The Syntactic Classification of Adverbs as an Update to COMLEX Syntax:An Addition to an On-line Resource for Research in Syntax

1. Introduction
COMLEX (COMmon LEXicon), a large on-line syntactic dictionary for use in Natural Language Processing was created at New York University [4] under the auspices of the Linguistic Data Consortium (LDC) and is available for research and commercial purposes to members of the LDC. This large (38,000 English base forms) dictionary has syntactic classes for the major parts of speech: nouns, adjectives and verbs. We have now added syntactic classes for adverbs to COMLEX Syntax. The new updated version of COMLEX has been available from the LDC since January, 1998.
This paper discusses the types of syntactic information on adverbs that we have been able to capture in the dictionary and the notation used in entering this information in the dictionary. We also discuss some of the uses to which the adverb entries can be put.
2. Classification
The nouns, adjectives and verbs in COMLEX have both syntactic complements and syntactic features. The adverbs instead are classified in two different ways: (1) by position and (2) by features. The adverb features are different from the features for the other parts of speech we classified, in that they may be based on semantics rather than on syntax.
2.1 Positional Classification
Our positional classification is based on the adverb categorization developed at the Linguistic String Project by Naomi Sager [8]. This classification expresses modification relations. Among these are pre-adv (occurs before an adverb), pre-adj (occurs before an adjective), post-noun (occurs following a noun) and pre-quant (occurs before a quantifier i.e. a number). The following are some examples of modification relations (the adverb in question is capitalized).
“very fervidly”
(pre-adv)
“deeply unhappy”
(pre-adj)
“months ago”
(post-noun)
“approximately five”
(pre-quant)
The other positional classification is that of clausal adverb. This type can occur in various sentential positions. They can occur sentence initially (init), between the subject and the auxiliary verb (subj-aux), before the verb (pre-verb), after the auxiliary (post-aux), between the verb and its complement (verb-obj) and sentence finally (final). Note that post-aux is a sub-class of pre-verb, therefore an adverb marked pos-aux may not also be classified as pre-verb. Below are some examples of clausal adverbs.
“generally, he felt well.”
(init)
“They easily caught the rabbit.”
(pre-verb)
“He never could type.”
(subj-aux)
“They have not been active.
(post-aux)
“He participated actively in sports.”
(verb-obj)
“They ran to her immediately.”
(final)

Other essentially syntactic features based on Sager’s classes are PRED-ADV and TIMETAG. PRED-ADV are adverbs which can occur as the predicate of “be” and instantiate the COMLEX complements ADVP-PRED, NP-ADVP-PRED, and ADVP-PRED-RS.
For example:
“We were upstairs.”
“We will stay around.
“They believed him there.
The classification TIMETAG is given to adverbs involved in time expressions. When an adverb marked TIMETAG is used with certain time nouns (NTIME1 in COMLEX Syntax) it enables them to occur as bare noun phrases in sentence modifier positions as follows:
“A year ago, he traveled to Spain.”
*“A year, he traveled to Spain.”
“The package arrived two days later”
*“The package arrived two days
Two rare syntactic features, also based on Sager’s classes are COMPARATIVE and PERM-ADV. Only the verbs ‘more’ and ‘less’ are COMPARATIVE, their distinctive feature being that they can be modified by PRE-COMPARATIVE adverbs, a positional class which includes ‘far’, and ‘any’, as in:
“He didn’t have any more”
“It gave me far less satisfaction.”
PERM-ADV adverbs occur in clause initial position and allow subject-aux permutation, e.g.:
“Seldom has he seen such joy.”
“Long may he live.”
2.2 Semantic and Meta Features
Our main considerations in the creation of our more semantic adverb features were: (1) the extent to which these new distinctions would be needed in work on selection restrictions, scope and discourse; (2) the usefulness of these classes for identifying semantic roles in information extraction tasks, (e.g., c.f. the Proceedings of the Sixth Message Understanding Conference); (3) the role of these features in determining whether a clause is factive or nonfactive; and (4) the prevalence of these distinctions in the literature on adverbs (e.g., [6], [3] and [5]).
Most of our semantic features are essentially subclasses of our positional classes. These are given below:
“John treated her well.” [PRED-ADV + EVAL-ADV]
“Mary put the book down.” [PRED-ADV + LOC&DIR-ADV]
“They ate brunch downstairs.” [CLAUSAL-ADV + LOC&DIR-ADV]
“John walked clumsily.”[CLAUSAL-ADV + MANNER-ADV]
“Suddenly, Mary left.”[CLAUSAL-ADV + TEMPORAL-ADV]
“John is quite angry.”[PRE-ADJ + DEGREE-ADV]
“John spoke quite angrily to Mary.” [PRE-ADV + DEGREE-ADV]
“Unfortunately, Celia fell.” [CLAUSAL-ADV + META-ADV]
PRED-ADV includes EVAL-ADV (evaluative) and LOC&DIR-ADV (locative/directional) subclasses, both subclasses based on [8]. These subclasses may provide the basis for differentiating complements assigned to verbs which take PRED-ADV complements. A subset of PRE-ADJ and PRE-ADV are also marked DEGREE-ADV, expressing the degree of an attribute. Clausal-adv includes MANNER-ADV (the manner in which an action is performed), TEMPORAL-ADV (time, frequency, etc.), LOC&DIR-ADV and META-ADV subclasses.
META-ADV (often referred to as sentential adverbs in the linguistics literature) items have the following two properties: (1) they can occur with any verb (other clausal adverbs impose selectional restrictions on the verb); (2) they are outside the scope of sentential negation. For example, “John will not presumably leave the country” means that it is presumable that John will not leave (“presumably” is not negated). In contrast, “John will not suddenly leave the country” means that it is not the case that John will leave suddenly (“suddenly” is negated). We recognize five subclasses of META-ADV (cf. [6], [3], and [5]), including: :CONJ (connects sentences), :ATTITUDE (expresses speaker’s opinion), :VIEWPOINT (puts sentence in a context), :PERFORMATIVE (expresses speaker’s manner of speaking or state of mind when speaking) and :EPISTEMIC (expresses likelihood), e.g.,

“Mary was evil. However, John was good.”
[META-ADV :CONJ T]
“Unfortunately, John will never win.”
[META-ADV :ATTITUDE T]
(the speaker thinks it is unfortunate)

“Biologically, John is a mammal.”
[META-ADV :VIEWPOINT T]
“Frankly, I don’t care.”
[META-ADV :PERFORMATIVE T]
(the speaker is being frank)

“Probably, they will arrive late.”
[META-ADV EPISTEMIC T]

GRADABLE adverbs are adverbs which can be modified by ‘more’, ‘most’, ‘less’ or ‘least’, or which can, in rare cases have ‘-er’ and ‘-est’ suffixes:
“more happily engaged”/“the most happily engaged couple”
“less fortunately favored.”/“the least adequately funded
“she left sooner than her mother did.”
“he threw farther than she.”
NEGATIVE adverbs (cf. [7], [2]) permit items requiring negative polarity and preclude positive polarity items.
“He did not go.”
“She will never win.”
“She could not bear the thought.”
“She could hardly bear the thought.”
*“She bore the thought.She seldom/barely/hardly ever ate broccoli.”
*“She ever ate broccoli.”
3. Notation and Entry
The notation for the adverbs corresponds to that of the rest of COMLEX: a Lisp-like notation consisting of feature-value pairs. Just as we did for COMLEX Syntax, we have endeavored to make this classification theory neutral, by using generally accepted linguistic terminology in defining our classes. In this way, the lexicon can be broadly used across different areas of study, by scholars from differing theoretical backgrounds.
Figure 1 shows some examples of adverb entries. :MODIF stands for modification position. The other classes have been explained above.

(ADVERB :ORTH “ago”
:MODIF ((POST-NOUN))

:FEATURES ((TIMETAG)))
(ADVERB :ORTH “there”
:MODIF ((CLAUSAL-ADV

:INIT T

:FINAL T) (PRED-ADV))

:FEATURES((LOC&DIR-ADV)))
(ADVERB :ORTH “evidently”
:MODIF((CLAUSAL-ADV

:PRE-VERB

:SUBJ-AUX T

:FINAL T :INIT T)

(PRE-ADJ))

:FEATURES ((META-ADV

:ATTITUDE T)))
Figure 1. Sample COMLEX Syntax Adverb Entrie

The adverbs were entered by our staff of two part-time Linguistics graduate students using the entry program developed for COMLEX Syntax. They consulted our on-line concordance and hard-copy dictionaries. Another important resource available to us, through the kindness of Lou Burnard, was the BNC (British National Corpus ) [1] which we accessed on the Web. We were able to get fifty random examples for any entry. The reason that this was so valuable to us was that we could specifically ask for adverbs. Adverbs and other parts of speech commonly overlap, for example, ‘about’ is an adverb, a preposition and a verb particle; while ‘hard’ is both an adjective and an adverb. Therefore, it was very handy in weeding out non-applicable examples. It was also useful in finding examples for rare adverbs, many of which did not appear at all in our corpus (e.g. ‘ad libitum’). Finding examples of rare words is possible because the BNC is so large (100 million words) and also balanced (it has many sources).
4. COMLEX Adverbs for Linguistic Research
One of the advantages of having a large number of adverbs classified and entered in a machine readable lexicon, is that they are available for on-line queries. By using our query facility for COMLEX you may enquire as to which classes overlap or which never co-occur. For example, if you believe that adverbs classified as META :CONJ T always occur in initial position, you might look for exceptions by posing the following query:
(get-feature-set ’(and (meta-adv :conj t)(not (clausal-adv :init t))))
This query produced three adverbs which did not have both classes
(‘else’ ‘thereby’ ‘too’)
The following query
(get-feature-set ‘meta-adv :conj t))
produced a list of thirty-six adverbs. The assumption here is that either the three above are misclassified or they are counter-examples. On closer scrutiny, we found that two adverbs (‘else’ and ‘thereby’) should have been classified as sentence initial adverbs. We have since reclassified these correctly. The remaining adverb (‘too’) is, in fact, a counter-example, it is a meta-adv which is a conjunction but does not occur in initial position. This can be seen in the following examples
.“Jane went. *Too, John went.”
“Jane went. John went too.”
“I like John. However, he is a simpleton.”
“I like John. He is a simpleton, however.”
5.Conclusion
The addition of syntactic/semantic classification of adverbs to COMLEX Syntax will be of significant value to researchers in Linguistics and other related fields. Easy access to a large amount of lexical data on adverbs can be of assistance in ongoing research as well as providing insights into new areas to explore. Using the query tool developed earlier for COMLEX Syntax allows complex logical searches to be performed quickly and efficiently. Having the data strictly organized according to well-defined classes facilitates searches using other types of on-line search engines, as well (grep, for instance). These types of searches are not possible using hard copy off-line dictionaries.
References:
1. G. Aston and L. Burnard The BNC Handbook Edinburgh University Press, 1998.

2. S. P. Colon and M. Evens An Adverbial Lexicon for Natural Language Processing Systems International Journal of Lexicography, Vol. 7 No. 3,1993.

3. Sidney Greenbaum Studies in English Adverbial Usage Coral Gables, Florida: University of Miami Press, 1969.

4. Ralph Grishman, Catherine Macleod, and Adam Meyers Comlex Syntax: Building a Computational Lexicon. The Proceedings of COLING94, Vol I, pages 268-272, Kyoto, Japan, August, 1994.

5. Shuan-Fan Huang A Study of Adverbs The Hague: Mouton & Co. B. V., 1975.

6. Sven Jacobson Adverbial Positions in English Uppsala Dissertation, Stockholm: Studentbok, 1964.

7. J. D. McCawley The Syntactic Phenomena of English. Vol. 1 The University of Chicago Press, 1988.

8. Naomi Sager Natural Language Information Processing: a Computer Grammar of English and its Applications Addison-Wesley Publishing Company, Inc., 1981.