On Lexical Cohesive Behavior of Heads of Deﬁnite

Descriptions: A Case Study

Beata Beigman Klebanov and Eli Shamir

School of Computer Science and Engineering,

The Hebrew University of Jerusalem, 91904, Israel,

Abstract. This paper uses materials from annotation studies of lexical cohe-

sion (Beigman Klebanov and Shamir, 2005) and of deﬁnite reference (Poesio

and Vieira, 1998; Vieira, 1998) to discuss the complementary nature of the two

processes. Juxtaposing the two kinds of annotation provides a unique perspective

for observing the workings of the reader’s common-sense knowledge at two lev-

els of text organization: in patterns of lexis and in realization of discourse entities.

1 Introduction

Introducing the notion of cohesion of a text, Halliday and Hasan [6] detail numerous

ways in which textual elements connect with each other, providing the perception of the

unity of the text. Among other kinds of texturizing devices, they cite lexical cohesion –

use of words with related meaning, and referential cohesion – the tendency of a text to

refer repeatedly to the same set of entities.

Subsequent research discovered that whereas the cohesive devices are often identi-

ﬁable on the basis of their form, the patterns they form with each other – what connects

to what – are by no means easy to establish by some systematic, algorithmic ’resolution’

procedure [4,10,12]. Readers were thought to do it effortlessly; however, reader-based

studies show cases of disagreement and difﬁculty [9].

In this paper, we will attempt to analyze the relationship between lexical and refer-

ential cohesion: whether and when they reinforce each other, and what can be learned

from their divergence. Human-annotated data is used to provide information about each

of the phenomena: The Wall Street Journal 1989 article ’Computers Start to Get Per-

sonal’ (shown as appendix A) was used both in Poesio and Vieira’s seminal annotation

study regarding referential behavior of deﬁnite noun phrases [9,11] and in Beigman

Klebanov and Shamir [1–3] lexical cohesion annotation experiment.

In what follows, we introduce the two annotation schemes, and sketch out our pre-

dictions of the interaction between the two types of cohesion (sections 2, 3). We then

proceed to testing the prediction on the basis of the annotated data.

2 Lexical Cohesion through Anchoring

In Beigman Klebanov and Shamir’s rendering, lexical cohesion between items in a text

arises on the basis of stereotypical, common-knowledge-based connections between the

We will use another annotated text which will be introduced in due course.

Shah H. (2006).

Chatterbox Challenge 2005: Geography of the Modern Eliza.

In Proceedings of the 3rd International Workshop on Natural Language Understanding and Cognitive Science, pages 109-119

DOI: 10.5220/0002505001090119

 SciTePress

relevant concepts [1–3]. Sensitive to text dynamics, their question to the subjects was:

”For every concept ﬁrst mentioned in the text, which previously mentioned concepts

help the easy accommodation of the current concept into the evolving story, if indeed

it is easily accommodated, based on the common knowledge as perceived by the sub-

ject” [3]. The preceding helper concept is called an anchor, and the relation is marked

anchored→anchor.

Only ﬁrst mentions of every lexical item in the text were subjected to anchoring

annotation. There were no limitations on the number of anchors per item or on the

textual distance between the anchored item and its anchor. Subjects were made aware

of the notion of referential cohesion through examples, and were asked not to mark

connections solely on the basis of identity of reference in the given text.

Beigman Klebanov and Shamir’s experiment resulted in 10 texts annotated for an-

choring relations by 22 readers. Detailed statistical analysis of the data identiﬁed 2 out-

liers, and set annotator agreement thresholds that ensure high degrees of reliability [1].

In particular, an item is reliably anchored if it was given some anchor by at least 13

out of 20 non-outliers

; an anchoring pair a→b is core, or strong, if the anchored item

a is reliably anchored, and the speciﬁc anchor b features in at least 6-7 annotations.

Examples will be given during case-study presentation; also see [2].

3 Deﬁnite Reference

The phenomenon of deﬁnite reference concerns both the text and the reader’s general

knowledge; it is thus particularly interesting to investigate its relationship with anchor-

ing. We will concentrate on one type of deﬁnite references - noun phrases quantiﬁed

by ’the’. Other deﬁnite references include demonstratives and possessives; however,

Poesio and Vieira’s study [9] only considered ’the’-deﬁnites.

According to the discourse analysis literature, ’the’-deﬁnites are used to mention

entities that should be uniquely identiﬁable by the hearer/reader on the basis of the

nominal alone in her mental representation of the current discourse augmented with her

general knowledge, as construed by the speaker/writer [5,8].

Unique identiﬁability is a cognitive status commanding an intermediate degree of

givenness in the following hierarchy [5]: In Focus > Activated > Familiar > Uniquely

Identiﬁable > Referential > Type Identiﬁable.

The sources for this degree of givenness are variable, and include linguistic context,

situational context, and general knowledge:

Linguistic: My sister has two children, a boy and a girl. The boy is an excellent student.

Situational: The boy sitting in the front row is misbehaving.

General Knowledge: When John came home, he saw that the door was already open.

In Poesio and Vieira’s study [9], deﬁnite descriptions were classiﬁed according to

the following taxonomy:

The gist of the example was that if in a given story a child’s father went out to the sea because

he is a sailor, then sailor→sea, but not sailor→father.

This corresponds to 99% reliability threshold.

The exact threshold depended on the text.

110

ENTITY referred to by a definite description

LINK

linked to previous mention

NO LINK

new in the text

mentioned before

(coreferential)

new but based on text

(bridging)

(associative)

presumably known to the average reader

(larger situation)

(hearer-old)

presumably new to the average reader

(unfamiliar)

(hearer-new)

Fig.1. Poesio and Vieira’s classiﬁcation of Deﬁnite Descriptions. Category descriptions are

shown according to Vieira’s [11] summary of instruction given to the annotators. Additional

category titles shown in parenthesis are used by Poesio and Vieira in their article, highlighting

the correspondence with other concepts commonly employed in studies of the phenomenon.

3.1 ’D’-type deﬁnites

Given the position of deﬁnites in the givenness hierarchy, the D type (no previous

knowledge of the entity) is not expected to be abundant. However, about 20% of the

deﬁnite descriptions annotated in Poesio and Vieira’s study belong to this type [11]. A

cursory inspection of cases unanimously classiﬁed as D yields numerous examples of

long deﬁnites with lexically rich post-modiﬁers of the head:

– the class of asbestos including crocidolite

– the unit of New York-based Loews Corp that makes Kent cigarettes

– the fact that New England proposed lower rate increases

In these cases, the reader probably does not know about the aforementioned classes,

units, and facts, but the rich description given on this ﬁrst mention is sufﬁcient for secur-

ing the unique identiﬁability. Thus, the identiﬁability is not based on either knowledge

or prevoius linguistic information, but on linguistic information within the NP itself.

In terms of lexical anchoring, we expect heads of D deﬁnites to be left un-anchored,

as they are judged to be new and not based on previous text. This is even more so for

the ’heavy-tail’ deﬁnites, where identifying information follows the head.

3.2 ’=’ type Deﬁnites

Deﬁnites that repeat an already mentioned entity constitute less than half of all ’the’-

deﬁnites in Poesio and Vieira’s study. A repeated mention of an entity can be done with

the same lexical head (as in ...a boy ...The boy ... ), or with a different lexical head (as

in ...a boy . ..the kid ...). In the ﬁrst case, anchoring information will not be available,

as the relevant item would constitute a lexical repetition.

In the different-head case where the current head is ﬁrst mentioned within the given

deﬁnite, we ask whether this head is anchored at all, and if yes, whether the anchor coin-

cides with the previous mention of the same entity (this would be the case if kid→boy in

the last example). In such cases, lexical and referential cohesion would go hand-in-hand,

111

intensifying the connectedness between the two items. Cases of divergent anchoring and

reference annotations would constitute examples of the item’s different positioning at

the two planes of textual structure.

3.3 ’R’ type Deﬁnites

These are cases where the entity is new but based on text, or, as Poesio and Vieira’s

guidelines [9] elaborate, ”based on, dependent on, related to some other idea or thing

in the text”, exemplifying by ”The Parks wanted to buy an apartment but the price

was very high”, where the price should be classiﬁed as R with an apartment as its

trigger/antecedent. This class is the smallest in their data, only 6-11% of all deﬁnites.

This class – often termed bridging references, or indirect anaphora – bears the most

similarity to anchoring. We point out that in reference annotation, the trigger/antecedent

itself is taken to be a noun phrase,

whereas anchoring is at the word, rather than phrase,

level, and the part-of-speech of the anchor is not constrained in any way.

Furthermore, anchoring guidelines allowed multiple anchors, whereas reference

guidelines ask for ”the previous related mention”, i.e. a single antecedent. Indeed, some

cases of annotator disagreement were due to different choices of antecedent [9].

Poesio and Vieira note that this was by far the most difﬁcult category to distinguish

from the rest. A point of interest is whether the difﬁculty was in the initial judgement

of relatedness to something, or in pinpointing a previous noun phrase that functions as

the relevant previous item. In the ﬁrst case, anchoring annotators might have a similar

difﬁculty; in the second, we expect to ﬁnd robust anchoring decisions often going to

something other than the head of the triggering/antecedent noun phrase.

3.4 ’K’ type Deﬁnites

K deﬁnites refer to an entity that ”was not mentioned in the text, not related to something

in the text, but it refers to something which is part of the common knowledge of the

writer and readers in general” [9]. This class covers 20-25% of deﬁnites.

Similarly to R type, non-relatedness is likely to be understood as non-relatedness to

some other entity mentioned in the text. Hence, K class is not necessarily exclusive of

anchoring, as long as the anchor is not a head of a noun phrase.

Alternatively, K deﬁnites could be references to entities that come-out-of-the-blue

as far as the text is concerned, but have a natural reference point in the reader herself,

as in Poesio and Vieira’s example: ”During the last 15 years housing prices increased

nearly ﬁvefold.” In such cases, the text would not provide any material for anchoring.

Table 1 summarizes prediction of anchoring according to reference classiﬁcation.

4 Reference vs. Anchoring - A Case Study

In the article ”Computers Start to Get Personal” – henceforth, PC-text, there are 6

examples of LINK-ENTITY (=/R) and 10 examples of NO-LINK (K/D) deﬁnites.

Poesio and Vieira speak about ”...the head noun of the deﬁnite description and the head noun

of its antecedent”, p. 200, our italics.

112

Table 1. Predicted relationship between referential and anchoring behavior.

Reference

Anchoring

Anchored Anchor Identity

= (Coref) Same Head —

Other Head Yes Possibly same as coreferent

R (Bridging) Yes Same or other than the trigger/antecedent

K (Knowledge) Yes Not a head of an NP

D (Unfamiliar) No

In our analysis of LINK-ENTITY deﬁnites, we expand the data with coreferential

deﬁnites from a NYT article ”Inspectors: TWA Explosion Did’t Originate Near Cock-

pit” (appendix B) – henceforth, TWA-text. This text was annotated for coreference

chains within MUC initiative [7], and used for anchoring annotation as well.

4.1 LINK-ENTITY Deﬁnites

Out of the 6 LINK-ENTITY deﬁnites in PC-text, 4 are unanimously classiﬁed as re-

peated reference. However, we do not get a chance to see anchoring at work, since in

none of the 4 cases is the head a newly introduced lexical item.

Table 2. LINK-ENTITY deﬁnites with lexically new heads. The heads are boldfaced. An-

tecedents are in square brackets. We show cases with majority =/R classiﬁcation. The last column

shows the head’s anchors produced by the parenthesized number of people (out of 20).

Deﬁnite NP P&V class Anchors

the Homebrew D hobbyists (6) homebrew (1) garage (1)

Computer Club R [hobbyists]

= [hobbyists]

the world leader in = [ibm] pioneer (4) led (3) world (3) new (2) ibm (2)

computers = [ibm] chairman (1) team (1) owners (1) after (1)

D example (1) triggered (1) market (1)

The two cases with judgement variability are shown in table 2. The deﬁnites are em-

bedded in a larger NP with which they co-refer: an appositive (IBM, the world leader

in computers) and a list construction with a single member (for hobbyists such as the

Homebrew Computer Club). There is an uncertainty in reference annotations - some

annotators marked embedded NP as coreferential with the whole of the appositive/list

construction, whereas others apparently decided that the embedded NP does not consti-

tute a second mention, but rather is part of the ongoing ﬁrst mention of the entity (D);

the single R annotation seems to hit the middle ground between the two solutions.

There is no reliable anchoring of the head of the deﬁnite, contrary to our prediction

for =/R types. In one case there is a tendency to anchor the head in its antecedent’s head

113

(club→hobbyists by 6 people), which is virtually nil in the other case (leader→ibm by 2

people). In the latter case, there are better anchors which are not heads of noun phrases:

pioneer is a modiﬁer inside the NP many pioneer PC contributors, and led is a verb.

For additional information on coreference and anchoring behavior of deﬁnites, we

turn to the TWA-text, and seek coreference chains that include coreferential deﬁnites

with ﬁrst-mention heads. There are four such chains; the relevant heads are boldfaced:

Crash-chain: the crash – the accident – the crash – the plane crash

Plane-chain: Trans World Airlines Flight 800 – the Boeing 747 – the jet – the plane

– the plane – plane – Flight 800 – plane – the airplane

Day-chain: Tuesday–the 20th day since Flight 800 exploded in midair off Long Island

Man-chain: James K. Kallstrom–the assistant director of the FBI’s New York ofﬁce

Table 3 presents the anchoring data. Let us consider the last two chains ﬁrst. For

day and director, the situation is similar to leader in PC-text: the items are embedded

in a coreferential appositive construction. Day shows negligible anchoring; director has

a stronger anchoring pattern, though it misses the core-reliability threshold.

Table 3. Anchoring behavior of coreferential deﬁnites, TWA-text.

Same-chain Anchoring Patterns

lexical heads Within the Chain Outside the Chain

crash – accident→crash (17) accident→ explosion (11) catastrophic (9)

– accident wreckage (7) survived (4) destroyed (2)

ﬂight 800 – boeing → ﬂight (15) boeing → airlines (14) twa (11) cockpit (11)

– Boeing 747– jet → ﬂight (17) boeing/747 (16) jet → airlines (16) cockpit (12) twa (9)

– jet – plane → jet (17) ﬂight (15) plane → airlines (12) cockpit (12) twa (10)

– plane – boeing/747 (16)

– airplane airplane → plane (19) jet (18) airplane → airlines (10) cockpit(9) twa(7)

boeing/747 (11) ﬂight (10) transportation (3) hangar (3) altitude (2)

Tuesday–day day → tuesday (2)

Kallstrom – director → chairman (7) assistant (4)

– director board (3) vice chairman (2) senior (1)

While the agreement on the best anchor for director is not high, chairman is the

preferred choice. This case shows clearly the distinction between referential and lexical

structures: the NP headed by chairman is not related referentially to the one headed by

director, as they refer to different people in different organizations. However, the direc-

tor is the second senior ofﬁcial named and quoted in this short article, so the items form

a pattern, for some readers. On the other hand, director is not cohesive anchoring-wise

with Kallstrom, as, presumably, this name does not feature in the readers’ knowledge.

We now turn to the crash- and plane-chains in table 3. None of the 7 deﬁnites in

question is embedded in its coreferent NP. Their behavior is radically different: all items

are anchored, with numerous strong anchors. Thus, accident in strongly anchored in its

coreferent’s head crash, but also in other items.

114

The pattern in the second chain is interesting: each member of the chain is strongly

anchored in each of the preceding members; all members of the chain

are strongly

anchored in the same additional items: airlines, cockpit, twa. We note that within-

chain anchors tend to be somewhat stronger than out-of-chain ones; this might mean

that coreference intensiﬁes the perceived lexical connectedness. The anchoring support

given to the referential structure makes it different from Kallstrom–director chain, in

that the latter is an accidental referential connection, whereas the items in plane-chain

have a lot of associative commonality. Anchoring also shows additional ’networking’

of the chain in the given text, connecting it to other aviation-related things.

To summarize the analysis of LINK-TYPE entities with respect to our predic-

tions: heads of coreferential deﬁnites that are not embedded within their coreferent NP

conﬁrm our expectation – they are anchored, and there is substantial anchoring texture

accompanying the coreference links. Anchoring provides additional information, show-

ing common-knowledge-based connection with items outside the coreference structure.

For heads of deﬁnites embedded in their coreferent NP, the anchoring support tends

to be weak or lacking. These are cases with an explicit syntactic construal of co-

referentiality, through an appositive. In the Gricean framework, this means that the

reader is not likely to be able to work out the connection without the explicit help;

Kallstrom-director is a good example. Hence, in terms of the involvement of the reader’s

knowledge, these cases are akin to D-deﬁnites – self-identifying lexically elaborate in-

troductions of unfamiliar entities. In fact, in both such cases in PC-text, there was a mi-

nority D annotation. The anchoring behavior of these deﬁnites matches what we would

expect of D-types, and is quite unlike that of other coreferential deﬁnites.

4.2 NO-LINK Deﬁnites

Table 4 shows the K/D deﬁnites from PC-text. Let us consider the 5 cases of majority K

annotations. Apart from Journal, all other items are reliably anchored, and have at least

one very strong anchor: century→ centennial, ofﬁce → business, drives→ computers,

drives→ disk, telephone → modems.

It is predicted that anchors should not be nominal heads in the text, so they are

not perceived as discourse entities. Centennial is an adjective; business and disk are

non-head noun modiﬁers. However, computers and modems are NP heads. It seems that

reference and anchoring are at odds regarding the relatedness of the entity to the text.

The deﬁnite the telephone is embedded in ’the internal modems that allow PCs to

share data via the telephone’, headed by the anchor modems. This is reminiscent of

the problem with embedded appositives – perhaps the annotators thought that an entity

which is still being introduced could not be used as a basis for an R-type connection.

The annotators knew about the disk drives for PCs, but took them to be unrelated

to previously discussed computers. Indeed, the three computers launched in 1977 did

not have disk drives, as the text implies (’could store about two pages of text in their

memories’ carries a scalar implicature that they could not store more). From the modern

perspective, these are rather non-typical computers. Perhaps, the reference annotators

assumed the 1977 perspective from which disk drives were a new development, whereas

including ﬂight, which is not listed as it was not introduced inside a coreferential deﬁnite NP

115

Table 4. NO-LINK deﬁnites. The heads are boldfaced. For minority =/R annotations, the an-

tecedent is in square brackets. Dots ... indicate ≥3 additional anchors, each marked by 1 person.

Deﬁnite NP P&V class Anchors

1 The Wall Street Journal K K K wall street (5)

2 the past century K K K centennial (19) year (7) past (2)

3 the face of personal computing K D D personal (1) changed (1)

4 the Apple II R [three computers] computers (17) ...

K D

5 the home and ofﬁce K K K business (11) computers (5) home (4)

pcs (2) desktop (4) owners (2) ...

6 the Altair = [b.-f.-kit types] pcs (1) types (1) apple ii (1)

K D commodore (1) tandy (1)

7 the team that developed D D D led (2) chairman (1)

the disk drives for PCs

8 the disk drives for PCs K K K computers (10) disk (9) keyboards (2)

pcs (2) data (2) technology (2) .. .

9 the internal modems that K/D D D computers (19) pcs (10)

allow PCs to share data keyboards (6) disk (4) technology (4)

via the telephone screens (4) drives (3) . ..

10 the telephone K K K modems (14) television (4)

technology (3) computers (2) .. .

anchoring annotators took their own current perspective, judging disk drives generally

related to computers. One difﬁculty for this explanation is the perspective of the text: it

mentions current, i.e. 1989, computers, which, presumably, did have disk drives.

Turning to the 3 cases where D type predominates, we note that two of them – items

3 and 7 – bear out our prediction of lack of anchoring. Additionally, case 6, coming from

’earlier built-from-kit types such as the Altair, Sol and IMSAI’ is a by now familiar case

of a deﬁnite embedded inside a NP with overlapping reference; the reference annotation

shows confusion, but the lack of anchoring places this case ﬁrmly in ’D’-type company.

Case 4 also shows uncertainty regarding reference. Apple II is one of the three

computers mentioned in the preceding sentence. The anchoring pattern is extremely

strong, suggesting that (some) computers are the entity to which Apple II is related. A

minority vote indeed opts for R; the K/D decisions are puzzling.

However, perhaps the most surprising case is that of modems (case 9): A robust D

type reference-wise, with an overwhelming anchoring connection to computers. The

discussion of perspective taken by annotators is possibly relevant here: The three com-

puter brands repeatedly mentioned in the text did not have modems. Still, surely the

connection between modems and computers is of readers’ knoweldge!

We suggest that the solution in the wording of instructions to reference annotators:

”...the D[eﬁnite]D[escription] is self-explanatory or it is given together with its own

identiﬁcation. In these cases it becomes clear to the general reader what is being talked

about even without previous mention in the text or without previous common knowl-

edge of it” [9]. Thus, D-types are not necessarily unfamiliar to the reader; rather, the

reader does not have to use the familiarity in order to achieve unique identiﬁcation. In

116

this case, the 12-word deﬁnite headed by modems is squarely within this category, even

though the common knowledge could be in place, too. The mere existence of cases like

this is surprising Grice-wise, as they seem to provide superﬂuous information that the

reader could have recovered on the basis of her knowledge. Modems are introduced as

an important invention of the past, as are disk drives; however, there is no elaboration

about what disk drives are. Possibly, the ’modems’ case is slightly over-indulging for

an up-to-date 21st-century reader; perhaps modems were still a rarity in 1989.

We thus see that D-class is not homogenous with respect to the role of knowledge:

it contains elements that are unfamiliar, so the reader has to use the material inside the

deﬁnite to reach the unique identiﬁability, and elements that are in fact quite familiar in

the reader’s common knowledge, but their presentation in the text is such that the reader

does not have to use her knowledge to interpret them.

In contrast, anchoring sides completely with the reader’s knowledge: a putative con-

nection is either supported by it or not, irrespective of what the text says about each en-

tity. This is because anchoring asks for intuitive stereotypical judgment, for the shallow

but robust load brought into the text by the mere use of a word, like modems. Anchoring

is meant to uncover how such ’loads’ organize into structures in the text, below the level

of the discourse-entity-based who-did-what-to-whom stories where reference operates.

5 Conclusion

This paper reported a case study of the relationship between referential behavior of

deﬁnite NPs and lexical anchoring of their heads, on the basis of juxtaposed relevant

annotations of two texts.

We observed a tendency for deﬁnites whose referent repeats or relates to some pre-

vious textual entity (=/R) to be anchored in their antecedents, as well as in other things,

providing additional text-based connections (plane vs. airlines, cockpit).

Even when an entity is judged to be referentially unrelated to the text, but of reader’s

knowledge (K-type), the anchoring pattern shows what could have triggered the relevant

knowledge earlier in the text (disk→computers, century→centennial). Often, the anchor

is not a nominal head, although we saw cases of heads as well.

When the entity is judged new in the text (D-type), we discerned two sub-types.

In case the entity is genuinely unfamiliar, the lexical anchoring texture for the head

is indeed meager. In case the entity is familiar, but in the current case could as well be

uniquely identiﬁed by the current description, the anchoring pattern reveals the familiar-

ity. Such discrepancies could be interesting from a historical perspective, as, assuming

Gricean cooperative framework, they detect potential knowledge mismatches between

text-creation-time and current audience.

In the other direction, lack of anchoring tends to corresponds to D-class, but also to

cases where the deﬁnite is embedded, through an appositive or a list structure, inside

an NP with overlapping reference. These cases often show disagreement in reference

type classiﬁcation, possibly reﬂecting confusion as to the availability of a referent-in-

the-making as an antecedent. There is usually a minority D annotation in these cases.

Clearly, additional parallel annotation is needed to check out these trends. Such

work is promising in exposing the intricate, multi-level workings of the reader’s knowl-

117

edge upon the text: not only does it help to consolidate the plot of the story by tracing

repeated and related referents, but also to prepare subtle, associative ground for intro-

duction of new things and ideas, and to strengthen the perceived unity of the text by

enriching the network of connections between its elements.

Acknowledgments

We thank Renata Vieira for giving us access to deﬁnite description annotations.

References

1. Beigman Klebanov, Beata: Using Readers to Identify Lexical Cohesive Structures in Texts. In

Proceedings of ACL Student Session (2005) 55-60

2. Beigman Klebanov, Beata, and Shamir, E: Lexical Cohesion: Some Implications of an Empiri-

cal Study. In Proceedings of 2nd International Workshop on Natural Language Understanding

and Cognitive Science, Miami, USA (2005) 13-21

3. Beigman Klebanov, Beata, and Shamir, E: Guidelines for Annotation of Concept Mention

Patterns. Technical Report 2005-8, Leibniz Center for Research in Computer Science, The

Hebrew University of Jerusalem, Israel (2005)

4. Ge, Niyu, Hale, John, and Charniak, Eugene: A Statistical Approach to Anaphora Resolution

In Proceedings of the 6th Workshop on Very Large Corpora (1998) 161-170

5. Gundel, J.K., Hedberg, N., and Zacharski, R: Cognitive Status and The Form of Referring

Expressions in Discourse. Language 69(2) (1993) 274-307

6. Halliday, M.A.K., and Hasan, R.: Cohesion in English. Longman Group Ltd. (1976)

7. Hirschman, Lynette: MUC-7 Coreference Task Deﬁnition, version 3. In Proceedings of the

7th Message Understanding Conference (1997)

8. Lyons, Christopher: Deﬁniteness. Cambridge, UK: Cambridge University Press (1999)

9. Poesio, Massimo, and Vieira, Renata: A Corpus-based Investigation of Deﬁnite Description

Use. Computational Linguistics 24(2) (1998) 183-216

10. Tetreault, Joel: Analysis of syntax-based pronoun resolution methods. ACL (1999) 602-605

11. Vieira, Renata: Deﬁnite Description Processing in Unrestricted Text. PhD Thesis, University

of Edinburgh (1998)

12. Vieira, Renata, and Poesio, Massimo: An Empirically-based System for Processing Deﬁnite

Descriptions. Computational Linguistics 26(4) (2000) 539-593

Appendix A: PC text

Computers Start to Get Personal, 1977

1989

Wall Street Journal

(During its centennial year, The Wall Street Journal will report events of the past century that

stand as milestones of American business history.)

Three computers that changed the face of personal computing were launched in 1977.

That year the Apple II, Commodore Pet and Tandy TRS-80 came to market. The computers

were crude by today’s standards. Apple II owners, for example, had to use their television sets as

screens and stored data on audiocassettes. But Apple II was a major advance from Apple I, which

was built in a garage by Stephen Wozniak and Steven Jobs for hobbyists such as the Homebrew

Computer Club. In addition, the Apple II was an affordable $1,298.

118

Crude as they were, these early PCs triggered explosive product development in desktop

models for the home and ofﬁce.

Big mainframe computers for business had been around for years. But the new 1977 PCs –

unlike earlier built-from-kit types such as the Altair, Sol and IMSAI – had keyboards and could

store about two pages of data in their memories. Current PCs are more than 50 times faster and

have memory capacity 500 times greater than their 1977 counterparts.

There were many pioneer PC contributors. William Gates and Paul Allen in 1975 developed

an early language-housekeeper system for PCs, and Gates became an industry billionaire six

years after IBM adapted one of these versions in 1981. Alan F. Shugart, currently chairman of

Seagate Technology, led the team that developed the disk drives for PCs. Dennis Hayes and Dale

Heatherington, two Atlanta engineers, were co-developers of the internal modems that allow PCs

to share data via the telephone.

IBM, the world leader in computers, didn’t offer its ﬁrst PC until August 1981 as many other

companies entered the market. Today, PC shipments annually total some $38.3 billion world-

wide.

Appendix B: TWA text

Investigators: TWA explosion didn’t originate near cockpit

1996

New York Times News Service

After picking apart some of the wadded remains of the cockpit of Trans World Airlines Flight

800, investigators concluded Tuesday that the catastrophic explosion that destroyed the Boeing

747 most likely did not originate inside the cockpit or in the electronics bay beneath it.

They were partly persuaded by a surprising discovery found in the ton of wreckage that had

been the jet’s cockpit: The circles of glass that cover many of the cockpit dials, and even a light

bulb above a staircase that led to the plane’s upper deck, had somehow survived the crash intact.

“You have this mass of wreckage and yet things from that area are relatively the way they

were before the accident,” said Robert Francis, vice chairman of the National Transportation

Safety Board. “There is no indication at this point of anything in that area that would give cause

for concern in terms of something having initiated there.”

A senior investigator who looked at the cockpit wreckage Tuesday said that one of the plane’s

altimeters — instruments that show the plane’s altitude — was frozen with a reading of 13,100

feet. Altimeters are mechanically driven instruments that do not depend on electricity to work,

so the ﬁnding suggests that the mechanics continued working for several seconds after the initial

explosion, at about 13,700 feet.

Federal investigators continued their search for the cause of the crash Tuesday, the 20th day

since Flight 800 exploded in midair off Long Island and plunged into the Atlantic Ocean, killing

all 230 people on board. On the seas and on the shore, investigators said they made a modest

amount of progress, though they still have not determined if the plane crash was caused by a

bomb, a missile attack or a mechanical malfunction.

At the former Grumman hangar in Calverton, investigators on Tuesday began piecing to-

gether the fractured parts of the airplane. They also pulled about one-third of the cockpit wreck-

age off the one-ton ball of metal, essentially unwrapping it.

James Kallstrom, the assistant director of the FBI’s New York ofﬁce, said he had sent many

agents who had been working in Suffolk County back to their home ofﬁces, mostly in New York

City. Criminal investigators are anxious for the cause to be determined, he said, adding: “We are

in a bit of a waiting pattern.”

119