The Use of Focus as a Tool for 
Disambiguation of Definite Noun Phrases 
by 
Candace L. Sidner 
M|T Artificial Intelligence Laboratory 
Cambridge, Massachusetts 
1. Introduction to the Problem 
When speakers I utter or write sentences, they use 
certain words in the sentence to refer to people, places, 
object, times, events and ideas which exist in the real world. 
When sentences are formed into units of two or more 
sentences, certain words refer back to other referring 
expressions in' the previous sentences. Among the words 
which can be used to refer to the real world as well as to 
refer back (which is called the process of co-reference) are 
noun phrases containing a definite article, such as the (called 
defnps, hereafter). Several aspects of comprehension of 
defnps are open problems: 
1.) What is the definition of the reference of a defnp? 
That is, what to we mean by reference in 
computational linguisitics? 
2.) How are defnps which are used to co-refer into a 
discourse distinguished from those which refer to 
real world objects outside the discourse? 
3.) What methods of search will distinguish the referent 
of a defnp which refers to an entity outside the 
discourse context? 
4.) What different wayscan defnps be used to co-refer 
to other entities in the discourse? 
5.) How can co-reference of defnps be detected? 
6.) What inferences and data structures will be needed 
for the detection process? The work of Winograd 
\[1971\], Charniak \[1972\] and Rieger \[1973\] suggest 
that inferencing is crucial to the interpretation of 
co-reference. 
This paper presents a viewpoint from which to 
answer these questions based on the concept of focus, as 
developed by Grosz \[1977\] and the author (Bullwinkle \[1977\]). 
This paper extends' Grosz' work by establishing a framework 
for communication and a set of rules for use of focus in 
discourse. The approach taken here represents an alternative 
to the inference driven schemes of Charniak and Rieger. 
2. The Communication Process and Focus 
The description of the communication process given 
here contains four simple assumptions which are generally true 
and will be taken as true in this work. First, the speaker is 
assumed to be communicating about something. This 
assumption implies that the speaker is not speaking gibberish, 
that there are referring expressions and either requests, 
questions, assertions or acknowledgements being made. The 
something which the communication is about will be called the 
focus of the discourse. 2 Second, the hearer is assumed to be 
able to identify what the focus of the discourse is. The 
speaker wants to communicate about something, and for the 
communication to occur, the hearer must be able to distinguish 
what the speaker is communicating about. Third, the speaker 
is not trying to confuse or deceive the hearer. The speaker 
uses referring expressions with the intention of referring to 
someone or something, or with the intention of describing 
something or some event. In Gricean 3 terms, the byword is 
"Be perspicuous." The final assumption claims that the speaker 
assumes the hearer knows certain knowledge about the 
real-world which can be referred to during the communication 
process, Recent research (Cohen \[1978\]), as well as the well 
known work of Searle \[1969\] and Austin \[\]962\], describe 
models of the speaker's knowledge of what the hearer 
believes. In this chapter, the weakest form of such a model is 
assumed: the speaker assumes the hearer has enough 
real-world knowledge in common with the speaker to know the 
entities which the speaker refers to, and that knowledge is 
what the speaker draws on in constructing a message for a 
hearer. These four assumptions will play an important part in 
the discussion of co-reference interpretation which follows. 
1. I will use the term speaker to refer to the producer of a 
spoken or written discourse and hearer to refer to the 
receiver of the discourse. 
2. ! don't want |o suggest that only one thing can be 
communicated in a discourse, for speakers do direct their 
attention from one thing to another. For the moment, I will 
speak of the focus as the first center of attention in a 
speaker's discourse. 
3. Grice, H.P. "Logic and Conversatron" etc. 
86 
This paper makes the claim that the focus acts as an 
index function for referring expressions. For those referring 
expressions which are anaph0ric, the focus indicates where to 
look for an antecendent. For those referring expressions 
which are names or descriptions of things in the world, the 
focus acts as a generation center for a process that chooses a 
representation of a real world entity which fits the name or 
description. However, the focus of a discourse alone is not 
sufficient to produce the indexing behavior. The focus must be 
used in conjunction with a hierarchical semantic network of 
associations. The network will indicate what other concepts 
are related to the focus. It is a codification of some of the 
general knowledge speakers and hearers have about the real 
world. The network is a dynamic structure because the hearer 
adds to his/her general knowledge in the process of 
interpreting a piece of discourse. Focus must also act with a 
third piece of computational machinery, an inferencing 
mechanism. It is used to infer from general knowledge and 
some suppositions that a certain proposition is true. 
An example will be helpful here. In the discourse 
below, the focus of discussion is the meeting of D0-1. 
DO-1 I want to schedule a meeting with Ira. 
2 It should be at 3 p.m. 
3 We can get together in his office. 
4 \]nvite John to come, too. 
All four sentences give information about the focussed entity. 
Thus in DO both sentences 3 and Zl make no direct reference 
to the meeting of D0-t. As human • hearers, we know that 
these sentences are related to the rest of DO because they 
give information about the focus meeting. \]n DO-3 there are 
three clues which connect this sentence and the rest of the 
discourse: the use of get together, the co-reference of we to 
the participants of the meeting, and his office establishing a 
place for a meeting. DO-4 introduces an additional participant • 
which can be surmised from the use of invite, and the fact that 
the ellipsis of the event that John is invited to is the focus. 
A piece of the hierarchical net needed for DO is 
• given below in figure !. A prototype meetings has associated 
places, times, pai'ticipants, and purposes. The relation 
between meeting and place is one of occurrence while the 
relation between meeting and purposes is one of causality. 
When DO-\] is encountered, the hierarchical net grows a new 
member: an instance of meeting from D0-1. It inherits the 
associated entities of meetings and some specific values for 
the participant entity. 1:)0-2 indicates that something (called it) 
will Occur at a particular time; The focus of D0-1 is meeting, 
so the focus, meeting, is proposed as the antecedent of it. To 
Fig. 1. Instances of a General Meeting Concept 
j meeting 
place ome par "" t. . i. icipant purpose 
meeting-with-St anoczyk I I "..~ 
u" / 
place: 801 ~/. 
time: Thursday-at:3 participants: Stanoczyk, Lewin 
confirm the proposal, the inference mechanism checks to see if 
meetings occur at times. Indeed they do, so the proposal of 
meeting as antecedent of it is accepted. 
The explanation about the use of focus is not really 
so simple because the focus of a discourse changes. The 
interpretation of focus requires a means of recognizing that 
the focus has changed to some other entity. In DO' the focus 
begins on meeting, but the it in D0'-3 has my office as its 
co-referent, not the meeting. Detecting this co-reference 
requires a means of noticing a shift of focus and using the 
inferencing mechanism to confirm the choice of co-referent. 
Focus shift detection will not be discussed here; the reader is 
referred to Bullwinkle \[1977\] for a discussion of focus shift 
where the term =sub-topic shift" is used. 
DO'-\] ! Want to schedule a meeting with George, Jim, 
Steve and Mike. 
2 We can meet in my office. 
3 It's kind of small, but the meeting won't last very 
long anyway. 
3. Reference in Computational Terms 
The theory presented here distinguishes two kinds 
of referring. The first is an internal reference between a noun 
phrase and some pre-existing database object. That database 
object represents a real world entity, in Figure 2 below 
internal reference links the noun phrase NP1 "Jimmy Carter" to 
a representation of Jimmy Carter (who is described as 
president of the US, etc.).. How the noun phrase and the 
database object refer to the real world is the classical 
semantic problem of reference (cf. Kripke \[1972\] among 
others) and is beyond the scope of this work. The other kind 
of referring is co-reference. Co-reference links a noun phrase 
to another noun phrase. The two noun phrases are said to 
co-refer, and both internally refer to the same database 
object, both refer to the real world object. In Figure 1, the 
dashed link from NP2 "Jimmy" to NPI is a co-reference link. 
The dot-dash link from NP2 to the database object is a virtual 
internal reference link which results from the co-reference link 
from NP2 to NP! and from the internal reference link from NPI 
• to the database object. Internal reference and co-reference 
links are distinguished because co-reference links can be 
established more easily using discourse context, which will be 
discussed in detail later in this paper. In the remainder of 
this paper when \[ speak of internal reference, \] will drop the 
phrase "internal" and use only "reference." 
A computational theory of co-reference 
comprehension must answer the following questions about the 
use of referential terms in natural language: 
(1) Does the expression refer to someone or something? 
(2) What conceptual entity in the memory or the 
database of the hearer's knowledge, if any, is 
denoted by the referring expression? 
(3) When does a given expression refer to the same 
entity as another referring expression? 
The expression Julius Caesar is used to refer, and can refer to 
the person represented in the heater's knowledge as Julius 
87 
Fil~. 2. Reference Links Between Noun Phrases 
/ co-reference 
NP\] "Jimmy Carter" NP2 "Jimmy" 
rnal virtual ~ 
erence internal reference ~ ,/ 
ld 
Database Representation of Jimmy Carter 
Name: Jimmy Carter 
occupation: President of US 
birthplace: Georgia 
Caesar. To answer the first question above, the hearer must 
decide that names are referring expressions. To answer the 
second question, the hearer must decide 1) whether Julius 
Caesar refers uniquely and 2) what conceptual entity in the 
bearer's memory represents the hearer's real-world referent. 
These two decisions together with the initial assumptions 
appear to make necessary and sufficient conditions for 
comprehension since by deciding that Julius Caesar refers 
uniquely and choosing a conceptual entity, the hearer has 
decoded what entity the speaker was referring to. 
There are, however, situations where the hearer's 
choices to the above decisions and the speaker's intended 
referent do not coincide. Suppose the hearer decides that 
Julius Caesar refers uniquely and refers to Julius Caesar, who 
was a Roman emperor. The speaker may also have intended it 
to refer uniquely, but to the author's deceased cat, whose 
name was Julius Caesar. Now there are three possibilities: 
either the hearer knew about Julius Caesar the cat, but 
decided the expression referred to Julius the emperor, or the 
hearer only knew about the emperor, or the hearer didn't 
know of either. In the last case, the hearer "found" a referent 
by a chance from randomly linking up the name and some 
memory representation. The last possibility does not fit a 
description of reference comprehension of any kind. Randomly 
hooking up information from one's memory to what appears to 
be a referring expression may be a cognitive act, but 
intuitively no one would call it reference comprehension. 
In the case wh~re the hearer only knew about the 
emperor, it seems safe to conclude that the reference may 
have been comprehended, but incompletely. As we shall see, 
there are many other clues in communication about the 
referent of terms than those given by referring expressions in 
isolation. Without these, reference comprehension is 
incomplete because the hearer has no means of knowing 
whether s/he may have the wrong referent. Even with the 
best set of clues, the hearer may still choose Julius the 
emperor. Here we will say that comprehension has taken 
place, completely but incorrectly, because the hearer has used 
all the relevant communication knowledge to decode the 
speaker's message. What can be concluded is that the 
speaker's rules for reference generation and/or the speaker's 
knowledge of the hearer is faulty (thereby contradicting the 
speaker's assumption above). 
In the case where the hearer knows of both 
possibilities and chooses the incorrect one, the hearer may 
have erred due to failure to follow other communication clues 
or again because the speaker's rules and knowledge were 
lacking. In conclusion, a referring expression is comprehended 
as intended, if and only if the same referent as that intended 
by the speaker is chosen from the entities in memory. The 
expression is otherwise just comprehended when the hearer 
chooses an entity from memory which is denoted by the 
referring expression using all the available communication 
clues but does not choose the same entity as intended by the 
speaker. An expression will be-considered incompletely 
comprehended if the hearer fails to use all the 'communication 
cues available at the time the communication occurs. 
So far ! have not considered the possibility of error 
on the part of the hearer because of the heater's beliefs. 
Suppose, for example, that the hearer believes the speaker 
hates to even speak of cats. Then the hearer may conclude 
that Julius Caesar is most likely a reference to the emperor of 
Rome. \] am not going to consider this possibility in the 
forthcoming discussion; instead I will restrict the discussion to 
cues from the communication process. Hearer beliefs raise a 
separate set of philosophical as well as computational 
problems and entends the scope of this study too broadly. 
However, the issues are significant in the total picture of 
reference and co-reference comprehension. 
In the remainder of this paper \[ will consider 
co-reference comprehension just from the hearer's point of 
view. Thus it1 discussing referential and co-referential 
expressions, \] wifl be concerned with a model of how the 
hearer disambiguates these expressions used in discourse. By 
symmetry, one might suppose that the generation of 
referential expressions by a speaker could make use of a 
similar model. Such a supposition will remain untested in this 
paper and is to be verified by later work. Furthermore, \] will 
not be concerned with comprehension as intended since this 
process requires the additional information of what the hearer 
believes that the speaker knows about, instead ! will point 
out at various times how the theory under discussion would 
need modification if hearer's beliefs were included. 
4. Problems with Definite Noun Phrases 
Definite noun phrases can be used to refer to 
entities in the real world. Russell \[1905\] says of the 
expression the author of Waverly that it denotes Sir Walter 
Scott, and that when it is strictly used, a defnp denotes 
uniquely. 4 Thus by us!ng a definite article, a speaker is saying 
in effect "there is one object in the world denoted by the 
phrase that follows and \] mean that one." Of course a defnp 
may be used to denote someone without actually denoting 
anyone, as is the case with the woman who wrote Waverly. 
This defnp is used to refer to someone, but there is no 
conceptual representation in the hearer's (or for that matter, 
the speaker's) memory which corresponds to a real world 
4. By strictly used Russell means used without ambiguity. 
88 
entity assuming the normal case. 5 Nothing in the syntactic or 
semantic form of the expression itself suggests that the 
expression has no denotation. How can the hearer determine 
whether the defnp refers to someone or not? Of course, if 
there exists a memory entity the author of Waverly, which is 
attributed as male, the hearer can decide that the expression 
does not refer to anyone on the basis of a contradiction. But 
if no memory entity exists, the hearer cannot decide whether 
the woman who wrote Waverly refers to anyone. This way of 
looking at defnps, however, fails to account for all the 
phenomena of defnps because it involves an assumption which 
is not true. 
The Russell!an analysis has difficulties because 
defnps are not always used to refer! The problem is not only 
whether a particular defnp actually denotes a real world 
object; it is also a question of whether the defnps is intended 
to refer at all. Even more surprising, a defnp may be used to 
refer, but the speaker may not intend for the hearer to know 
the referent of the defnp; the defnp form is used to indicate 
that the referent is knowable, but possibly not significant for 
the communication at hand. Donnel~an \[1977\] points out that 
some defnps are used attributively, if we happen upon Smith 
who lies dead with foul wounds, one can say "Smith's murderer 
is insane." Used attributively, Smith's murderer does not refer 
to anyone, and the phrase does not describe a particular 
person, it is as if to say, Smith was murdered and the 
murderer, whoever that may be, is insane. Thus the speaker 
using an attributive defnp does not assume that someone fits 
the description, whereas with a referential defnp the speaker 
expects the hearer to realize who is being pointed to. 
The other distinction a speaker can make is to use a 
defnp to indicate that the referent is knowable. Thus if one 
says: 
(1) Larry read a lot of linguistics in the hospital. 
(2) Larry read a lot of linguistics in a hospital. 
the (2) usage is not the same as the (1). While the hearer 
does not know which hospital the hospital refers to, it is clear 
it refers to some particular one. Comprehension of the 
referential term does not involve finding a memory entity 
which represents the real world entity that the expression 
refers to. For reference comprehension, this concern is 
considerable since the Chinese government in (3) does not 
demand reference disambiguation, while (4) does; the 
disambiguation is difficult because the expression can refer to 
more than one thing. 
(3) John got help from the Chinese government in 
adopting an Oriental child. 
(4) Get a visa for your trip from the Chinese 
government. 
Another difficulty with defnps is that sometimes they 
are used not to refer to or to describe specific individuals or 
objects, but to characterize a class of entities with the 
5. Possible world semantics will not be discussed here. 
Issues of transworld identity and designation by definite 
descriptions may require more machinery than is considered 
here. 
properties of th e head noun phrase and any of its modifiers. 
Thus (el) used in this way does not refer to an individual. It 
characterizes a member of the class of individuals who are 
men and book writers. (el) is similar to attribution except that 
the description applies to a class. 
(el) the man who writes books 
So far, then, the following classes of defnps can be stated. A 
defnp that is used to refer uniquely to one entity, whether or 
not such an entity exists in the real world, is a specific defnp. 
A defnp that characterizes a class of entities by means of an 
individual whose properties are delineated by the properties 
of the head noun phrase and its modifiers is a generic defnp. 
A defnp is attributive if it describes an entity without 
referring. A defnp can be ambiguous in use (u-ambiguous 
hereafter) if its use as a specific, attributive or generic is not 
identifiable, while a defnp is ambiguous in reference 
(r-ambiguous hereafter) if it is used specifically and there is 
more than one entity fitting the description of the defnp. 
5. General Role of Context in Disambiguation 
Little of language, if any at all, is said without some 
surrounding contexts of information. For example, most 
conversations happen in a location where there are other 
objects present. Most stories have at least the context of 
there being a story teller, a hearer and the story being told. 
There are contexts, with more presumed common knowledge, 
such as what the hearer knows of the speaker's own identity 
or some shared additional information between them (e.g. they 
have children or parents in common). 
Contexts are needed to determine what a defnp 
refers to. If \] say (5), when I am standing in my kitchen with a 
friend, the defnp, if specific, must refer to some unique object 
in the world. 
(5) Get me the hot dish holder. 
There may be lots of hot dish holders denoted by conceptual 
entities in my friend's mind, but ! am referring to a specific 
one. Since nothing in (.5) distinguishes the one ! mean from 
the whole collection, either | have misused the language, or 
there is a context which contains only one such hot dish 
holder, and my friend is aware of that context at the time of 
my saying (5). In this case, the necessary context is the 
kitchen, and the referent is probably an item in the kitchen. 
Reference made to an object external to the conversion in 
called extra-sentential reference. It is discussed here to 
exemplify the role of one class of contexts used in reference 
determination. \] call contexts of reference which exist in 
additional to the one created by the discourse implicit 
contexts. In this paper I will show how use of implicit 
contexts can avoid the problem of searching a general 
database for the entities denoted by defnps. 
Other defnps make use of different implicit contexts. 
Instead of an implicit context consisting of objects near the 
speaker, the implicit context may be events that the speaker 
believes are common to the hearer. The speaker who opens a 
dialogue with (6) below is assuming some previous context (a 
discussion with the hearer or some other situation) where the 
reference of the A.L Lab Language Group was first established. 
In (7), the speaker is again assuming a pro-established 
referent, but since the hearer may know of several different 
89 
dogs, some specific context must be chosen that will 
distinguish a single dog. Later in this paper some heurisitics 
for choosing a context will be discussed. 
(6) The A.L lab Language Group wants to meet next week. 
(7) The dog is sick again. 
Contextual information of yet another kind appears 
in story telling. At the beginning of a story, the hearer 
expects characters to be introduced. Sometimes this is done 
with indefinite noun phrases, which are a way all discourses 
introduce new items, but often a story-teller uses names or 
defnps as (8) below shows. 
(8) The heiress lived the life Of a recluse. She died 
under mysterious circumstances, but the murderer 
was never found. 
/ 
(8) is not a case of cataphoric referencing (referring forward 
in a text) since the phrase the heiress can fully specify an 
object itself. However, hearers of (8) do not have to search 
their memories for a referent to the heiress in (8). They use 
the context of story beginning to guide them in reference 
disambiguation. 
6. Distinguishing Generics 
Defnps must be disambiguated as generic or 
non-generic. As will be shown later, generics in the midst of a 
discourse can be easily disambiguated, but in an initial 
sentence only implicit contexts may exist in which a 
co-referent can be found to use in disambiguation. Implicit 
contexts may be helpful in some cases, but in general they are 
not sufficient to indicate the int~erpretation. However certain 
rules can be postulated based on observable sentence data. 
Thisdata indicates that there are several levels of sentential 
and phrasal information used for disambiguation. The rules 
which will be summarized 6 here give preferences for generic 
and non-generic readings. 
Some rules govern whether the defnp itself is 
preferred as a generic or non-generic reading. A small 
collection of phrases like the sun, the moon and the president 
default to specific, well defined entities. Which entities are 
defaulted to depends on the presence or absence of an implicit 
context to which the phrase may co-refer. For other defnps, a 
"yes" answer to (1) of the following questions indicates a 
specific reading preference, while a "yes" to (2) and (3) 
indicates a generic: 
1.) Is there a specific individual so described? 
2.) Is there a class use acceptable for this pair? 
3.) Are there many individuals described but none 
outstanding? 
"No" answers to all causes a preferred specific reading. 
Question (2) is necessary because generic readings are 
difficult to obtain for certain classes of entities. Thus color 
words like black, red, yellow and white applied to man 
describe a class specification while the other color words 
indicate a non-generic description. 
6. See Sidner \[forthcoming\] for a full specification of the 
interpretation rules. 
in addition to phrasal preferences, predicate 
argument relations for certain verbs may indicate a preferred 
non-generic reading. In these cases, a u-ambiguous defnp will 
be taken as a non-generic, and a generic defnp will cause an 
odd sentence. 
(9) The black man Was moving towards the window. 
(10) The woman who reads Total Woman is coming to 
dinner on Saturday. 
Some classes of speech acts are also distinctly generic. The 
is-a sentence below is always generic; this reading may result 
from the use of is-a to indicate further characteristics of the 
subject. The announcement speech act in (12) is generic as 
long as an implicit context does not exist which contains an 
acceptable co-reference for the defnp. 
(11) The elephant is a large mammal. 
(12) I want to tell you about the orangetang. 
Speaker-hearer assumptions about perspicuity can 
force a reading to be generic or non-generic. The defnp in 
(13) is forced to be read generically because a specific reading 
would be r-ambiguous and therefore not perspicuous. On the 
other hand, (14) is odd since invite requires a non-generic 
object. However, because the speaker is assumed to be 
maintaining perspicuity, the hearer may attempt to read the 
defnp in (14) as a non-generic. 
(13) Bill considers the black man to be the source of 
Boston's social unrest. 
(14) Invite the man who readsThe Grapes of Wrath to 
dinner. 
I want to emphasize that the noun phrase, verb phrase and 
sentence level rules are only preferences for readings. |n the 
worst case, as (15) shows, an initial sentence may contain a 
u-ambiguous phrase which, while preferred as generic, can be 
used either way. 
(15) The robot is replacing the car. 
7, The Explicit Backwards Co-reference Rule for Defnp 
Disambiguation 
Using the concept of focus, rules governing the 
co-referential use of defnps in discourse can be stated. The 
rules for defnp co'reference which follow depend Upon the 
ability of the hearer to identify focus. This process is a 
complex one and will not be discussed here. The reader is 
referred to Sidner \[forthcoming\] for full details. In brief, the 
focus of a sentence depends upon predicate argument 
relations and in some cases, special syntactic forms, such as 
cfefts and pseudoclefts. 
In the simplest formulation, the rules for defnp 
co-reference states: the discourse focus provides a reference 
point for the co-references of defnps. As \[ will show below, 
the rule contains several subparts which must be stated 
separately. In this paper \] will refer to cases of a defnp used 
anaphorically as explicit backwards co-referencing (EBC). The 
EBC rule states that a defnp with the same noun phrase head 
as the focus, and which appears in a sentence following the 
sentence with the focus, is co-referential to the focus. The 
more common forms of explicit backwards co-referencing are 
found in D\] and D2 below: 
9O 
D1-1 I want to have a_ big partYi with lots of guests. 
2 Th__ee party! ought to be on Saturday so everyone 
can come. 
D2-1 I'm going to tell you about the elephantj. 
2 The elephantj is the largest of the jungle 
mammals. 
3 Hej weighs over 3000 pounds. 
4 ~one point in itsj existence, the elephantj had to 
protect itself from-~e lion, 
5 but now its: herds are so large, that most lions 
won't even venture near. 
What the reader will notice about D2 is not only the 
co-referenciality of focus for the second and third uses of the 
elephant, but also the co-referenciality implies that these uses 
are generic. Where defnps in isolation are often ambiguous on 
the generic-specific classification, in discourse context, this 
rarely occurs since the focus provides the class type for the 
defnp. As stated, the EBC rule makes a true prediction about 
u-ambiguous defnps which occur in sentences following the 
focus: they are c¢)-referential with the focus, and hence 
disambiguated as non-generic. D3-2 below contains a defnp 
which is u-ambiguous in isolation, but in the discourse context, 
it refers to George's elephant, the reference of the focus. 
D3-1 I sent George an elephant last year for a birthday 
present. 
2 Th.__ee elephant likes potatoes for breakfast. 
The EBC rule is inaccurate when applied to strictly 
generic defnps, and where it fails, the role of phrasal and 
sentential level processing in co-reference comprehension is 
indicated. D4 is an indication of the problem. D4-2 is generic 
in isolation. Even in the context of D4, where the focus is 
Mary's ferret, hearers interpret the underlined defnp as 
generic. 
D4-1 Mary got a ferret for Christmas last year. 
2 The ferret is a very rare animal. 
The context cues of discourse are not strong enough to 
reverse a strongly generic reading of a defnp. In order for 
this to be so, sentential level processing must have occurred 
without consideration of the demands of the context. Since the 
EBC rule as stated predicts co-reference in cases like D4, it 
must be revised: specific and u-ambiguous defnps which 
contain the same noun phrase head as the focus, and which 
follow the focus in the discourse, co-refer with the focus. 
A further refinement on the EBC rule is needed. 
Consider the fairy tale book in 175-2. The EBC rules predicts it 
will be co-referential with the focus of book in DS-1. In fact, 
English speakers find D5-2 an odd sentence in the discourse 
because it is not clear what the fairy tale book has to do with 
the rest of DS. 
DS-I I bought a book today. 
2 The fairy tale book is by the Brothers Grimm. 
3 It is really well illustrated. 
It seems that defnps which co-refer with the focus cannot 
contain anymore information than is known abo~Jt the focus. 
Thus one could say following 175-1, "The book ! bought is a 
fairy tale book by the Brothers Grimm" (since 175-\] states that 
the speaker bought the book), but one cannot say D5-2. Why 
can't a defnp that contains more information than the focus 
co-refer to the focus? Returning to the discussion of 
focus-shift earlier, a referring expression following the focus 
is either co-referential to the focus or introduces an entity 
which is the potential new focus of the discourse. The 
difficulty with phrases like the fairy tale book is that one 
cannot tell if it is intended to co-refer, or because it is 
somewhat different from the focus, intended to be used as a 
potential new focus. The EBC rule must be revised to state: 
specific and u-ambiguous defnps which contain the same noun 
phrase head as the focus, which follow the focus in the 
discourse, and which do not contain more information than is 
known about the focus co-refer with the focus. 
The EBC rule explains why a negative existential 
cannot be referred to using a defnp. A sample case, from 
Karttunen \[1968\], is given in D6. D6-2 is generally regarded 
as an unacceptable sentence following D6-1. The sentence is 
certainly grammatical, so the assumption by Karttunen is that 
the referential term the car is being used in some 
inappropriate manner. 
D6-1 \[ don't have a car. 
2 ~= The car is black. 
The EBC rule predicts that the car co-refers with the focus in 
D6-1. But a car in that sentence does not have a referent 
(because the speaker has just said so). Thus the use of the 
defnp in D6-2 causes the hearer to expect a reference when 
in fact there is no referential entity. 
A similar case, (16), also from Karttunen, does not 
involve negative existentials, but entities within modal 
contexts: 
(16) = Mary expected a present from John although th._ee 
present was expensive. 
(17) Mary expected a present from John although th._.ee 
present wasn't the thing that worried her. 
The defnp in (16) according to the EBC rule must co-refer with 
the focus. What is significant is that the Co-reference is 
acceptable, as (17) shows. What is odd about the second 
clause of (16) is the predication. This paper cannot give an 
account of such semantics, but intuitively, it seems odd to 
predicate the property of being expensive to something one 
expects. Thus as long as there is a co-referent entity 
specified by the focus, a defnp may be used, but the 
predication about the defnp must be semantically meaningful. 
Another form of explicit backwards co-re~erencing 
is slightly different than the previous examples. |~: involves 
the use of lexical generalization of the focus. Grosz \[~.977\] 
first categorized the relation of focus and defnps with a more 
general noun in the noun phrase head. In D7, t~e poor old 
beast is a lexical generalization of the dog, that is, its head 
noun is a term which is a class generalization of the focus. 7 
7. This term comes from the observation of HaUiday and 
Hasan \[1976\] that lexical cohesion includes the use of 
reiteration of four types: same word; synonym, superordinate, 
and general word. 
91 
Determining the class generalization of the focus is possible 
when the focus is represented in the way that is assumed in 
this paper: as an association network with an is-a hierarchical 
structure. Using that hierarchy, it is possible to determine 
whether a phrase like beast is hierarchically related to 
Salamut. 
D7-I Harold took hi_~s Salamut; to the vet yesterday. 
2 The poor old beast i was quite lame. 
One might expect that some constraint on the amount 
of information in the lexical generalization of the focus is 
needed. This is the case, since the underlined defnp in D7-2' 
is unacceptable following D7-1 as a co-referent with the focus. 
D7-2' The beast who is old was quite lame. 
-2" The mangy, snarling, unfriendly beast was quite 
lame. 
It appears from all the cases \[ can find that any post-nominal 
modifiers on a noun which is a lexical generalization of the 
focus force the defnp to be non-coreferential with the focus, 
while pre-nominal modifiers, no matter how complex, preserve 
co-referentiality. It is unclear why pre-nominal modifiers and 
post-nominal modifiers have these different behaviors. 
8. Implicit Backwards Co-reference 
Many definite noun phrases which occur in discourse 
are not cases of backwards co-reference to the focus. Grosz 
\[1977\] suggested that the focus brings other items implicitly 
into focus as well, by means of association. Such defnps are 
related to the focus in one of several ways. Since the focus is 
well specified, these relationships can be easily determined. 
The focus acts as an anchor point for finding references for 
such defnps. In DS, the defnp the time refers to the time of 
the discourse focus, the meeting. This defnp use | will call 
implicit backwards co-reference. Such cases are to be 
distinguished from explicit backwards co-referencing because 
the defnp is co-referential with an entity that is closely 
associated with the focus rather than to the focus itself. The 
phenomenon of association between two noun phrases has 
been cited by Norman and Rumelhart \[1975\]. 
D8-1 The pa group wants to have a meeting. 
2 The time will be 3 p.m. on Tuesday. 
implicit backwards co-referencing is constrained by 
the association network surrounding the focus. Any entity 
closely associated with the entity which represents the focus 
can be mentioned using a simple defnp. Thus in Dg~ sentences 
with acceptable defnps as well as ones with unacceptable 
defnps are given. 
D9-1 i went to a new restaurant with Sam. 
2 The waittress was nasty. 
3 The food was great. 
4 The soup was salty, but the wine was good. 
5 ,~ The rug was ugly. 
Non-simple defnps have infinitely more variety because the 
modifiers can specify the relation of the defnp to the focus at 
hand as in D9-6. Non-simple defnps which do not suggest 
some connection are less acceptable, but hearers, in reliance 
on the perspicuity maxim, may attempt some connection. Thus 
if D9 included D9-7 below, some hearers might attempt to 
connect the defnp with focus. 
D9-6 i like the band that plays there. 
-7 The elephant with the green tutu danced an 
impressive jig. 
Another use of focus is as an inference point for 
inferred co-references, inferred co-references, like the 
murderer in (8), presented here as DI0, are not mentioned 
explicitly in the previous discourse nor can they be considered 
closely associated to the focus on general principles. Their 
use reflects an inference about the focus on the part of the 
speaker. 
DIO-I The heiress lived the life of a recluse. 
2 She died under mysterious circumstances, but 
the murderer was never found. 
In D6, the murderer represents an inference that the heiress 
death was due to a specific type of circumstance, a murder. 
Such an inference is possible given a Fahlman \[1977\] type net 
with two inference points like heiress and murderer (and the 
information associated to heiress from the context thus far); 
from the net, the relation of the murderer to the heiress can 
be inferred. Such an inference does not produce a real-world 
entity to which the murderer refers. Instead, the inferred 
relation of murderer and heiress provides sufficient 
information to produce the entity if it exists in the database. 
When a denotation does not exist in the database, the 
inference between the murderer and the heiress sugests that 
the speaker is attributing of some individual that s/he is a 
murderer. 
A concrete example will illustrate my point. Suppose 
the hearer knows that the heiress was killed by Jones. Then 
on hearing DI0, the hearer not only concludes that the 
murderer refers to the murderer of the heiress, but also based 
on that conclusion, the hearer decides that Jones is denoted by 
the referring expression: However, another hearer upon 
hearing DtO and not knowing what the first hearer knew, could 
only conclude that murdel"er is attributed of a person who is 
assumed to have murdered the heiress. The referent is not 
known to the second hearer, but if someone were to tell 
him/her that Jones murdered the heiress, the hearer could 
conclude who the murderer refers to. in effect, the defnp 
used in this way points out the attributional use of expressions 
which Donnellen has observed. The argument presented here 
is not only about the nature of focus; it is a statement of what 
information is sufficient to make up a description which can 
denote a unique entity. Viewing inferred co-reference defnps 
as attributions has an implication for a computational model 
which disambiguates such defnps. This model must be able to 
use an expression without knowing its referent and be able to 
link up the denotation to the referring expression if some 
knowledge makes that denotation available at a later point. 
implicit and inferred co-reference at first glance 
appear to be one in the same thing. The discourse below, 
from Karttunen \[1968\], will indicate just how the two differ. 
DI 1-1 \[ was driving on the freeway the other day. 
2 Suddenly the engine began to make a funny 
noise. 
92 
3 I stopped the car. 
4 When I opened the hood, I saw that the radiator 
was boiling over. 
With focus of freeway in DI 1-1, the relation of the engine can 
be found since vehicles are driven on freeways and vehicles 
have engines. The association chain here suggests that the 
connection between DII-I and the engine involves a few 
inferences. These inferences are part of a heater's general 
knowledge and true of the worleJ. They are part of the 
knowledge in the association network. With D10, however, the 
inference about the murderer involves a supposition which is 
not necessarily true, since dying under" mysterious 
circumstances does not necessarily imply murder. The 
distinction between implicit and inferred co-reference can thus 
be stated: imp!icit co-reference involves inferences which are 
true about the world, while inferred co-reference involves a 
supposition which the speaker has made which is not 
necessarily true. 
Another kind of implicit focus-defnp relation exists in 
D12. ! call this relation the set-element relation since the 
clown with a unicycle is an element of the set of clowns which 
the focus denotes. 
Dl2-1a I went downtown today, 
lb and there were clowns performing in the 
square. 
2 The clown with a unicycle did this really 
fantastic stunt. 
As with inferred references, the focus does not make it 
possible to identify a specific denotation with the referring 
expression. Instead the focus is the set of which the referent 
of that phrase is a member. These cases are easier to 
distinguish than those of inferred reference because the head 
noun is the singular of the noun phrase represented in the 
focus. Unlike defnps using the EBC rule, set-element 
co-reference demands a modifier that distinguishes it from the 
focus. Without the modifier, there is no means of determining 
which member of the set is being discussed. Grosz \[1977\] 
says of cases like the set-element relation that an inference is 
needed to establish additional properties of an object in focus. 
However, because the head noun phrase is the same as the 
focus, while the modifier is different, the relation can be 
established without the need for inference. 
Another kind of focus relation, which I call computed 
reference, can be seen in DI3. Here the last meeting does not 
refer to the meeting mentioned in the previous sentence, but 
that meeting can be used as a point for determining a las._.~t 
meeting, if one is known in the database, else it is a 
description of the entity required, as with inferred references 
and set-element references. Several modifiers - first, last, 
next, second and the other ordinals - are used in this way. 
DI3-1 Aunt Het's Sewing Bee wants to have a meeting 
this week. 
2 The meeting should beon Tuesday. 
3 The last meeting, which was at 5, was too late, 
so schedule this one earlier. 
From these examples, the nature of focus in 
discourse can be re-emphasized. It is the focus which 
connects sentences of the discourse. In the process of 
determining the focus relation between a defnp and the focus, 
the link in the discourse is created. From these examples, one 
can predict that there ought to be cases of defnps which bear 
none of these relations to the focus and which result in 
unacceptable discourses. This prediction is accurate as D12" 
shows. 
Dt2"-la ! went downtown today, 
Ib and there were clowns performing in the square. 
2 • \] saw the chair. 
The difficulty with DI2' is that a chair is not associated with 
clowns, and the discourse does not suggest any suppositions 
that would link chairs and clowns inferentially. 
There are, however, cases of defnps which do not 
bear any relation to the focus and which are perfectly 
acceptable in discourse. Consider the chairman of the math 
department in D14 below. 
D14-1 George wants to have a semin.~r to discuss 
representation in frame-like languages. 
2 He wants to invite the chairman of the math 
department. 
The focus in D14 is the seminar of 014-1. The math 
department chairman is not directly related to the seminar. 
However, the focus does provide an important piece of 
information; it is the source of the ellided event to which the 
chairman is invited. Hence D\]4 is different from D9-6 or DI2' 
where there is no link between the focus and the sentence in 
question. What can be concluded is that the focus is not a 
useful reference point for the referent of the chairman of the 
math department. The sentence is not odd because of focus 
links. Thus something outside of the discourse context must 
contain the needed denotation. This conclusion points to the 
limitation of th e focus: it captures only those reference 
relations which are internal to the discourse. In a sense, the 
focus is a summary of the discourse context and what the 
hearer knows about it. 
How can the denotation of the chairman of the math 
department be found? Since the denotation of the chairman of 
the math department lies outside the discourse context, a more 
global context such as that of the speaker's situtation in time 
and space must be used to determine a context of possible 
referents. This context must be limited because there are 
potentially many math department chairmen in the speaker's 
and heater's memories. I do not intend to describe just what 
such a context will look like, but I do want to indicate that it 
may be "grown" from a search through the associative network 
to other entities which are related to any of the objects in the 
discourse, including the speaker and hearer. The association 
net includes not only abstract representations of general 
classes of real world entities, it also contains representations 
of real world objects. The associations between real world 
objects can be gathered by a search method which collects 
associations close to the focus and then extends for other 
associations until one is found that matches the defnp in 
question. Of course, it is possible that no entity will be found: 
such a circumstance is yet another example of the hearer 
knowing that a defnp refers without being able to tell who the 
93 
speaker had intended as the reference. The implications of 
extra-discourse contexts for computational models is clear: 
models are needed of what the speaker assumes his/her 
hearer knows about, so that the speaker does not produce 
referring expressions which the hearer cannot disambiguate, 
and models are needed of what the hearer assumes the 
speaker has beliefs about so that the hearer can decide what 
to do with referring expressions which s/he cannot 
disambiguate. 
Earlier in this paper \] mentioned the use of such 
defnps as the sun, the moon, and the Earth. These defnps 
have default referents in initial sentences of a discourse. 
Inside a discourse, they can act in one of two ways: related to 
the focus as either a backwards co-reference, implicit 
co-reference, inferred reference, set-element reference or 
computed reference; or they may act as references to entities 
outside the context. The EBC rule predicts that such defnps 
will be taken as co-referring when the focus has the same 
defnp head. Thus a defnp like the sun or the moon will refer 
to its default value only if the focus does not predict a 
referent based on,one of the five co-reference relations 
discussed here. These kinds of defnps are especially 
noteworthy because they are a clear example of a defnp that 
can be used in either role. 
The four related co-reference relations specify ways 
in which a defnp can be predicted as co-reference to an entity 
associated with the focus. Other defnps either refer to 
objects outside the discourse context or the defnp fails to 
refer as intended by the speaker. The former case is marked 
by the presence of discourse links elsewhere in the sentence 
to the focus or by the use of default reference. In those 
cases where the reference as intended fails, the hearer may 
attempt to create a connection to the focus, as was suggested 
with DIO-7, and thereby maintain the perspicuity contract. 
Alternatively, the hearer may fail to understand the referent. 
Focus can be used for disambiguation of generic 
defnps in a manner similar to the cases presented above. As 
discussed earlier, the EBC rule predicts that a defnp which is 
u-ambiguous will be generic or not based on the focus, and 
that a strictly generic defnp is independent of the focus. The 
strictly generic defnp case, as shown in DIS, presents one 
means of shifting' focus in a discourse with defnps. 
D15-1 I got a new ASR 33 this week. 
2 The ASR 33 is an old but reliable output device. 
If D15-3 were "It was available long before the newer, 
electronic consoles," the focus would have shifted from the 
ASR33 which the speaker had gotten to the generic entity of 
ASR33 teletypes. 
The strictly generic defnp used after a non-generic 
focus is just one case of implicit backwards co-reference using 
associations. Other associations occur as well. However, 
implicitly related defnps are not distinguishable as generic 
unless a full modifying noun phrase is attached, as D16 shows. 
The time as a simple defnp can be used only as an implicit 
co-reference to the focus of party. Only the complex noun 
phrase has the syntactic and semantic distinctions which 
reflect the generic usage. The simple defnp used implicitly 
takes its generic/specific classification from the focus. 
DI6-1 I want to have a party. 
la The time of a party is hard to decide on. 
2b The time is hard to decide on. 
Set-element implicit co-reference exists for generic 
foci as well as specific foci. A significant difference is that the 
foci for the generic case can be a singular defnp, or they can 
be a plural noun phrase with either a definite article or no 
article. The set membership is indicated by a distinguishing 
modifier, just as with specific set-element implicit 
co-reference. An example of generic set-element 
co-reference is given below with both a singular defnp focus 
and a plural unspecified focus. 
D\]7-! The Austrailian aborigine represents an almost 
extinct hunter-scavenger social group. 
l' Austrailian aborigines represent an almost 
extinct hunter-scavenger social group. 
2 The aborigine in the southern sections of 
Austrailia sometimes gathers food, butthe other 
aborigines do not. 
Inferred generic co-references also occur. In D18, 
the owner of a motorcycle is a generic defnp: 
D18-I Alfonso was in an accident with a motorcycle last 
week. 
2 I think the owner of a motorcycle ought to be 
required to take driving lessons. 
The owner of a motorcycle is generically related to the first 
sentence by an inference of what happened to the agent. The 
same kind of machinery that is used for specific inferred 
co-references can be used for making these inferences as well. 
How can generic inferred co-references be distinguished from 
specific inferred co-reference? A strictly generic defnp as in 
D18-2 remains generic. Those defnps which are u-ambiguOus 
at the sentential level, as with DlO-2, can be disambiguated as 
specific because of the relation to the focus. 
The use of a semantic network with a focus and 
inference mechanism results in a computational theory of 
co-reference which makes use of representation properties 
such as prototype of entities, hierarchical connections and 
associative links between entities in the representation. The 
use of focus for co-reference rules such as the EBC rules, 
implicit backwords co-reference rules and inferred 
co-reference relies on this net representation. An inference 
mechanism is necessary both to verify co-reference 
predictions and to test suppositions used in inferential 
co-reference. With the net and a focus, predictions about 
acceptable co-reference for noun phrases has been shown and 
verified with linguistic evidence. Psychological predictions, 
such as implicit co-reference requiring more time than explicit 
co-reference, canalso be tested although the related 
psychological literature has not been discussed in this paper. 
The limits of focus as a co-reference mechanism suggest that 
focus is central for co-reference of noun phrases related to 
previous discourse. For noun phrases that refer outside the 
discourse, focus may also be used to generate a context of 
entities from which a co-referent may be chosen. Further 
research can extend the focus mechanism to rules involving 
94 
other types of noun phrases and the personal pronoun 
anaphors. 
9. Acknowledgements 
This report describes research done at the Artificial 
Intelligence Laboratory of the Massachusetts Institute of 
Technology. Support for the laboratory's artificial intelligence 
research is provided in part by the Advanced Research 
Projects Agency of the Department of Defense under the 
Office of Naval Research under Contract Number 
NO0014-75-C-0643. 
I0. References 
Austin,J.L. \[1962\] How to Do Things With Words, Oxford 
University Press. 
Bullwinkle, C. \[1977\] Levels of Compexity in Discourse for 
Anaphora Disambiguation and Speech Act /nterpretation, 
Proceedings of the Fifth International Joint Conference in 
Artificial Intelligence, August 1977; also M.I.T.A.I. Lab 
Memo 413. 
Charniak, E. \[1972\] Toward a Model Of Children's Story 
Comprehension, M.I.T.A.I. Lab TR-266. 
Cohen, Philip R. \[1978\] On Knowing What to Say: Planning 
Speech Acts, University of Toronto Department of 
Computer Science, T.R. No. \] 1.8. 
Donnellan, Keith S. \[1977\] Reference and Definite Descriptions 
in Naming, Necessity, and Natural Kinds, S.P. Schwartz (ed.), 
Cornell University Press, ithaca. 
Fahlman, Scott E. \[1977\] A System for Representing and Using 
Real-World Knowledge, M.I.T.A.I. Laboratory AI-TR-450, 
December. 
Grosz, Barbara \[1977\] The Representation and Use of Focus in 
Dialogue Understanding. Stanford Research Institute 
Technical Note 151, Menlo Park, California 
Halliday, M.A. and ~uqaiya Hasan \[1976\] Cohesion in English 
London, Longman press. 
Karttunen, Lauri \[1968\] What Makes Definite Noun Phrases 
Definite? Paper p-3871, The" RAND corporation, Santa 
Moncia, California. 
Kripke, Saul A. \[1972\] Naming and Necessity. inLSemantics of 
Natural Language, Davidson and Harman le-'~s.)~eid'~ 
Publishing Co., Boston. 
Norman, Donald A. and David E. Rumelhart \[1975\] Explorations 
in Cognition W.H. Freeman, San Francisco. 
Rieger, Charles J. \[\]974\] Conceptual Memory: A Theory and 
Computer Program for Processing the Meaning Content of 
Natural Language Utterances. Stanford Artificial Intelligence 
Lab Memo AIM-233. 
Russell, Bertrand \[1905\] On Denoting, reprinted in Logic an.__dd 
Knowledge~ R.W. March (ed.), Allen and Unwin, London, 
1956. 
Searle, J.R. \[1969\] Speech Acts: An Essay in the Philoshpy of 
Language, Cambridge University Press. 
Sidner, C. \[forthcoming\] A Computational Model of 
Co-reference Comprehensio'n in English. Ph.D. dissertation, 
M.I.T. 
Ninograd, Terry \[1971\] Procedures as a Representation for 
Data in a Computer Program for Understanding NaturM 
Language. M.I.T. dissertation. 
95 
