SCALE AND THE CLASSIFICATORY DIMENSION

A Linguistic Approach to Contextual IR

Esben Alfort

Department of International Language Studies and Computational Linguistics

Copenhagen Business School, Copenhagen, Denmark

Keywords: Contextual IR, Ontologies, Subjectivity, Context, Scale, perspectives, Basic level, Individuation.

Abstract: Users of information retrieval systems presumably have (subconscious) reasons for choosing certain ways

of formulating their queries. Consequently, the words used may tell us something about the users’

intentions. However, researchers in Contextual IR have a strong emphasis on computational solutions and

tend to ignore a careful linguistic analysis of the actual queries. As a first step towards such a linguistic

treatment, I suggest that we treat classification as a dimension parallel to space and time and learn from our

experience with these dimensions in trying to cope with subjectivity in connection with IR systems.

1 RELEVANCE

We are now in the fortunate position of having

access to ready-made ontologies for use in semantic

applications. It is therefore regrettable that

companies producing semantic software cannot

benefit fully from these resources, because the

resulting ontologies simply contain far too much

information. A search engine relying on an elaborate

ontology with many kinds of relations would return

far too many hits, because most of them would

simply not be relevant to a given user in the context

at hand. We therefore find ourselves forced to

exploit only a fraction of the ontological information

available, and even such straightforward relations as

subsumption may have to be disregarded in some

cases, because the subcategories are judged to be

irrelevant to most potential users searching for that

particular concept. However, in some cases these

subcategories might nonetheless be important. It

would thus be preferable to be able to include all the

information related to a concept and only exclude

some of it from consideration at the last moment if it

is judged irrelevant in the case at hand. For instance,

when users search for furniture on an e-trading site

using a semantic search engine based on an

ontology, chances are that they are not interested in

This paper is partly based on a longer version to be

presented at a workshop at the Terminology and

Knowledge Engineering conference in Dublin 2010.

a complete list of documents on every kind of

furniture available. Rather, a link to a page with an

overview of furniture in store would be a lot more

helpful in most cases. This is because furniture is a

highly general and heterogeneous category, which

does not refer to any specific type of furniture and

thus can only be used in generalizations.

Consequently, the simple subsumption relation is not

very relevant in this case, whereas in other

circumstances it may be extremely useful.

Pilot studies of log files from semantic search

engines powered by Ankiro (www.ankiro.com)

confirm that different user groups characterized by

different intentions use search terms from very

different classificatory levels (i.e. general vs.

specific terms), and that different kinds of sites are

visited by different combinations of such user types.

Very general categories are characteristic of sites

concerned with laws and principles, because rules

are generalizations and refer to generalized types. If

I were to search for pets in a municipal context, I

would most likely be looking for laws or suggestions

concerning pets in general, and a list of documents

on guinea pigs and golden retrievers would be quite

unhelpful. Texts on rats would probably be

completely irrelevant, dealing mostly with pest

control. On a pet shop site, on the other hand, a list

of pet types would in all probability be just what I

was looking for.

Deciding what is irrelevant is exceedingly tricky,

due to the countless facets of relevance (Borlund,

370

Alfort E..

SCALE AND THE CLASSIFICATORY DIMENSION - A Linguistic Approach to Contextual IR.

DOI: 10.5220/0003078803700373

In Proceedings of the International Conference on Knowledge Engineering and Ontology Development (KEOD-2010), pages 370-373

ISBN: 978-989-8425-29-4

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

2003). It seems obvious, though, that we should at

least pay close attention to how users actually state

their queries. Their linguistic structure may tell us

whether they represent a quest for information or a

transactional or navigational query (Daoud et al.,

2009). To some degree, it may even reveal what

kind of information is sought. Unfortunately, this

latter possibility is rarely pursued, due to lack of a

thorough, linguistic analysis (Tamine-Lechani et al.,

2010, p. 6). Instead, the focus has been on

developing computational solutions to the problem

of identifying the user’s context. Whenever the

linguistic structure of the query is at all consulted,

features noted are rather superficial ones like the

number of words in a query, as well as the possible

presence of interrogative words in the string (Jansen

et al., 2008). Also, certain lexical items that signal a

transactional intention (buy, download, software,

and so on) are utilized. However, having decided

that a request is of the informational type – which is

the case in more than 80 % of web search queries

according to Jansen et al. (2008) – the actual query

is then put to one side, and recourse is in general

made to user activity or other sources of information

on user interests and preferences. Very frequently,

the query is then modified by the addition of

specifying and disambiguating terms (e.g. Phinitkar

& Sophatsathit, 2010). It is my endeavour to

contribute a linguistic view of the matter, in which

the query itself plays the main part.

Information retrieval researchers usually look at

what documents X are relevant in context C, given a

query Q. I shall instead be exploring what

expression Q a user would be likely to use in

reference to a potential document (or set, or type of

documents) X believed to satisfy his or her

information need, given the context C. This

enhances the understanding of the user’s choice of

search terms. On the other hand, it masks the fact

that users may actually need something different

from what they think they do. However, we can

hardly solve that problem until we have a

satisfactory understanding of the queries themselves.

2 CHOOSING ONE’S WORDS

One is easily led into thinking that every word

capable of reference corresponds more or less

directly to a certain class of referents. This is in no

way the case (Brown, 1958). Even though there are

in many cases strong preferences for calling a

certain type of entity by one specific name, countless

others are possible in principle. When such

alternatives are chosen, this is a meaningful act

performed by the speaker, and we should understand

this as an important piece of information on how the

reference is to be interpreted. There are special

situations when I might describe a rat by the words

pet, rodent, mammal, animal, or even object. I might

also call it a nuisance, friend, or test individual. The

sets of potential referents to these expressions differ

hugely; consequently, a search containing such

terms should differ correspondingly. The focus of

my research is currently on why a user chooses a

certain hierarchical level in reference to an entity.

I have found that classificatory information can

profitably be considered in a way similar to spatial

and temporal data, allowing us to apply the

understanding that linguists have already acquired of

perspectival phenomena in connection with space

and time to classification and learn important new

lessons about this complex phenomenon. It might

not be obvious at first glance how this is possible. I

would remind the reader of the fact that the temporal

dimension is replete with spatial metaphors like “a

short time”, “time moves”, “we are approaching the

future”, “from the point of view of fifty years ago”,

and so on. We do not hesitate in treating the

temporal and the spatial dimensions alike,

irrespective of whether we consider them to be

physically identical or not. This last question is

irrelevant here, because ontologies are a cognitive

phenomenon, so it is enough that the two

phenomena are treated alike in cognition. I propose

that we take the next step and recognize

classification as yet another dimension. In doing so,

I am not suggesting that classification is a physical

dimension, but simply that we seem to treat it as

such cognitively, and that consequently we can learn

much about classification by drawing on our

knowledge of spatial and temporal phenomena.

Again, I might point to the existence of common

metaphors such as “from the point of view of

forestry...”, in which case forestry is not to be

understood as an entity but rather as a topic or

mindset, on which all other classification is based (if

in any way relevant to forestry). We also speak of

“broad topics”, “closing in on a topic”, and so on,

statements that testify to the fact that we think of

descriptions (i.e. classification) as something

navigable - a plane of sorts.

It would seem that we treat all phenomena in the

world that we want to portray linguistically as sets of

spatial, temporal and classificatory coordinates.

I shall consider all points regions (cf. Tarski,

1929), since if you zoom in on them they generally

turn out to be regions that were simply too small for

SCALE AND THE CLASSIFICATORY DIMENSION - A Linguistic Approach to Contextual IR

371

the internal distinctions to be relevant in the context

in question. What corresponds to a region in space

or a period of time in classificatory terms is the

category - or rather the set of potential referents to

which a term applies. The classificatory region is

small if a very specific term is used, but includes an

increasingly large number of potential referents

when more general terms are resorted to.

In portraying situations, we readily zoom in on

specific points in space and time that we are

concerned about, whereas we generalize and let the

distinctions go blurred if we consider them irrelevant

in the context at hand. The same is true of

classification; we can be highly specific about the

type to which a certain referent belongs, or we can

get away with very general descriptions if we have

no reason to be more explicit.

We can only portray situations as if observed

from some point of view. We place an imaginary

observer at a chosen point in space and time (which

I shall call the host of the observer) and portray the

situation as it appears from that point of view

(Alfort, 2009). The host is situated within a domain

of relevance, which includes the places, times and

categories considered, or in other words those

regions in all of these dimensions that constitute

contrasts to any focused subregion of special

interest. The phenomenon of scale (e.g. Langacker,

2006, p. 116) is caused by the difference in size

between the domain of relevance and the focused

subregion.

3 CLASSIFICATORY

HOMOGENEITY

A referent with a spatial extent may be

homogeneous if it shows relatively little variation

across its spatial region. If we were to zoom in on

such a region, we would most likely find that it was

not entirely homogeneous, because there are usually

small distinctions, which simply pale into

insignificance from a larger perspective. However,

even greater heterogeneity would be obtained if we

were to zoom out, say from a region consisting

solely of (a) rock to a section of the seabed on which

it was lying. The spatial region would now

incorporate other substances such as sand, water and

crab. A category may likewise be homogeneous if

all potential referents within the domain of interest

show relatively little variation. Zooming in on them

will make their relative differences grow in

importance, and subtypes will emerge. If we zoom

out and generalize, however, the category is likely to

become even more heterogeneous, as it includes

more potential referents of clearly different types.

The world is rarely really homogeneous, so the

only way of attaining complete apparent

homogeneity is in fact by excluding deviating

instances. If I posit that “Italy is a beautiful

country”, this is of course a generalization that

excludes some less appealing regions such as the

occasional refuse tip. Similarly, if someone is

“walking in a forest”, he or she is not in fact within

anything. Rather, the person in question is walking

among the trees that make up the forest that we

choose to consider a homogeneous entity for the

sake of this statement. In a similar manner, I might

say, “If I were a bird, I’d just fly away”, though

strictly speaking, if I were an ostrich this would

simply not be an option. I thus effectively exclude

ostriches from consideration by not including them

in the relevant classificatory region behind the term

bird. I do this in order to be able to treat the category

as homogeneous, which is a precondition for a

generalization to work - the statement must apply to

all relevant instances. Note that such generalizations

are of course impossible in reference to an ostrich,

because the homogeneous classificatory region

behind the category bird would not include

ostriches. Consequently, people for whom the ability

to fly is a crucial property of birds probably rarely

refer to ostriches with the term bird, unless the

domain of relevance excludes all prototypical kinds

of birds, in which case the category would be

homogeneous without further generalization. It is for

the same reason that we tend not to call boxing

gloves gloves (Murphy & Lassaline, 1997, p. 110) -

such subcategories are simply too deviant to allow

generalization across a heterogeneous domain.

4 CLASSIFICATORY LEVELS

AND INDIVIDUATION

In connection with space, we are rarely interested in

heterogeneous regions; we mostly concern ourselves

with the regions that correspond exactly to

individuals (at least in connection with prototypical,

physical objects). There is a similar tendency to

concentrate on homogeneous classificatory regions.

This is a phenomenon known traditionally as basic

level (Rosch et al., 1976). However, I suggest that

basic level categories are in fact what one might call

classificatory individuals, i.e. regions in the

classificatory dimension that are relatively

KEOD 2010 - International Conference on Knowledge Engineering and Ontology Development

372

homogeneous, and at the same time clearly

delimited from surrounding regions. It is important

to note that what appears to be an individual is a

subjective judgment (cf. Wisniewski et al., 2003, p.

587). When seen from afar, a flock of sheep is

considered an individual; it may move about on the

hillside, and it may disintegrate into separate sheep,

which are also individuals, but as long as there is a

flock, it is considered an individual, homogeneous

flock of sheep. However, if a shepherd is looking for

one particular sheep, he will focus on the individual

animals. The flock has become heterogeneous,

because different animals are treated as having

different coordinates in the spatial dimension.

Individuation is a highly subjective phenomenon

depending on the granularity of the portrayal as well

as the size of the domain of relevance. A vet caring

for an injured sheep would focus on that particular

animal and would distinguish individual muscles

and bones, while the sheep as a whole would be

highly heterogeneous for his or her purposes. In the

classificatory dimension, the vet would hardly find it

useful if all parts of the animal were referred to as

instances of sheep, since everything within his or her

domain of relevance would be sheep. In the same

way, the shepherd would hardly refer to separate

animals as instances of flock, even though they

would be, just as drops of water are instances of that

liquid.

The fact that individuals are homogeneous

regions in space, time and classification means that

they dissolve if we zoom too much in on the details,

because this makes them heterogeneous. Certainly,

individuals have an extremely privileged cognitive

status to all humans (Bloom & Kelemen, 1995, p. 7),

but this status is not restricted to those entities that

appear as individuals in an everyday human context.

Rather, whenever we encounter individuals, whether

they be flocks, sheep, or muscles, they receive the

same privileged status in our consciousness. General

and specific terms are used under very different

circumstances and for distinct purposes. We can

only interpret a query by reference to the context,

including the perspective from which the situation is

seen in spatial, temporal and classificatory terms. If

a very general term is used in a restricted domain of

relevance, its meaning is highly dependent on

context, because the apparent homogeneity is a

product of the restricted domain of relevance rather

than an absence of contrast across domains.

Compared to this, specific terms are much more

straightforward. Consequently, I am currently

researching the contexts that allow general terms to

be used, and the meanings and intentions behind

such expressions. If we can establish the size of the

domain of relevance as well as the generality of a

term, then this will hint at whether subordinate

categories are likely to be relevant to the user

providing it.

REFERENCES

Alfort, E., 2009. Subjektiv og objektiv

referenceklassifikation i analyser af kommunikation og

tekst. Master's Thesis, University of Copenhagen.

Bloom, P. & Kelemen, D., 1995. Syntactic cues in the

acquisition of collective nouns. Cognition, 56, pp. 1-

30.

Borlund, P., 2003. The concept of relevance in IR. Journal

of the American Society for Information Science and

Technology, 54 (10), pp. 913-925.

Brown, R., 1958. How shall a thing be called?

Psychological Review, 65, pp. 14-21.

Daoud, M., Tamine-Lechani, L., Duy, D. B. &

Boughanem, M., 2009. Towards a graph-based user

profile modeling for a session-based personalized

search. Knowledge and Information Systems, 21, pp.

365-398.

Jansen, B.J., Booth, D.L. & Spink, A., 2008. Determining

the informational, navigational, and transactional

intent of Web queries. Information Processing and

Management, 44, pp. 1251-1266.

Langacker, R.W., 2006. On the continuous debate about

discreteness. Cognitive Linguistics, 17 (1), pp. 107-

151.

Murphy, G. L., & Lassaline, M. E., 1997. Hierarchical

structure in concepts and the basic level of

categorization. In K. Lamberts, & D. Shanks (eds),

Knowledge, Concepts, and Categories, pp. 93-132.

Psychology Press.

Phinitkar, P. & Sophatsathit, P., 2010. Personalization of

search profile using ant foraging approach. Lecture

Notes in Computer Science 6019, International

Conference on Computation Science and Its

Applications (ICCSA2010), pp. 209-224. Springer

Verlag.

Rosch, E., Mervis, C., Gray, W., Johnson, D. & Boyes-

Braem, P., 1976. Basic level objects in natural

categories. Cognitive Psychology, 8, pp. 382-439.

Tamine-Lechani, L., Boughanem, M. & Daoud, M., 2010.

Evaluation of contextual information retrieval

effectiveness: overview of issues and research.

Knowledge and Information Systems, 24, pp. 1-34.

Tarski, A., 1929. Foundations of the geometry of solids,

reprinted in Tarski, 1982, pp. 24-29.

Tarski, A., 1982. Logic, Semantics, Metamathematics.

Hackett Publishing Co, 2

edition.

Wisniewski, E. J., Lamb, C. A., & Middleton, E. L., 2003.

On the conceptual basis for the count and mass noun

distinction. Language and Cognitive Processes, 18

(5/6), pp. 583-624.

SCALE AND THE CLASSIFICATORY DIMENSION - A Linguistic Approach to Contextual IR

373