Noun-Phrase Analysis in Unrestricted Text for Information Retrieval 
David A. Evans, Chengxiang Zhai 
Laboratory for Computational Linguistics 
Carnegie Mellon Univeristy 
Pittsburgh, PA 15213 
dae@cmu.edu, cz25@andrew.cmu.edu 
Abstract 
Information retrieval is an important ap- 
plication area of natural-language pro- 
cessing where one encounters the gen- 
uine challenge of processing large quanti- 
ties of unrestricted natural-language text. 
This paper reports on the application of a 
few simple, yet robust and efficient noun- 
phrase analysis techniques to create bet- 
ter indexing phrases for information re- 
trieval. In particular, we describe a hy- 
brid approach to the extraction of mean- 
ingful (continuous or discontinuous) sub- 
compounds from complex noun phrases 
using both corpus statistics and linguistic 
heuristics. Results of experiments show 
that indexing based on such extracted sub- 
compounds improves both recall and pre- 
cision in an information retrieval system. 
The noun-phrase analysis techniques are 
also potentially useful for book indexing 
and automatic thesaurus extraction. 
1 Introduction 
1.1 Information Retrieval 
Information retrieval (IR) is an important applica- 
tion area of naturaManguage processing (NLP). 1 
The IR (or perhaps more accurately "text retrieval") 
task may be characterized as the problem of select- 
ing a subset of documents (from a document col- 
lection) whose content is relevant to the informa- 
tion need of a user as expressed by a query. The 
document collections involved in IR are often gi- 
gabytes of unrestricted natural-language text. A 
user's query may be expressed in a controlled lan- 
guage (e.g., a boolean expression of keywords) or, 
more desirably, a natural language, such as English. 
A typical IR system works as follows. The doc- 
uments to be retrieved are processed to extract in- 
dexing terms or content carriers, which are usually 
(Evans, 1990; Evans et al., 1993; Smeaton, 1992; Lewis 
& Sparck Jones, 1996) 
single words or (less typically) phrases. The index- 
ing terms provide a description of the document's 
content. Weights are often assigned to terms to in- 
dicate how well they describe the document. A 
(natural-language) query is processed in a similar 
way to extract query terms. Query terms are then 
matched against the indexing terms of a document 
to determine the relevance of each document to the 
quer3a 
The ultimate goal of an IR system is to increase 
both precision, the proportion of retrieved docu- 
ments that are relevant, as well as recall, the propor- 
tion of relevant document that are retrieved. How- 
ever, the real challenge is to understand and rep- 
resent appropriately the content of a document and 
quer~ so that the relevance decision can be made ef- 
ficiently, without degrading precision and recall. A 
typical solution to the problem of making relevance 
decisions efficient is to require exact matching of in- 
dexing terms and query terms, with an evaluation 
of the 'hits' based on a scoring metric. Thus, for 
instance, in vector-space models of relevance rank- 
ing, both the indexing terms of a document and the 
query terms are treated as vectors (with individual 
term weights) and the similarity between the two 
vectors is given by a cosine-distance measure, es- 
sentially the angle between any two vectors? 
1.2 Natural-Language Processing for IR 
One can regard almost any IR system as perform- 
ing an NLP task: text is 'parsed" for terms and 
terms are used to express 'meaning'--to capture 
document content. Clearly, most traditional IR sys- 
tems do not attempt to find structure in the natural- 
language text in the 'parsing' process; they merely 
extract word-like strings to use in indexing. Ide- 
ally, however, extracted structure would directly re- 
flect the encoded linguistic relations among terms-- 
captuing the conceptual content of the text better 
than simple word-strings. 
There are several prerequisites for effective NLP 
in an IR application, including the following. 
2 (Salton & McGill, 1983) 
17 
1. Ability to process large amounts of text 
The amount of text in the databases accessed by 
modem IR systems is typically measured in gi- 
gabytes. This requires that the NLP used must 
be extraordinarily efficient in both its time and 
space requirements. It would be impractical 
to use a parser with the speed of one or two 
sentences per second. 
2. Ability to process unrestricted text 
The text database for an IR task is generally 
unrestricted natural-language text possibly en- 
compassing many different domains and top- 
ics. A parser must be able to manage the many 
kinds of problems one sees in natural-language 
corpora, including the processing of unknown 
words, proper names, and unrecognized struc- 
tures. Often more is required, as when spelling, 
transcription, or OCR errors occur. Thus, the 
NLP used must be especially robust. 
3. Need for shallow understanding 
While the large amount of unrestricted text 
makes NLP more difficult for IR, the fact that 
a deep and complete understanding of the text 
may not be necessary for IR makes NLP for IR 
relatively easier than other NLP tasks such as 
machine translation. The goal of an IR system 
is essentially to classify documents (as relevant 
or irrelevant) vis-a-vis a query. Thus, it may 
suffice to have a shallow and partial represen- 
tation of the content of documents. 
Information retrieval thus poses the genuine chal- 
lenge of processing large volumes of unrestricted 
natural-language text but not necessarily at a deep 
level. 
1.3 Our Work 
This paper reports on our evaluation of the use of 
simple, yet robust and efficient noun-phrase analy- 
sis techniques to enhance phrase-based IR. In par- 
ticular, we explored an extension of the ~phrase- 
based indexing in the CLARIT TM system ° using 
a hybrid approach to the extraction of meaning- 
ful (continuous or discontinuous) subcompounds 
from complex noun phrases exploiting both corpus- 
statistics and linguistic heuristics. Using such sub- 
compounds rather than whole noun phrases as in- 
dexing terms helps a phrase-based IR system solve 
the phrase normalization problem, that is, the prob- 
lem of matching syntactically different, but semanti- 
cally similar phrases. The results of our experiments 
show that both recall and precision are improved by 
using extracted subcompounds for indexing. 
2 Phrase-Based Indexing 
The selection of appropriate indexing terms is criti- 
cal to the improvement of both precision and recall 
in an IR task. The ideal indexing terms would di- 
rectly represent the concepts in a document. Since 
'concepts' are difficult to represent and extract (as 
well as to define), concept-based indexing is an 
elusive goal. Virtually all commercial IR systems 
(with the exception of the CLARIT system) index 
only on "words', since the identification of words in 
texts is typically easier and more efficient than the 
identification of more complex structures. How- 
ever, single words are rarely specific enough to sup- 
port accurate discrimination and their groupings 
are often accidental. An often cited example is the 
contrast between "junior college" and "college ju- 
nior". Word-based indexing cannot distinguish the 
phrases, though their meanings are quite different. 
Phrase-based indexing, on the other hand, as a step 
toward the ideal of concept-based indexing, can ad- 
dress such a case directly. 
Indeed, it is interesting to note that the use 
of phrases as index terms has increased dramat- 
ically among the systems that participate in the 
TREC evaluations. ~ Even relatively traditional 
word-based systems are exploring the use of multi- 
word terms by supplementing words with sta- 
tistical phrases--selected high frequency adjacent 
word pairs (bigrams). And a few systems, such 
as CLARIT--which uses simplex noun phrases, 
attested subphrases, and contained words as in- 
dex terms--and New York University's TREC 
systemS--which uses "head-modifier pairs" de- 
rived from identified noun phrases--have demon- 
strated the practicality and effectiveness of thor- 
ough NLP in IR tasks. 
The experiences of the CLAR1T system are in- 
structive. By using selective NLP to identify sim- 
plex NPs, CLARIT generates phrases, subphrases, 
and individual words to use in indexing documents 
and queries. Such a first-order analysis of the lin- 
guistic structures in texts approximates concepts 
and affords us alternative methods for calculating 
the fit between documents and queries. In particu- 
lar, we can choose to treat some phrasal structures 
as atomic units and others as additional informa- 
tion about (or representations of) content. There are 
immediate effects in improving precision: 
1. Phrases can replace individual indexing words. 
For example, if both "dog" and "hot" are used 
for indexing, they will match any query in 
which both words occur. But if only the phrase 
"hot dog" is used as an index term, then it will 
only match the same phrase, not any of the in- 
dividual words. 
3(Evans et al., 1991; Evans et al., 1993; Evans et al., 
1995; Evans et al., 1996) 
4 (Harman, 1995; Harman, 1996) 
5 (Strzalkowski, 1994) 
18 
2. Phrases can supplement word-level matches. 
For example, if only the individual words "ju- 
nior" and "college" are used for indexing, both 
"junior college" and "college junior" will match 
a query with the phrase "junior college" equally 
well. But if we also use the phrase "junior col- 
lege" for indexing, then "junior college" will 
match better than "college junior", even though 
the latter also will receive some credit as a 
match at the word level. 
We can see, then, that it is desirable to distinquish-- 
and, if possible, extract--two kinds of phrases: 
those that behave as lexical atoms and those that re- 
flect more general linguistic relations. 
Lexical atoms help us by obviating the possibility 
of extraneous word matches that have nothing to 
do with true relevance. We do not want "hot" or 
"dog" to match on "hot dog". In essence, we want to 
eliminate the effect of the independence assumption 
at the word level by creating new words--the lexical 
atoms--in which the individual word dependencies 
are explicit (structural). 
More general phrases help us by adding detail. 
Indeed, all possible phrases (or paraphrases) of ac- 
tual content in a document are potentially valuable 
in indexing. In practice, of course, the indexing 
term space has to be limited, so it is necessary to se- 
lect a subset of phrases for indexing. Short phrases 
(often nominal compounds) are preferred over long 
complex phrases, because short phrases have bet- 
ter chances for matching short phrases in queries 
and will still match longer phrases owing to the 
short phrases they have in common. Using only 
short phrases also helps solve the phrase normal- 
ization problem of matching syntactically different 
long phrases (when they share similar meaning). 6 
Thus, lexical atoms and small nominal com- 
pounds should make good indexing phrases. 
While the CLARIT system does index at the level 
of phrases and subphrases, it does not currently 
index on lexical atoms or on the small compounds 
that can be derived from complex NPs, in particular, 
reflecting cross-simplex NP dependency relations. 
Thus, for example, under normal CLARIT process- 
ing the phrase "the quality of surface of treated 
stainless steel strip "7 would yield index terms such 
as "treated stainless steel strip", "treated stainless 
steel", "stainless steel strip", and "stainless steel" 
(as a phrase, not lexical atom), along with all the 
relevant single-word terms in the phrase. But the 
process would not identify "stainless steel" as a po- 
tential lexical atom or find terms such as "surface 
quality", "strip surface", and "treated strip". 
To achieve more complete (and accurate) phrase- 
based indexing, we propose to use the following 
6 (Smeaton, 1992) 
ZThis is an actual example from a U.S. patent 
document. 
four kinds of phrases as indexing terms: 
1. Lexical atoms (e.g., "hot dog" or 
2. 
3. 
4. 
perhaps 
"stainless steel" in the example above) 
Head modifier pairs (e.g., "treated strip" and 
"steel strip" in the example above) 
Subcompounds (e.g., "stainless steel strip" in 
the example above) 
Cross-preposition modification pairs (e.g., 
"surface quality" in the example above) 
In effect, we aim to augment CLARIT indexing with 
lexical atoms and phrases capturing additional (dis- 
continuous) modification relations than those that 
can be found within simplex NPs. 
It is clear that a certain level of robust and effi- 
cient noun-phrase analysis is needed to extract the 
above four kinds of small compounds from a large 
unrestricted corpus. In fact, the set of small com- 
pounds extracted from a noun phrase can be re- 
garded as a weak representation of the meaning of 
the noun phrase, since each meaningful small com- 
pound captures a part of the meaning of the noun 
phrase. In this sense, extraction of such small com- 
pounds is a step toward a shallow interpretation 
of noun phrases. Such weak interpretation is use- 
ful for tasks like information retrieval, document 
classification, and thesaurus extraction, and indeed 
forms the basis in the CLARIT system for automated 
thesaurus discovery. 
3 Methodology 
Our task is to parse text into NPs, analyze the noun 
phrases, and extract the four kinds of small com- 
pounds given above. Our emphasis is on robust 
and efficient NLP techniques to support large-scale 
applications. 
For our purposes, we need to be able to identify 
all simplex and complex NPs in a text. Complex 
NPs are defined as a sequence of simplex NPs that 
are associated with one another via prepositional 
phrases. We do not consider simplex NPs joined by 
relative clauses. 
Our approach to NLP involves a hybrid use of 
corpus statistics supplemented by linguistic heuris- 
tics. We assume that there is no training data (mak- 
ing the approach more practically useful) and, thus, 
rely only on statistical information in the document 
database itself. This is different from many cur- 
rent statistical NLP techniques that require a train- 
ing corpus. The volume of data we see in IR tasks 
also makes it impractical to use sophisticated statis- 
tical computations. 
The use of linguistic heuristics can assist statis- 
tical analysis in several ways. First, it can focus 
the use of statistics by helping to eliminate irrele- 
vant structures from consideration. For example, 
syntactic category analysis can filter out impossible 
19 
word modification pairs, such as \[adjective, adjec- 
tive\] and \[noun, adjective\]. Second, it may improve 
the reliability of statistical decisions. For example, 
the counting ofbigrams that occur only within noun 
phrases is more reliable for lexical atom discovery 
than the counting of all possible bigrams that occur 
in the corpus. In addition, syntactic category anal- 
ysis is also helpful in adjusting cutoff parameters 
for statistics. For example, one useful heuristic is 
that we should use a higher threshold of reliability 
(evidence) for accepting the pair \[adjective, noun\] 
as a lexical atom than for the pair \[noun, noun\]: a 
noun-noun pair is much more likely to be a lexical 
atom than an adjective-noun one. 
The general process of phrase generation is illus- 
trated in Figure 1. We used the CLARIT NLP mod- 
ule as a preprocessor to produce NPs with syntactic 
categories attached to words. We did not attempt 
to utilize CLARIT complex-NP generation or sub- 
phrase analysis, since we wanted to focus on the 
specific techniques for subphrase discovery that we 
describe in this paper. 
I Raw Text 
~Np CLARIT 
Extractor I 
NPs 
NP Parser ~ 
I ' ~( Lexical Atoms 9 
/ Structured/~k Attested Terms ,NPs / ~ 
Subcompound / 
Generator / 
Meaningful Subcompounds 
Figure 1: General Processing for Phrase Generation 
After preprocessing, the system works in two 
stages--parsing and generation. In the parsing 
stage, each simplex noun phrase in the corpus is 
parsed. In the generation stage, the structured noun 
phrase is used to generate candidates for all four 
kinds of small compounds, which are further tested 
for occurrence (validity) in the corpus. 
Parsing of simplex noun phrases is done in mul- 
tiple phases. At each phase, noun phrases are par- 
tially parsed, then the partially parsed structures are 
used as input to start another phase of partial pars- 
ing. Each phase of partial parsing is completed by 
concatenating those most reliable modification pairs 
together to form a single unit. The reliability of a 
modification pair is determined by a score based 
on frequency statistics and category analysis and 
is further tested via local optimum phrase analysis 
(described below). Lexical atoms are discovered at 
the same time, during simplex noun phrase parsing. 
Phrase generation is quite simple. Once the struc- 
ture of a noun phrase (with marked lexical atoms) 
is known, the four kinds of small compounds can 
be easily produced. Lexical atoms are already avail- 
able. Head-modifier pairs can be extracted based on 
the modification relations implied by the structure. 
Subcompounds are just the substructures of the NP. 
Cross-preposition pairs are generated by enumerat- 
ing all possible pairs of the heads of each simplex 
NP within a complex NP in backward order. 8 
To validate discontinuous compounds such as 
non-sequential head-modifier pairs and cross- 
preposition pairs, we use a standard technique of 
CLARIT processing, viz., we test any nominated 
compounds against the corpus itself. If we find 
independently attested (whole) simplex NPs that 
match the candidate compounds, we accept the 
candidates as index terms. Thus for the NP "the 
quality of surface of treated stainless steel strip", 
the head-modifier pairs "treated strip", "stain- 
less steel", "stainless strip", and "steel strip", and 
the cross-preposition pairs "strip surface", "surface 
quality", and "strip quality", would be generated 
as index terms only if we found independent evi- 
dence of such phrases in the corpus in the form of 
free-standing simplex NPs. 
3.1 Lexical Atom Discovery 
A lexical atom is a semantically coherent phrase 
unit. Lexical atoms may be found among proper 
names, idioms, and many noun-noun compounds. 
Usually they are two-word phrases, but sometimes 
they can consist of three or even more words, as 
in the case of proper names and technical terms. 
Examples of lexical atoms (in general English) are 
"hot dog", "tear gas", "part of speech", and "yon 
Neumann". 
However, recognition of lexical atoms in free text 
is difficult. In particular, the relevant lexical atoms 
for a corpus of text will reflect the various discourse 
domains encompassed by the text. In a collection 
of medical documents, for example, "Wilson's dis- 
ease" (an actual rheumatological disorder) may be 
used as a lexical atom, whereas in a collection of 
general news stories, "Wilson's disease" (reference 
to the disease that Wilson has) may not be a lexi- 
cal atom. Note that in the case of the medical us- 
age, we would commonly find "Wilson's disease" 
as a bigram and we would not find, for example, 
8 (Schwarz, 1990) reports a similar strategy. 
2O 
"Wilson's severe disease" as a phrase, though the 
latter might well occur in the general news corpus. 
This example serves to illustrate the essential obser- 
vation that motivates our heuristics for identitying 
lexical atoms in a corpus: (1) words in lexical atoms 
have strong association, and thus tend to co-occur 
as a phrase and (2) when the words in a lexical atom 
co-occur in a noun phrase, they are never or rarely 
separated. 
The detection of lexical atoms, like the parsing 
of simplex noun phrases, is also done in multiple 
phases. At each phase, only two adjacent units 
are considered. So, initiall~ only two-word lexical 
atoms can be detected. But, once a pair is deter- 
mined to be a lexical atom, it will behave exactly 
like a single word in subsequent processing, so, in 
later phases, atoms with more than two words can 
be detected. 
Suppose the pair to test is \[W1, W2\]. The first 
heuristic is implemented by requiring the frequency 
of the pair to be higher than the frequency of any 
other pair that is formed by either word with other 
words in common contexts (within a simplex noun 
phrase). The intuition behind the test is that (1) in 
general, the high frequency of a bigram in a simple 
noun phrase indicates strong association and (2) we 
want to avoid the case where \[W1, W2\] has a high 
frequency, but \[W1, W2, W\] (or \[W, W1, W2\]) has an 
even higher frequency, which implies that W2 (or 
W1) has a stronger association with W than with 
W1 (or W2, respectively). More precisely, we re- 
quire the following: 
F(W~, W2) > Maa:LDF(W~, W2) 
and 
F(W~, W2) > Ma3:RDF(W1, W2) 
Where, 
MaxLDF(W1, W2) = 
Maxw( U in( F(W, W1), DF(W, W2))) 
and 
MaxRDF(W1, W2) = 
Maxw( U in( DF(W1, W), F(W2, W) ) ) 
W is any context word in a noun phrase and F(X, Y) 
and DF(X, Y) are the continuous and discontin- 
uous frequencies of \[X, Y\], respective135 within a 
simple noun phrase, i.e., the frequency of patterns 
\[...X, Y...\] and patterns \[...X, ..., Y...\], respectively. 
The second heuristic requires that we record all 
cases where two words occur in simplex NPs and 
compare the number of times the words occur as 
a strictly adjacent pair with the number of times 
they are separated. The second heuristic is simply 
implemented by requiring that F(W1, W2) be much 
higher than DF(W1, W2) (where 'higher' is deter- 
mined by some threshold). 
Syntactic category analysis also helps filter out 
impossible lexical atoms and establish the thresh- 
21 
old for passing the second test. Only the follow- 
ing category combinations are allowed for lexical 
atoms: \[noun, noun\], \[noun, lexatom\], \[lexatom, 
noun\], \[adjective, noun\], and \[adjective, lexatom\], 
where "lexatom" is the category for a detected lexi- 
cal atom. For combinations other than \[noun, noun\], 
the threshold for passing the second test is high. 
In practice, the process effectively nominates 
phrases that are true atomic concepts (in a par- 
ticular domain of discourse) or are being used 
so consistently as unit concepts that they can be 
safely taken to be lexical atoms. For example, the 
lexical atoms extracted by this process from the 
CACM corpus (about 1 MB) include "operating 
system", "data structure", "decision table", "data 
base", "real time", "natural language", "on line", 
"least squares", "numerical integration", and "fi- 
nite state automaton", among others. 
3.2 Bottom-Up Association-Based Parsing 
Extended simplex noun-phrase parsing as devel- 
oped in the CLARIT system, which we exploit in our 
process, works in multiple phases. At each phase, 
the corpus is parsed using the most specific (i.e., 
recently created) lexicon of lexical atoms. New lex- 
ical atoms (results) are added to the lexicon and are 
reused as input to start another phase of parsing 
until a complete parse is obtained for all the noun 
phrases. 
The idea of association-based parsing is that by 
grouping words together (based on association) 
many times, we will eventually discover the most 
restrictive (and informative) structure of a noun 
phrase. For example, if we have evidence from the 
corpus that "high performance" is a more reliable 
association and "general purpose" a less reliable 
one, then the noun phrase "general purpose high 
performance computer" (an actual example from 
the CACM corpus) would undergo the following 
grouping process: 
general purpose high performance computer =~ 
general purpose \[high=performance\] computer =~ 
\[general=purpose\] \[high=performance\] computer =~ 
\[general=purpose\] \[\[high=performance\]=computer\] =~ 
\[\[general=purpose\]=\[\[high=performance\]=computer\]\] 
Word pairs are given an association score (S) ac- 
cording to the following rules. Scores provide ev- 
idence for groupings in our parsing process. Note 
that a smaller score means a stronger association. 
1. Lexical atoms are given score 0. This gives the 
highest priority to lexical atoms. 
2. The combination of an adverb with an adjec- 
tive, past participle, or progressive verb is given 
score 0. 
3. Syntactically impossible pairs are given score 
100. This assigns the lowest priority to those 
pairs filtered out by syntactic category analysis. 
The 'impossible' combinations include pairs 
such as \[noun, adjective\], \[noun, adverb\], \[ad- 
jective, adjective\], \[past-participle, adjective\], 
\[past-participle, adverb\], and \[past-participle, 
past-participle\], among others. 
4. Other pairs are scored according to the formu- 
las given in Figure 2. Note the following effects 
of the formulas: 
When /;'(W1,W2) increases, S(W1,W2) de- 
creases; 
When DF(W1, W2) increases, S(Wx, W2) de- 
creases; 
When AvgLDF(W~, W2) or AvgRDF(W~, W2) 
increases, S(W1, W2) increases; and 
When F(Wx)- F(W1,W2) or F(W2)- 
F(W1, W2) increases, S(W1, W2) decreases. 
S(W1 W2)= I+LDF(W,,W2)+RDF(W1,W=) A(W1,W2) XlxF(W1,W2)+DF(W1,W,~) X 
Min(F(W, W1),DF(W,W',)) AvgLDF(Wa, W2) = ~-..,WeLD 
ILD\[ 
5-" Min( F( W2,W),D F( W1,W)) AvgRDF(W1, W2) = ~-..,WCRD 
IRDI 
A(W1, W2 ) = ~ F(W1)+F(W2)--2×F(WI,W2)+X2 
Where 
• F(W) is frequency of word W 
• F(W1, W2) is frequency of adjacent bigram \[W1,W2\] 
(i.e ..... W1 W2 ...) 
• DF(W1, W2) is frequency of discontinuous bigram 
\[W1,W21 (i.e ..... W1...W2...) 
• LD is all left dependents, i.e., 
{W\]min(F(W, Wl), DF(W, W2)) ~ 0} 
• RD is all right dependents, i.e., 
{WJmin( D F(W1, W), F(W2, W) ) ¢ 0} 
• ),1 is the parameter indicating the relative contribu- 
tion of F(W1,W2) to the score (e.g., 5 in the actual 
experiment) 
• A2 is the parameter to control the contribution of 
word frequency (e.g., 1000 in the actual experiment) 
Figure 2: Formulas for Scoring 
The association score (based principally on fre- 
quency) can sometimes be unreliable. For example, 
if the phrase "computer aided design" occurs fre- 
quently in a corpus, "aided design" may be judged 
a good association pair, even though "computer 
aided" might be a better pair. A problem may arise 
when processing a phrase such as "program aided 
design": if "program aided" does not occur fre- 
quently in the corpus and we use frequency as the 
principal statistic, we may (incorrectly) be led to 
parse the phrase as "\[program (aided design)\]". 
One solution to such a problem is to recompute 
the bigram occurrence statistics after making each 
round of preferred associations. Thus, using the ex- 
ample above, if we first make the association "com- 
puter aided" everywhere it occurs, many instances 
of "aided design" will be removed from the corpus. 
Upon recalculation of the (free) bigram statistics, 
"aided design" will be demoted in value and the 
false evidence for "aided design" as a preferred as- 
sociation in some contexts will be eliminated. 
The actual implementation of such a scheme re- 
quires multiple passes over the corpus to generate 
phrases. The first phrases chosen must always be 
the most reliable. To aid us in making such decisions 
we have developed a metric for scoring preferred 
associations in their local NP contexts. 
To establish a preference metric, we use two statis- 
tics: (1) the frequency of the pair in the corpus, 
F(W1, W2), and (2) the number of the times that 
the pair is locally dominant in any NP in which the 
pair occurs. A pair is locally dominant in an NP 
iff it has a higher association score than either of 
the pairs that can be formed from contiguous other 
words in the NP. For example, in an NP with the se- 
quence \[X, Y, g\], we compare S(X, Y) with S(Y, g); 
whichever is higher is locally dominant. The prefer- 
ence score (PS) for a pair is determined by the ratio 
of its local dominance count (LDC)--the total num- 
ber of cases in which the pair is locally dominant--to 
its frequency: 
LDC(WI 1W2) PS(W1, W2) = r(Wl,W~) 
By definition all two-word NIPs score their pairs 
as locally dominant. 
In general, in each processing phase we make only 
those associations in the corpus where a pair's PS 
is above a specified threshold. If more than one as- 
sociation is possible (above theshold) in a particular 
NP, we make all possible associations, but in order 
of PS: the first grouping goes to the pair with high- 
est PS, and so on. In practice, we have used 0.7 as 
the threshold for most processing phases. 9 
4 Experiment 
We tested the phrase extraction system (PES) by us- 
ing it to index documents in an actual retrieval task. 
In particular, we substituted the PES for the default 
NLP module in the CLARIT system and then in- 
dexed a large corpus using the terms nominated by 
the PES, essentially the extracted small compounds 
and single words (but not words within a lexi- 
cal atom). All other normal CLARIT processing-- 
weighting of terms, division of documents into 
subdocuments (passages), vector-space modeling, 
etc.--was used in its default mode. As a baseline 
°When the phrase data becomes sparse, e.g., after six 
or seven iterations of processing, it is desirable to reduce 
the threshold. 
22 
for comparison, we used standard CLARIT process- 
ing of the same corpus, with the NLP module set to 
return full NPs and their contained words (and no 
further subphrase analysis).l 0 
The corpus used is a 240-megabyte collection 
of Associated Press newswire stories from 1989 
(AP89), taken from the set of TREC corpora. There 
are about 3-million simplex NPs in the corpus and 
about 1.5-million complex NPs. For evaluation, 
we used TREC queries 51-100, ll each of which 
is a relatively long description of an information 
need. Queries were processed by the PES and nor- 
mal CLARIT NLP modules, respectively, to gener- 
ate query terms, which were then used for CLARIT 
retrieval. 
To quantify the effects of PES processing, we used 
the standard IR evaluation measures of recall and 
precision. Recall measures how many of the rele- 
vant documents have been actually retrieved. Pre- 
cision measures how many of the retrieved docu- 
ments are indeed relevant. For example, if the total 
number of relevant documents is N and the system 
returns M documents of which K are relevant, then, 
Recall = K IV 
and 
Precision = ~-. 
We used the judged-relevant documents from the 
TREC evaluations as the gold standard in scoring 
the performance of the two processes. 
suggests that the PES could be used to support other 
IR enhancements, such as automatic feedback of the 
top-returned documents to expand the initial query 
for a second retrieval step) 2 
CLARIT Retrieved-Rel Total-Rel Recall 
Baseline 2,668 3,304 80.8% 
PES 2,695 3,304 81.6% 
Table 1: Recall Results 
Baseline Rel.Improvement 
0.6819 4% 
Recall PES 
0.00 0.7099 
0.10 0.5535 0.5730 
0.20 0.4626 0.4927 
0.30 0.4098 0.4329 
0.40 0.3524 0.3782 
0.50 0.3289 0.3317 
0.60 0.2999 0.3026 
0.70 0.2481 0.2458 
0.80 0.1860 0.1966 
0.90 0.1190 0.1448 
1.00 0.0688 0.0653 
3.5% 
6.5% 
5.6% 
7.0% 
0.5% 
0.9% 
--0.9% 
5.7% 
21.7% 
-5.0% 
Table 2: Interpolated Precision Results 
5 Results 
The results of the experiment are given in Tables 1, 
2, and 3. In general, we see improvement in both 
recall and precision. 
Recall improves slightly (about 1%), as shown in 
Table 1. While the actual improvement is not sig- 
nificant for the run of fifty queries, the increase in 
absolute numbers of relevant documents returned 
indicates that the small compounds supported bet- 
ter matches in some cases. 
Interpolated precision improves significantly5 as 
shown in Table 2. The general improvement in 
precision indicates that small compounds provide 
more accurate (and effective) indexing terms than 
full NPs. 
Precision improves at various returned-docu- 
ment levels, as well, as shown in Table 3. Initial 
precision, in particular, improves significantly. This 
1°Note that the CLARIT process used as a baseline does 
not reflect optimum CLARIT performance, e.g., as ob- 
tained in actual TREC evaluations, since we did not use a 
variety of standard CLARIT techniques that significantly 
improve performance, such as automatic query expan- 
sion, distractor space generation, subterm indexing, or 
differential query-term weighting. Cf. (Evans et al., 1996) 
for details. 
1 ~ (Harman, 1993) 
Do, c-Level 
5 docs 
10 docs 
15 docs 
20 docs 
30 docs 
100 docs 
200 docs 
500 docs 
1000 docs 
Baseline PES Rel.Improvement 
0.4255 0.4809 13% 
0.4170 0.4426 6% 
0.3943 0.4227 7% 
0.3819 0.3957 4% 
0.3539 0.3603 2% 
0.2526 0.2553 1% 
0.1770 0.1844 4% 
0.0973 0.0994 2% 
0.0568 0.0573 1% 
Table 3: Precision at Various Document Levels 
The PES, which was not optimized for pro- 
cessing, required approximately 3.5 hours per 20- 
megabyte subset of AP89 on a 133-MHz DEC alpha 
processor) 3 Most processing time (more than 2 of 
every 3.5 hours) was spent on simplex NP parsing. 
Such speed might be acceptable in some, smaller- 
scale IR applications, but it is considerably slower 
than the baseline speed of CLARIT noun-phrase 
identification (viz., 200 megabytes per hour on a 
100-MIPS processor). 
l~ (Evans et al., 1995; Evans et al., 1996) 
13Note that the machine was not dedicated to the PES 
processing; other processes were running simultaneously. 
23 
6 Conclusions 
The notion of association-based parsing dates at 
least from (Marcus, 1980) and has been explored 
again recently by a number of researchers. TM The 
method we have developed differs from previous 
work in that it uses linguistic heuristics and local- 
ity scoring along with corpus statistics to generate 
phrase associations. 
The experiment contrasting the PES with baseline 
processing in a commercial IR system demonstrates 
a direct, positive effect of the use of lexical atoms, 
subphrases, and other pharase associations across 
simplex NPs. We believe the use of N-P-substructure 
analysis can lead to more effective information man- 
agement, including more precise IR, text summa- 
rization, and concept clustering. Our future work 
will explore such applications of the techniques we 
have described in this paper. 
7 Acknowledgements 
We received helpful comments from Bob Carpen- 
ter, Christopher Manning, Xiang Tong, and Steve 
Handerson, who also provided us with a hash table 
manager that made the implementation easier. The 
evaluation of the experimental results would have 
been impossible without the help of Robert Lefferts 
and Nata~a Mili4-Frayling at CLARITECH Corpo- 
ration. Finally, we thank the anonymous reviewers 
for their useful comments. 

References 
David A. Evans. 1990. Concept management in text via 
natural-language processing: The CLARIT approach. 
In: Working Notes of the 1990 AAAI Symposium on "Text- 
Based Intelligent Systems", Stanford University, March, 
27-29, 1990, 93-95. 
David A. Evans, Kimberly Ginther-Webster, Mary Hart, 
Robert G. Lefferts, Ira A. Monarch. 1991. Automatic 
indexing using selective NLP and first-order thesauri. 
In: A. Lichnerowicz (ed.), Intelligent Text and Image Han- 
dling. Proceedings of a Conference, RIAO "91. Amsterdam, 
NL: Elsevier, pp. 624-644. 
David A. Evans, Robert G. Lefferts, Gregory Grefenstette, 
Steven K. Handerson, William R. Hersh, and Armar A. 
Archbold. 1993. CLARIT TREC design, experiments, 
and results. In: Donna K. Harman (ed.), The First Text 
REtrieval Conference (TREC-1). NIST Special Publication 
500-207. Washington, DC: U.S. Government Printing 
Office, pp. 251-286; 494-501. 
David A. Evans, and Robert G. Lefferts. 1995. CLARIT- 
TREC experiments Information Processing and Manage- 
ment, Vol. 31, No. 3, 385-395. 
David A. Evans, Nata~a Mili4-Frayling, Robert G. Lef- 
ferts. 1996. CLARIT TREC-4 experiments. In: Donna 
14 (Liberman et al., 1992; Pustejovsky et al., 1993; Resnik 
et al., 1993; Lauer, 1995) 
K. Harman (ed.), The Fourth Text REtrieval Conference 
(TREC-4). NIST Special Publication. Washington, DC: 
U.S. Government Printing Office. 
Donna K. Harman, ed. 1993. The First Text REtrieval 
Conference (TREC-1) NIST Special Publication 500-207. 
Washington, DC: U.S. Government Printing Office. 
Donna K. Harman, ed. 1995. Overview of the Third Text 
REtrieval Conference (TREC-3 ), NIST Special Publication 
500-225. Washington, DC: U.S. Government Printing 
Office. 
Donna K. Harman, ed. 1996. Overview of the Fourth Text 
REtrieval Conference (TREC-4), NIST Special Publica- 
tion. Washington, DC: U.S. Government Printing Of- 
fice. 
Mark Lauer. 1995. Corpus statistics meet with the noun 
compound: Some empirical results. In: Proceedings of 
the 33th Annual Meeting of the Association for Computa- 
tional Linguistics. 
David Lewis and K. Sparck Jones. 1996. Natural language 
processing for information retrieval. Communications of 
the ACM, January, Vol. 39, No. 1, 92-101. 
Mark Liberman and Richard Sproat. 1992. The stress and 
structure of modified noun phrases in English. In: I. 
Sag and A. Szabolcsi (eds.), Lexical Matters, CSLI Lec- 
ture Notes No. 24. Chicago, IL: University of Chicago 
Press, pp. 131-181. 
Mitchell Maucus. 1980. A Theory of Syntactic Recognition 
for Natural Language. Cambridge, MA: MIT Press. 
J. Pustejovsky, S. Bergler, and P. Anick. 1993. Lexical 
semantic techniques for corpus analysis. In: Compu- 
tational Linguistics, Vol. 19(2), Special Issue on Using 
Large Corpora II, pp. 331-358. 
P. Resnik, and M. Hearst. 1993. Structural Ambiguity and 
Conceptual Relations. In: Proceedings of the Workshop on 
Very Large Corpora: Academic and Industrial Perspectives, 
June 22, Ohio State University, pp. 58-64. 
Gerard Salton and Michael McGill. 1983. Introduction to 
Modern Information Retrieval, New York, NY: McGraw- 
Hill. 
Christoph Schwarz. 1990. Content based text handling. 
Information Processing and Management, Vol. 26(2), 
pp. 219-226. 
Alan F. Smeaton. 1992. Progress in application of natu- 
ral language processing to information retrieval. The 
Computer Journal, Vol. 35, No. 3, pp. 268-278. 
T. Strzalkowski and J. Carballo. 1994. Recent develop- 
ments in natural language text retrieval. In: Donna 
K. Harman (ed.), The Second Text REtrieval Conference 
(TREC-2). NIST Special Publication 500-215. Washing- 
ton, DC: U.S. Government Printing Office, pp. 123-136. 
