CAN COMPUTERS HANDLE ADVERBS?* 
SUMALI P1N-NGERN CONLON 
Dept. of Management and Marketing 
University of Mississippi 
University, MS 38677, U.S.A. 
601-232-5470, mksumali@vm.ec.olemiss.edu 
MARTHA EVENS 
Dept. of Computer Science 
Illinois Institute of Technology 
Chicago, IL 60616, U.S.A. 
312-567-5153, csevens@iitvax.bitnet 
ABSTRACT 
The adverb is the most complicated, and 
perhaps also the most interesting part of speech. 
Past research in natural language processing, 
however, has not dealt seriously with adverbs, 
though linguists have done significant work on 
this word class. The current paper draws on 
this linguistic research to organize an adverbial 
lexicon which will be useful for information 
retrieval and natural language processing 
systems. 
1. INTRODUCTION 
the semantic and syntactic properties of these 
different classes of adverbs. 
In this study, we give some indication of 
how a sophisticated and well structured 
adverbial lexicon can help NLP systems to 
handle adverbs. Our lexicon builds on the 
work of several linguists. Their work has 
helped us to organize information about adverbs 
from the Collins English Dictionary (CED) and 
other sources. The adverbial lexicon is 
organized using the Oracle Relational Database 
Management System. 
There have been many studies on nouns, 
verbs, and adjectives in NLP systems. Adverbs 
have received far less attention. 
It is very difficult for natural language 
processing (NLP) systems to handle adverbs 
because of the large number of syntactic roles 
that adverbs can assume in sentences. Adverbs 
can modify verbs, adjectives, adverbs, clauses, 
or the whole sentences. Adverbs obey 
complicated positional and selectional 
restrictions. Many adverbs change meanings 
when they change their positions in sentences. 
Despite this complexity, linguists have made 
significant progress developing a comprehensive 
set of adverbs classifications, and elucidating 
* This research was partially supported by the 
National Science Foundation under grant IRI- 
8704619. Our thanks also to Collins Publishers 
tot permission to use the dictionary tapes. 
Semantic classes of adverbs and 
selectional restrictions are discussed in Section 
2. Section 3 treats functional classes and 
describes some of the properties of certain 
functional classes of adverbs. Section 4 
presents some possible applications. Finally, 
Section 5 concludes. 
2. SEMANTIC CLASSIFICATION OF 
ADVERBS AND SELECTIONAL 
RESTRICTIONS. 
With minor variations, Jacobson \[1964\] 
and Quirk et al. \[1985, p438-74\] classify 
adverbs semantically as: adverbs of time, 
manner, degree, location, direction, transition 
and hedges. Adverbs of time, for example, 
answer questions such as "When?," "How 
often?," "How long?." Some examples of 
adverbs in this class are "before," "during," 
"now," "forever," "seldom," etc. Adverbs of 
manner answer questions like "How?" and "In 
what manner (way)?" Examples are "quickly," 
AcrEs DE COLING..92, NANTES, 23-28 AO~r 1992 1 i 9 2 PROC. OF COLING-92, NANTES, AU¢I. 23-28. 1992 
"like," "together," and "without." Hedges 
indicate fuzzy restrictions such as "essentially," 
"sort of," and "virtually" \[Lakoff, 1972\]. 
We have classified adverbs by 
systematically analyzing CED definitions. 
More details on this process can be found in 
Conlon et al., \[1990\] and Pin-Ngern \[1990\]. 
These semantic classes are stored in our 
adverbial lexicon as in Table 1. 
AD V_ CED (_W__QR~ ,H / G #, E~_I~S__F~, 
8UBSENSE#,SEMANTIC CLASS) 
about 1 2 l DEGREE, TIME 
actually 1 1 1 MOOD 
between 1 2 2 PLACE, DIRECTION 
down 1 2 7 POSITION 
evidently 1 1 1 MOOD 
Table 1. Adverb classification from CED 
The first line indicates that in CED, the 
adverb "about" with homograph number 1, 
sense number 2, and subsense number 1 
belongs to the flggr~ and time class. (To 
facilitate exposition, the tables here and below 
are not normalized.) 
Huang \[1975\] provides a similar 
classification of adverbs based on semantic 
usage with verbs. These classes of adverbs 
include: 
State of ~ ~. Adverbs in this 
group indicate a participant's state of mind. 
Examples are ~ and ~ in the 
following sentences: 
He rf_aLf.Ra~ stepped into the room. 
Mary was F_JllLegl2 taken advantage of... 
vA.d.x.cz~. These adverbs 
express an evaluation of the actor's 
participation in the event such as "clumsily" in 
"He clumsily stepped on the snail." 
of \]~5.UR. These describe 
aspects of the result of an event such as the 
adverb "correctly" in the sentence "He spelled 
the words correctly." 
Huang's classifications can also be 
incorporated into Table 1. 
These adverb classifications can be 
extremely useful to NIP systems, in particular, 
in the statement of selectional restrictions (or 
preferences). Huang describes a number of 
selectional restrictions between adverbs and 
other parts of speech. Some examples are: 
Adverbs of manner or direction can modify 
verbs of motion. 
Adverbs of result are proper to 
accomplishment and achievement verbs. 
Time and locative adverbs usually occur in 
action sentences and specify the temporal and 
spatial circumstances of the events. 
Degree adverbs express the degree or extent 
of a certain quality or state and presuppose an 
analysis of grading properties in the semantics 
of adjectives in particular. 
Conjunctive adverbs are used in almost the 
same way as conjunctions. 
This type of information should help 
text generation systems to select appropriate 
types of adverb to suit nouns, verbs, adjective, 
and adverbs which have already been chosen. 
It may also be useful for resolution of 
ambiguity in natural language understanding 
applications (see Radford \[1988, p372\] for an 
example of a sentence in which selectional 
restrictions resolve ambiguity). Information 
about selectional restriction are stored as in 
Table 2. 
ACTES I)E COLING-92, NAN'rE.S, 23-28 AO~" 1992 1 1 9 3 PROC. OF COLING-92, NANTES, AUG. 23-28, 1992 
ADV_MANNER~,H/G#,~F~, 
~UBSENSE#,VERB,SENTENCE) 
clumsily 1 1 1 dynamic action 
slowly 1 2 1 dynamic action 
Table 2. Selectional Restrictions for Adverb 
of Manner 
From the above examples, "clumsily" 
and "slowly" can modify dynamic verbs and 
action sentences. 
3. FUNCTIONAL CLASSIFICATION 
AND THE SYNTACTIC BEHAVIOR OF 
ADVERBS 
Jacobson \[1964\] groups adverbs into 
four major functional classes: modifying 
adverbs ("L.I/¢...~ he is ~ careful"), 
complementary adverbs ("Please come ~'), 
referential adverbs (" This section, in particular, 
is important"), and conjunctive adverbs ("He 
will not leave unless she asks him to"). 
Adverbs that function as modifiers, in 
turn, can modify verbs ("He spelledr,~_rLeC~"), 
adjectives ('They are ~ tall"), other adverbs 
("He works ~ well"), or even clauses 
and the whole sentences ("ForCunat¢ly he was 
not drunk"). 
These functional classes help to organize 
knowledge representation schemes involving 
adverbs, and are also essential for 
understanding the syntactic properties of 
adverbs. Jackendoff, for example, classifies 
adverbs as subject-oriented, speaker-oriented 
(sentence), and manner adverbs, and represents 
the meaning structures of these adverbs in 
essentially prolog form as follows: 
Adverb class: Sentence 
Prolog-like formula: ADJ(f(Np1,...,NPa)) 
Example: EVIDENT(WALKED(JOHN,IN)) 
It is evident that John walked in. 
Evidently, John walked in. 
Adverb class: Subject-oriented 
Prolog-like formula: ADJ(Npi, f(NP 1 ..... NPa)) 
Example: CLUMSY(FRED, 
DROPPED(FRED,THE BOOK)) 
It was clumsy of Fred to drop the book. 
Clumsily, Fred dropped the book. 
Adverb class: Manner 
Prolog-like formula: \[f/ADV\](Np1,...,NP n) 
Example: \[RAN/QUICKLY\] (SUE,HOME) 
Sue quickly ran home. 
Figure 1. 
Huang uses similar prolog-like formulas to 
represent the semantic structure of adverbs. 
This same functional classification also 
allows Jackendoff to derive positions of adverbs 
by using parse trees. According to Jackendoff, 
sentence adverbs must be attached to an S- 
node, while manner adverbs are attached to VP 
nodes. Thus the sentence adverb "evidently" 
can appear in positions such as the following: 
S 
ADV NP VP A 
V PP 
Evidently John walked in the store. 
Figure 2. 
ACRES DE COLING-92, NANTES, 23-28 not~W 1992 I 1 9 4 PROC. OF COLING-92, NArcrEs, Act;. 23-28, 1992 
S 
NP ADV VP 
V PP 
"d 1 walked in the store. John evl ent y 
Figure 3. 
A similar parse tree could be constructed for 
"John walked in the store, evidently." On the 
other hand, the sentence *'John walked 
evidently in the store," is impossible because 
the parse tree does not allow "evidently" to be 
attached to the S node (see figure 4.) 
NP ~ VP 
V ADV PP 
*John walked evid fly to the store 
Figure 4. 
For a similar discussion, see McCawley 
\[1988, p631=40\] and Radford \[1988, p73, 93\]. 
Classification of adverbs by functional class 
is gwen in tables such as Table 3. 
ADV_MOD W(_WQ_~_,H_H_LC.~,SENS~, 
SUBSENSE#,MODIFYING_CLASS) 
evidently 1 1 1 sentence 
happily 1 1 1 sentence, speaker, manner 
clumsily 1 1 1 subject, manner 
Table 3. Classifications of Modifying 
Adverb Based on Modification Class 
Additional information about adverbial 
positions in sentences has been collected by 
Jacobson \[1964\]. This information should 
supplement the information in Table 3, and 
should also be helpful for information retrieval 
and NLP applications. This information is 
organized as in Table 4. 
ADV-JACOBSON(~_,H/G#,SENS~, 
SUBSENSE#,POSITION) 
after 1 2 1 E4 
constantly 1 1 1 M1, M3, M4, El, E5 
exactly 1 1 2 M3, M4 
otherwise 1 2 1 M3, El, E4, E5 
Table 4. Adverb Positions in Sentences 
Collected by Jaeobson 
Thus, "constantly" is found in "middle" and 
"end" positions. More details about each 
position can be found in Jacobson \[1964\]. 
4. APPLICATIONS 
A sophisticated adverbial lexicon should help 
NLP systems to handle adverbs appropriately. 
NLP systems will be able to use information in 
the adverbial lexicon in parsing, language 
understanding, and text generation systems. 
Suppose, for example, that a parser has to 
parse a sentence of the form: "John \[Adv\] 
walked to the store." The parser can determine 
how to treat the adverb by looking in the 
lexicon to see whether the adverb is a sentence 
or a verb phrase adverb. For example, if 
"quickly" is the adverb in the sentence, it is a 
manner adverb, so it should be attached to the 
verb or verb phrase node. On the other hand, 
if the adverb is "evidently" then it is a mood 
adverb (speaker oriented sentence modifier), 
and so, should be attached to the S node. 
ACRES DE COLING-92, NANll~S, 23-28 AO~' 1992 1 1 9 5 t'ROC. O1: COLING-92, NANTES, AUG. 23-28, 1992 
Once the sentence has been successfully 
parsed, the language understanding system can 
then create prolog-like formulas of the form 
indicated in Figure 1. Similarly, the text 
generation system should be designed to take 
prolog-like formulas such as those in Figure 1 
and successfully translate them into natural 
language. The functional class then places 
restrictions on the position of the adverbs, as in 
Figures 2, 3, and 4. 
Finally, selectional restrictions can be used 
to reduce lexical ambiguity in language 
understanding systems, and assist in word 
choice in text generation (see Section 2). 
5. CONCLUSION 
NLP systems can handle adverbs, if the 
adverbial lexicon provides enough relevant 
information with appropriate knowledge 
representation techniques. The tables in our 
lexicon include, among others, 
Table 1: Semantic classifications of each 
adverb. 
Table 2: Selectional restrictions. 
Table 3: Modifying class (sentence, speaker, 
manner, etc.). 
Table 4. Jacobson's dictionary of adverb 
placement. 
This adverbial lexicon can be used in 
combination with Tables for other parts of 
speech to implement NLP systems. 
These tables, with others, should allow NLP 
systems to break down the intimidating 
complexity of the adverb class, and so, handle 
parsing of sentences with adverbs, adverb 
placement, knowledge representation related to 
adverbs, and selectional restrictions between 
adverbs and other parts of speech. 
6. REFERENCES 
Conlon, Sumali Pin-Ngern, Martha Evens, and 
Thomas Ahlswede, 1990. "Generating a 
Lexical Database for Adverbs." In Proceedings 
of the University of Waterloo Centre for the 
New Oxford English Dictionary. October 28- 
30, p95-I09. 
Huang, Shuan-Fan, 1975. A Study of Adverbs. 
Mouton, the Hague. 
Jacobson, Sven, 1964. Adverbial Positions in 
English. Dissertation, Uppsala, AB Studentbok, 
Stockholm. 
Jackendoff, Ray, 1972. Semantic Interpretation 
in Generative Grammar. MIT Press. p47-107. 
Lakoff George. 1972. "Hedges: A Study in 
Meaning Criteria and the Logic of Fuzzy 
Concepts." Proceedings of the Eight Regional 
Meeting of the Chicago Linguistic Society. 
April 14-16, p183-228. 
McCawley, James D. 1988. The Syntactic 
Phenomena of English. The University of 
Chicago Press. 
Pin-Ngem, Sumali. 1990. A Lexical Database 
for English to Support Information Retrieval, 
Parsing, and Text Generation. Unpublished 
Ph.D. Dissertation, Computer Science Dept., 
Illinois Institute of Technology. Chicago, IL. 
Quirk, Randolph, Sidney Greenbaum, Geoffrey 
Leech, Jan Svartvik. 1985. A Comprehensive 
Grammar of the English Language. Longman. 
Radford, Andrew. 1988. Transformational 
Grammar. Cambridge University Press. 
AcrEs DE COLING-92, NANTES, 23-28 AOUr 1992 1 1 9 6 PRoc. OF COLING-92, NANTES, AUG. 23-28, 1992 
