggregation 
for efficiency. Even though the lexicon is accessed twice in our system, CASPER prunes the search 
space drastically by delaying expensive detailed lexical decisions after it knows• about how many 
concepts are involved in the desired sentence. Efficiency issues in generation were also addressed 
in \[McDonald et al., 1987, Elhadad et al., •1997\]. 
5 Paratactic Operators 
We will use an imaginary human resource report system for a technical support team as an example 
to illustrate our paratactic algorithm. The example shown in Figure 4 has the following slots: PRED, 
ARC1, ARC2, MOD-BENEFICIARY,. MOD-TIME. We Currently have two approaches to combine 
propositions using coordinate constructions. In the first approach, adjacent propositions that have 
only 1 slot containing distinct elements are collapsed into one proposition with one conjoined slot 
containing the distinct elements. For example, the following sentence is the result of collapsing 
• two propositions with distinct elements in their MOD-BENEFICIARY slot: "Alice installed Quicken 
for Mary and Peter on Tuesday." \[McCawley, 198!\] described this •phenomenon as •Conjunction 
142 
il 
I 
! 
I 
I 
I 
I 
I 
I 
I Alice installed Excel for John on 
Bob removed WordPerfect for John on 
Alice installed Powerpoint for John on 
Cindy removed Access for John on 
Monday. 
Tuezday. 
Monday. 
Monday. 
Figure 5: A sample of input propositions in surface form. 
Alice insta~.le~_Excel for John on 
Alice installed Powerpoint for John on 
Cindy removed Access for John on 
Bob removed WordPerfect for John on 
Monday. 
Monday. 
Monday. 
Tuesday. 
Figure 6: The propositions in surface form after Stage 1. 
Reduction. In the second approach, the conjoined propositions have distinct elements in more than 
I 
I 
I 
I 
I 
i 
i 
I 
I 
i 
I 
I 
one slot. To combine them, each conjoined proposition is generated, but deletion rules (described 
later in Section 5.4) are used to ensure the resulting sentence has the correct ellipsis. In the following 
sentence, the two propositions are distinct at both PRED and ARG2: "John finished his work and 
\[John\] went home. ''1 The ARG1 in second proposition "John" is deleted. 
Due to limited space, we only describe the algorithm used in CASPER to produce sentences 
with coordinations. For a more detailed discussion with relevant linguistic motivations, please see 
\[Shaw, 1998\]. We have divided the algorithm into four stages~ where the first three stages take 
place in the sentence planner and the last stage takes place in the lexical chooser: 
Stage 1: group propositions and order them according to their similarities while 
satisfying pragmatic and contextual constraints. 
Stage 2: determine recurring elements in the ordered propositions that will be 
combined. 
Stage 3: create a sentence boundary when the combined clause reaches pre- 
determined thresholds. 
Stage 4: determine which recurring elements are redundant and should be deleted. 
We will go into detail of each Stage in the following 4 sections. 
5.1 Group and Order Propositions 
Coordination allows the deletion of recurring entities at the surface level, but only if they are 
adjacent; that is, the propositions containing the entities are sequentially next to each other. As a 
result, the sequential order of the propositions being coordinated affects the length of the output 
text. In Step 1, CASPER sequentializes the propositions to allow the maximum number of adjacent 
recurring entities to produce concise text. 
For the proposition in Figure 5, the semantic representations have the following slots: PRED, 
ARG1, ARG2, MOD-BENEFICIARY, and MOD-TIME. To identify which slot has the most similarity 
among its elements, we calculate the number of distinct elements (NDE) in each slot across the 
propositions. For the purpose of generating concise text, CASPER prefers to group propositions 
which result in as many slots with NDE = 1 as possible. For the propositions in Figure 5, the NDE 
of MOD-BENEFICIARY is 1 because all the beneficiaries are "John"; the NDEs for both PRED 
and MOD-TIME are2 because there are two actions, "install" and "remove", which occurred on 
either "Monday" or "Tuesday"; the NDE for ARG2 is 4 because it contains "Excel", "WordPerfect", 
"Powerpoint", and "Access"; similarly, the NDE of ARG1, the agent, is 3. 
1The string enclosed in symbols \[ and \] are deleted from the surface expression, but these concepts exist in the semantic representation. 
143 
I 
I 
-((pred c-and) (type 
(elts 
-(((pred ((pred 
(argl ((pred 
(arg2 ((pred 
(mod ((pred 
(argl 
((pred ((pred 
(argl ((pred 
(arg2 ((pred 
(mod ((pred (argl 
LIST) 
"installed") (type EVENT) (status RECURRING))) 
"Alice") (TYPE THING) (status RECURRING))) 
"Excel") (type THING))) 
"on") (type TIME) 
((pred "Monday") (type TIME-THING)))))) 
"installed") (type EVENT) (status RECURRING))) 
"Alice") (TYPE THING) (status RECURRING))) 
"Outlook") (type THING))) 
"on") (type TIME) 
( (pred "Friday" ) (type TIME-THING) ) ) ) ) ) ) ) ) 
Figure 7: The simplified representation for "Alice installed Excel on Monday and Outlook on Friday." 
I 
I 
I 
I 
The algorithm re-orders the propositions by sorting the elements in each slots using compar- 
ison operators which can determine that Monday is Smaller than Tuesday, or "Alice" issmaller 
than "Bob" alphabetically. Starting from the slots with highest NDE to the lowest, the algorithm 
re-orders the propositions based on the elements of each particular slot. In this case, proposi- 
tions will ordered according to their ARG2 first, followed by ARG1, MOD-TIME, PRED, and MOD- 
BENEFICIARY. The sorting process will put similar propositions adjacent to each other as Shown 
in Figur e 6. 
5.2 Identify Recurring Elements 
The current algorithm tries to combine only two propositions at once. In Stage 2, CASPER is 
concerned with how many slots have distinct values and • which slots they are. When multiple 
• adjacent propositions have only one slot with distinct elements, these propositions are 1-distinct. 
Propositions that are 1-distinct can be replaced with one proposition with one slot conjoining the 
distinct elements of that slot. In our example, the first and second propositions are 1-distinct at 
ARG2, and they are combined into a semantic structure representing "Alice installed Excel and 
Powerpoint for John on Monday." 
When propositions have more than one distinct slot or their 1-distinct slot is different from 
the previous 1-distinct slot, the two propositions are said to be multiple-distinct. Our approach 
in combining multiple-distinct propositions is different from previous linguistic analysis. Instead 
of removing recurring entities immediately based on transformation or substitution, the current 
system generates every conjoined multiple-distinct proposition. During the lexicalization of the 
conjoined sentence, the lexical chooser prevents the realization component from generating any 
string for the redundant elements. Our multiple-distinct coordination produces what linguists 
describe as •ellipsis and gapping. Figure 7 shows the result combining two propositions that will 
result in !'Alice installed Excel on Monday and Outlook on Friday." Some readers might notice 
that PRED and ARG1 in both propositions are marked as RECURRING. The process to delete only 
subsequent recurring elements at surface level will be explained in Section 5.4. 
• 5.3 Determine Sentence Boundary 
Unless combining the next proposition into the result proposition will exceed the pre-determined 
parameters for the complexity of a sentence, the algorithm will keep on combining more propositions 
into the result proposition using 1-distinct or multiple-distinct coordination. Based on looking at 
PLANDoc output, we limit the number of propositions conjoined by multiple-distinct coordination 
to two in normal cases. Higher threshold renders some of the sentences difficult to comprehend. 
In special cases where the same slots across nmltiple propositions are multiple-distinct, the pre- 
'-I 
II 
II 
II 
II 
II 
144 
I 
I 
! 
I 
i 
i 
I 
I 
I! 
I 
II 
II 
! 
I 
I 
! 
II 
! 
determined limit is ignored. By taking advantage of parallel structures, these propositions can be 
combined using multiple-distinct procedures without making the coordinate structure more difficult 
to understand. For example, the sentence "John took aspirin on Monday, penicillin on Tuesday, 
and Tylenol on Wednesday." is long but quite understandable. Similarly, conjoining a long list 
of 3-distinct propositions produces understandable sentences too: "John played tennis on Monday, 
drove to school on Tuesday, and won the lottery on Wednesday." These constraints allow CASP~,R 
to produce easily understandable complex sentences containing a lot of information. 
5.4 Delete Redundant Elements 
Stage 4 handles ellipsis. In the previous stages, adjacent elements that occur more than once among 
the propositions are marked as RECURRING, but the actual deletion decisions have not been made 
because CASPER lacks the necessary information. T15e importance of the surface sequential order 
can be demonstrated by the following example. In the sentence "On Monday, Alice installed Excel 
and \[on Monday,\] \]Alice\] removed Lotus 123.", the elements in MOD-TIME delete forward (i.e. 
the subsequent occurrence of the identical constituent disappears). When MOD-TIME elements 
are realized at the end of the clause, the same elements in MOD-TIME delete backward (i.e. the 
antecedent occurrence of the identical constituent disappears): "Alice installed Excel \[on Monday,\] 
and \[Alice\] removed Lotus 123 on Monday." In general, if a slot is realized at the front or medial 
of a clause, the recurring elements in that slot delete forward. In the first example, MOD-TIME 
is realized as the front adverbial while ARG1, "Alice", appears in the middle of the clause, so 
elements in both slots delete forward. On the other hand, if a slot is realized at the end position of 
a clause: the recurring elements in such slot delete backward, as the MOD-TIME in second exanlple. 
Our extended directionality constraint, an extension of \[Tai, 1969\]'s Directionality Constraint, also 
applies to conjoined premodifiers and postmodifiers as well, as demonstrated by ':in Aisle 3 and \[in 
Aisle\] 4;', and "at 3 \[PM\] and \[at\] 9 PM". 
Using the algorithm just described, the result is concise and easily understandable: "On Monday: 
Alice installed Excel and Powerpoint and Cindy removed Word for John. Bob removed WordPerfect 
for John on Tuesday." Further discourse processing can replace the beneficiary"John" in the second 
sentence with a pronoun "him". 
6 Related Work 
Both hypotactic and paratactic constructions described in this paper have received a lot of attention • 
in linguistics \[Quirk et al., 1985, Halliday, 19941 Carpenter, 1998\]. Much generation literature on 
aggregation was disguised under the topic "revision" \[Meteer, 1991, Robin, 1995\] 
\[Callaway and Lester, 1997\]. We consider clause aggregation as an integral part of a text gen- 
eration system, not as a revision. The term "revision" implies that something has been generated 
and then improved upon, which is not the case in these systems. We prefer the term optimization 
used by \[Dale, 1992\], which describes the phenomenon of aggregation more appropriately - it use 
fewer words to convey the same amount of information. 
In earlier systems, clause aggregations are implemented in strategic component 
\[Mann and Moore, 1980, Dale, 1992, Horacek, 1992\]. Logical derivations were used to combine 
clauses and remove easily inferable clauses in \[Mann and Moore, 1980\]. In such systems, ag- 
gregation decisions are made without lexical information. Newer systems, such as \[Shaw, 1995, 
Wanner and Hovy, 1996, Huang and Fiedler, 1997\]: use a sentence planner to make decisions at 
clause level between the strategic and tactical component. 
With the exception of \[Scott and de Souza, 1990\] and \[Robin, 1995\], most research in aggrega- 
tion did not transform clauses into modifiers, such as adjectives, PP, or relative clauses, in a sys- 
tematic manner. \[Scott. and de Souza, 1990\] proposed heuristics for carrying out clause combining 
based on RST and specifically identified which rhetorical relations are appropriate for "embedding" : 
145 
which corresponds to our hypotactic operators. We will incorporate rhetorical aggregation in the 
future. Robin's work on revision operators •is similar to ours. We have describe his work earlier in 
Section 4. 
Because sentences with coordination constructions can express a lot of information with few 
words, many text generation systems have implemented the generation of coordination expres- 
sions with various •complexities \[Dale, 1992, Dalianis and Hovy, 1993, Huang and Fiedler, 1997, 
Shaw, 1995, Callaway and Lester, 1997\]. Most systems handles simple coordination which con- 
tains only one conjoined syntactic constituents, such as subject, verb, or object. None of them 
handles ellipsis as CASPER does. CASPER tries to systematically find the most concise way to ex- 
press the propositions by looking through all the • propositions. In contrast, aggregation operators 
proposed in other work are local and does handle complex cases. In addition, the possibility •of too 
much information in a sentence has not received much attention. Most research simply ignores this 
possibility because the input to their sentence planners never exceeds a few clauses. 
7 Conclusion 
We describe how hypotactic •operators combine clauses using lexical information and how paratactic 
operators produce sentences with coordination. Through the use of look-ahead into the lexicon 
during the aggregation process to guarantee expressibility and by performing the task of sentence 
delimitation before lexical choice, the system can generate complex sentences efficiently. Since 
hypotactic, and paratactic operators are reusable, further speed-up in future generation system 
development is expected. 
8 Acknowledgments 
The author would like to thank Kathleen McKeown for her valuable advice and encouragement. 
This work is supported by DARPA Contract DAAL01-94-K-0119, the Columbia University Center 
for Advanced Technology in High Performance Computing and Comnmnications in Healthcare 
(funded by the New York State Science and Technology • Foundation) and NSF Grants GER-90- 
2406. 

References 
\[Callaway and Lester, 1997\] Callaway, C. B. and Lester, J. C. Dynamically improving explanations: A 
revision-based approach to •explanation generation. In Proc. of the 15th IJCAI, pages 952-958, Nagoya, 
Japan. . - ..... 
\[Carpenter, 1998\] Carpenter, B. Distribution, collection and quantification: A type-logical account. To 
appear in Linguistics and Philosophy. 
\[Dalai et al., 1996\] Dalai, M., Feiner, S., McKeown, K., Jordan, D., Allen, B., ar/d alSafadi, Y. MAGIC: An 
experimental system for generating multimedia briefings about post-bypass patient status. In Proc. 1996 
AMIA Annual Fall Syrup, pages 684--688, Washington, DC. 
\[Dale, 1992\] Dale, R. Generating Referring Expressions: Constructing Descriptions in a Domain of Objects 
and Processes. MIT Press, Cambridge, MA. 
\[Dalianis and Hovy, 1993\] Dalianis, H. and Hovyl E. Aggregation in natural language generation. In Proc. 
of the  th European Workshop on Natural Language Generation, Pisa, Italy. 
\[Elhadad et al., 1997\] Elhadad, M., McKeown, K., and Robin, J. Floating constraints in lexical choice. 
Computational Linguistics, 23(2):195-239~ 
\[Halliday, 1994\] Halliday , M. A. K. An Introduction to Functional Grammar. Edward Arnold, London, 2nd 
edition. 
\[Horacek, 1992\] Horacek, H: An integrated view of text planning. In Aspects of Automated Natural Language 
Generation, Lecture Notes in Artificial Intelligence, 587, pages 29-44. Springer-Verlag, Berlin. 
\[Hovy, 1993\] Hovy, E. H. Automated discourse generation using discourse structure relations. Artificial 
Intelligence, 63. Special Issue on NLP. 
\[Huang and Fiedler, 1997\] Huang, X. and Fiedler, A. Proof verbalization as an application of NLG. In Proc. 
of the 15th IJCAI, pages 965-970, Nagoya, Japan. 
\[Jackendoff, 1990\] Jackendoff, R. Semantic Structures. MIT Press, Cambridge, MA. 
\[Kaplan and Bresnan, 1982\] Kaplan, R. M. and Bresnan, J. Lexical-functional grammar: A formal sys- 
tem for grammatical representation. In Bresnan, J., editor, Th_e Mental Representation of Grammatical 
Relations, chapter 4. MIT Press. 
\[Kukich et al., 1994\] Kukich, K., McKeown, K., Shaw, J., Robin, J., Lim, J., Morgan, N., and Phillips, 
J. User-needs analysis and design methodology for an automated document generator. In Zampolli, A., 
Calzolari, N., and Palmer, M., editors, Linguistica Computazionale, Vol. IX-X, pages 109-115. Kluwer 
Academic Publishers, Norwell, MA. 
\[Mann and Moore, 1980\] Mann, W. C. and Moore, J. A. Computer as author - results and prospects. 
Technical Report RR-79-82, USC Information Science Institute, Marina del Rey, CA. 
\[Mann and Thompson, 1986\] Mann, \V. C. and Thompson, S. A. Rhetorical•structure theory: Description 
and construction of text structures. Technical Report RS-86-174, USC Information Sciences Institute, 
Marina Del Rey, CA. 
\[McCawley, 1981\] McCawley, J. D. Everything that linguists have always wanted to know about logic (but 
were ashamed to ask). University of Chicago Press. 
\[McDonald et al., 1987\] McDonald, D. D., Meteer, M. h'i., and Pustejovsky, J. D. Factors contributing to 
efficiency in natural language generation. In Kempen, G., editor, Natural Language Generation: New 
Results in Artificial Intelligence, Psychology and Linguistics, NATO ASI Series - 135, pages !59-182. 
Martinus Nijhoff Publishers, Boston. 
\[McKeown et al., 1994\] McKeown, K., Kukich, K., and Shaw, J. Practical issues in automaticdocumentation 
• generation. In Proc. of the 4th ACL Conference on Applied Natural Language Processing, pages 7-14, 
Stuttgart. 
\[McKeown et al., 1997\] McKeown, K., Pan, S., Shaw, J, Jordan, D., and Allen, B. Language generation for 
multimedia healthcare briefings. In Proc. of the Fifth ACL Conf. on ANLP, pages 277-282. 
\[Meteer, 1991\] Meteer, M. The implications of revisions for natural language generation. In Paris, C. L., 
Swartout, W. R., and Mann, W. C., editors, Natural Language Generation in Artificial Intelligence and 
Computational Linguistics, pages 155-178. Kluwer Academic Publishers, Boston. 
\[Quirk et al., 1985\] Quirk, R., Greebaum, S., Leech, G., and Svartvik, J. A Comprehensive Grammar of the 
English Language. Longman Publishers, London. 
\[Robin, 1995\] Robin, J. Revision-Based Generation of Natural Language Summaries Providing Historical 
Background. PhD thesis, Columbia University. 
\[Scott and de Souza, 1990\] Scott, D. R. and de Souza, C. S. Getting the message across in RST-based 
text generation. In Dale, R., Mellish, C., and Zock, M., editors, Current Research in Natural Language 
Generation, pages 47-73. Academic Press, New York. 
\[Shaw, 1995\] Shaw, J. Conciseness through aggregation in text generation: In Proc. of the 33rd A CL (Student 
Session), pages 329-331. 
\[Shaw, 1998\] Shaw, J. Segregatory coordination and ellipsis in text generation. In To appear in Proc. of the 
17th COLING and the 36th Annual Meeting of the ACL. 
\[Tai, 1969\] Tai, J. H.-Y. Coordination Reduction. PhD thesis, Indiana University. 
\[Wanner and Hovy, 1996\] V~ranner, L. and Hovy, E. The HealthDoc sentence planner. In Proc. of the 8th 
International Natural Language Generation Workshop, pages 1-10: Sussex, UK. 
