An Electronic Lexicon of Nominalizations: NOMLEX

New York University (NYU) has recently completed Nomlex, a dictionary containing detailed syntactic information about 1000 common English nominalizations. This dictionary, which has been developed for use in natural language processing, is freely available from NYU.

History:

Previous dictionary work at NYU includes COMLEX Syntax [Macleod'97], a large syntactic dictionary with detailed information on the syntactic properties of English words, especially with regard to complement structure. This dictionary is available through the Linguistic Data Consortium (LDC) and is being presently used by several natural language processing (NLP) groups.

Nominalizations present a special problem, in that one wants not only to analyze their syntactic structure, but also to relate them to corresponding verbal structures. This is essential for natural language interpretation in applications such as question answering and information extraction. For example, the answer to a question, "How badly was the city bombed?" could be phrased "The city was destroyed completely" or "The destruction of the city was complete." These answers though syntacticly diverse contain exactly the same information.

Dictionary structure:

The challenge in designing the NOMLEX entries is to provide, in reasonably compact form, all the lexically-specific syntactic information required to relate nominal arguments to the corresponding verbal arguments. This task is complicated because of the wide range of nominal argument structures, which makes a direct enumeration of all possible correspondences hopelessly unwieldy. In particular, the core arguments and oblique arguments behave differently in this respect.

The core arguments of the verb (subject, object, indirect object) may appear as possessive determiners (DET-POSS), noun noun modifiers (N-N-MOD) or in a prepositional phrase commonly preceded by "of" (PP-OF). Examples of this are: "His death" (where "his" represents the one who dies and therefore the subject of the verb "die"), "the price adjustment" (where "price" is the object of the verb "adjust"), and "the discussion of the case" (where "case" is the object of the verb "discuss" and the analysis would be "X discuss the case").

More complex verbal complementation such as sentential and verbal complements are found following the nominalization and often retain the same structure. For instance, the complement "that he came" is the same for the nominalization as for the verb, seen in the following examples.

"Someone report(ed) that he came."
"The report that he came."

Some verbal complements must have an introductory preposition in order to appear as nominalization complements. For example, "Someone questioned whether it was a wise plan." versus "The question of whether it was a wise plan". We encode all these possibilities in the lexical entries of the nominalizations. The NOMLEX entry for "appointment" is as follows.

(NOM :ORTH "appointment"
:VERB "appoint"
:PLURAL "appointments"
:NOUN ((EXISTS))
:NOM-TYPE ((VERB-NOM))
:VERB-SUBJ ((N-N-MOD) (DET-POSS))
:SUBJ-ATTRIBUTE ((COMMUNICATOR))
:OBJ-ATTRIBUTE ((NHUMAN))
:VERB-SUBC ((NOM-NP :OBJECT ((DET-POSS) (N-N-MOD) (PP-OF)))
(NOM-NP-PP :OBJECT ((DET-POSS) (N-N-MOD) (PP-OF))
:PVAL ("for" "to"))
(NOM-NP-TO-INF-OC :OBJECT ((DET-POSS) (PP-OF)))
(NOM-NP-AS-NP :OBJECT ((DET-POSS) (PP-OF)))))

This sample entry, in combination with a set of defaults, provides us with a lot of information, including the following:

A homograph noun of the nominalization "appointment" exists

("a dental appointment" is not the act of "appointment")
The subject of the verb tends to be a human, company or other entity capable of communication.
The object of the verb is a human.
The nominalization is of type VERB-NOM, meaning that the nominalization does not play the grammatical role of any of the arguments of the verb. In contrast, "fighter" is a SUBJECT nominalization and "interviewee" is an OBJECT nominalization.
The subject of the verb can occur in three positions:

N-N-MOD - "The IBM appointment of Mary Smith"
DET-POSS - "IBM's appointment of Mary Smith"
PP-by - "The appointment of Mary Smith by the president"
(PP-by is assumed as a subject marker by default.)
When the verb is followed by one noun phrase and no other complement phrases, that object may occur in the nominalized phrase (NOM-NP) in one of three positions:
DET-POSS - "Mary Smith's appointment"
N-N-MOD - "The Mary Smith appointment" <
PP-OF - "The appointment of Mary Smith by the president"
Object plus prepositional phrase complements ("for" and "to") share the same object mappings as simple NP complements.
In addition, the prepositions in the verb complement are realized as the same preposition in the nominalization complement.
Complements consisting of an NP object plus either an infinitival complement or an AS-NP complement allow only DET-POSS and PP-OF complements. In addition, the verbal complement phrases TO-INF-OC and AS-NP map to these same types of phrases in the nominalization (by default).
Due to space considerations, we must leave out a lot of detail regarding the interpretation of this dictionary entry. In particular, we have a system of rules and defaults which prevent spurious ambiguity and allow us to keep the lexical entries compact. The example cited above is that PP-by is so often a marker of the subject of the verb, that we treat this as a default, marking nominalizations that do not allow this mapping with NOT-PP-by. For more detail about NOMLEX, please see our web site: <http://cs.nyu.edu/cs/projects/proteus/nomlex/index.html>
Using NOMLEX:

To demonstrate the utility of NOMLEX we constructed two NLP applications that depend on NOMLEX entries to analyze nominalization phrases.

A program that transforms a nominalization phrase into one or more sentences, corresponding to the possible senses of the nominalization phrase. For example, "Rome's destruction of Carthage" has exactly one sense, which our program would paraphrase as "Rome destroyed Carthage". The program takes a grammatical analysis (a parse) of a nominalization phrase as input and uses NOMLEX to create grammatical analyses of the corresponding sentences, copying noun phrases and complement phrases from the various nominal positions (N-N-MOD, DET-POSS, post-noun, etc.) into the appropriate sentential positions (SUBJECT, OBJECT, etc.). Actual sentences are then generated from these parses. The output of this program could be used as input to any NLP application designed to operate on full sentences.
A program which converts an information extraction pattern designed for sentences into a set of information extraction patterns designed for nominalization phrases ([Meyers'98]). For our purposes, an information extraction pattern is a pattern used to identify events in text and correctly mark the participants of these events. One such system extracts information about corporate hirings, firings, resignations, etc., including the identification of who left which company, who joined which company, the positions they left, the positions they attained, dates, etc. Our program converts a pattern that analyzes "IBM appointed Alice Smith as Vice President" into a pattern that analyzes: "IBM's appointment of Alice Smith as President", "Alice Smith's appointment by IBM", "IBM's Alice Smith appointment", "IBM's appointee", "The appointee of IBM", etc. Nomlex helps determine where information identified by a sentence pattern would be found in a nominalization phrase. This would enable an information extraction tool to easily extract information from nominalizations.
Process:
The menu-based entry program used for COMLEX was adapted to enter NOMLEX. As for COMLEX, the entry program gave us access to a large corpus of text (including the Brown Corpus and large amounts of newspaper articles) and we had limited access to the British National Corpus (BNC) as well. Two half-time ELFs (Enterer of Lexical Features) worked for 5 months and then one half-time ELF completed the lexicon in two years.

Concluding Remarks:

We have created a dictionary with the goal of solving a pervasive problem in NLP. Most grammatical analyses are designed to process sentences, but in order not to miss information these need to be applied to nominalization phrases, as well. NOMLEX provides a bridge in the form of: (1) a set of rules and defaults; and (2) a dictionary record of idiosyncratic mappings between nominalizations and their related verbs. Before the creation of NOMLEX, developers and researchers have had to adapt processes which handle sentences to enable them to handle nominalizations also. Due to the high overhead of this endeavor, many systems do not handle nominalizations at all or use simple, but imprecise heuristics. NOMLEX makes it less expensive to fully integrate nominalizations into an NLP system.

Future work:

In order to make NOMLEX an even more useful resource, we plan to add support verbs to the entries. The use of support verbs changes the relationship of the nominal and verbal arguments. For example, in "his visit to Mary" the possessive determiner is the subject of "visit" (i.e. "he visit Mary"); in "he made a visit to Mary" the subject of the support verb "make" is now the subject of "visit" (i.e. "he visited Mary."). The demonstration that this is idiosyncratic and a matter for lexical interpretation is the appearance of "he had a visit from Mary" where the object of the preposition "from" is the subject of "visit" and the subject of the support verb becomes the object of "visit" (i.e. "Mary visited him"). We have been exploring Mel'cuk's notation as a means of capturing these relationships for NOMLEX.

Bibliography

Macleod, C., Grishman, R. and Meyers, A. (1997/1998) COMLEX Syntax: A Large Syntactic Dictionary for Natural Language Processing. Computers and the Humanities (CHUM), Kluwer Academic Publishers, Vol. 31 No. 6.
Meyers, A., Macleod, C., Yangarber, R., Grishman, R., Barrett L., and Reeves, R. (1998). Using NOMLEX to Produce Nominalization Patterns for Information Extraction. Coling-ACL98 workshop Proceedings: the Computational Treatment of Nominal.