33 The Process of Describing Resources

Describing Instances or Describing Collections

It is simplest to think of a resource description as being associated with another individual resource. As we discussed in Resources in Organizing Systems, it is challenging to determine what to treat as an individual resource when resources are themselves objects or systems composed of other parts or resources. For example, we sometimes describe a football team as a single resource and at other times we focus on each individual player. However, after deciding on resource granularity, the question remains whether each resource needs a separate description.

Libraries and museums specialize in curating resource descriptions about the instances in their collections. Resource descriptions are also applied to classes or collections of resources, because a collection is also a resource (“The Concept of “Collection””). Archives and special collections of maps are typically assigned resource descriptions, but each document or map contained in the collection does not necessarily have its own bibliographic description. Similarly, business and scientific datasets are invariably described at collection-level granularity because they are often analyzed in their entirety.

Furthermore, the granularity of description for a collection of resources tends to differ for different users or purposes. An investor who owns many different stocks focuses on their individual prices, while other investors put their money in index funds that combine all the separate prices into a single value.

Many web pages, especially e-commerce product catalogs and news sites, are dynamically assembled and personalized from a large number of information resources and services that are separately identified and described in content management and content delivery systems. However, a highly complex collection of resources that comes together in a single page is treated as a single resource when that page appears in a list of search engine results. Moreover, all of the separately generated pages can be given a single description when a user creates a bookmark to make it easy to return to the home page of the site.

Abstraction in Resource Description

We can also associate resource descriptions with an entire type or domain of resources. (See “Preserving Resource Types” and “Resource Domain”.) A collection of resource descriptions is vastly more useful when every resource is described using common description elements or terms that apply to every resource. A schema (or model, or metadata standard) specifies the set of descriptions that apply to an entire resource type. Sometimes this schema, model, or standard is inferred from or imposed on a collection of existing resources to ensure more consistent definitions, but more often, it is used as a specification when the resources are created or generated in the first place. (See What about “Creating” Resources? in “Introduction”.)

A relational database, for example, is easily conceptualized as a collection of records organized as one or more tables, with each record in its own row having a number of fields or attributes that contain some prescribed type of content. Each record or row in the database table is a description of a resource—an employee, a product, anything—and the individual attribute values, organized by the columns and rows of the table, are distinct parts of the description for some particular resource instance, like employee 24 or product 8012C.^[1]

The information resources that we commonly call documents are, by their nature, less homogeneous in content and structure than those that can be managed in databases. Document schemas, commonly represented in SGML or XML, usually allow for a mixture of data-like and textual descriptive elements.

XML schema languages have improved on SGML and XML by expressing the description of the document schema in XML itself, making it easy to create resources using the metadata as a template or pattern. XML schemas are often used as the specifications for XML resources created and used by information-intensive applications; in this context, they are often called XML vocabularies. XML schemas can be used to define web forms that capture resource instances (each filled-out form). XML schemas are also used to describe the interfaces to web services and other computational resources.^[2]

Scope, Scale, and Resource Description

If we only had one thing to describe, we could use a single word to describe it: “it.” We would not need to distinguish it from anything else. A second resource implies at least one more term in the description language: “not it.” However, as a collection grows, descriptions must become more complex to distinguish not only between, but also among resources.

Every element or term in a description language creates a dimension, or axis, along which resources can be distinguished, or it defines a set of questions about resources. Distinctions and questions that arise frequently need to be easy to address, such as:

Therefore, as a collection grows, the language for describing resources must become more rigorous, and descriptions created when the collection was small often require revision because they are no longer adequate for their intended purposes.^[3]

This co-evolution of descriptive scope and description complexity is easy to see in the highly complex bibliographic descriptions created by professional catalogers. The commonly used Anglo-American Cataloguing Rules(AACR2) cataloging standards distinguish 11 different categories of resources and specify several hundred descriptive elements. AACR2 has recently been superseded by the Resource Description and Access(RDA) standards, which make fine-grained distinctions about content, media type, and carrier (technology).^[4]

Because the task of library resource description has been standardized at national and international levels, cataloging work is distributed among many describers whose results are shared. The principle of standardization has been the basis of centralized bibliographic description for a century.

Centralized resource description by skilled professionals works for libraries, but even in the earliest days of the web many library scientists and web authoring futurists recognized that this approach would not scale for describing web resources. In 1995, the Dublin Core(DC) metadata element set with only 15 elements was proposed as a vastly simpler description vocabulary that people not trained as professional catalogers could use. Since then, the Dublin Core initiative has been highly influential in inspiring numerous other communities to create minimalist description vocabularies, often by simplifying vocabularies that had been devised by professionals for use by non-professionals.^[5]

Of course, a simpler description vocabulary makes fewer distinctions than a complex one; replacing “author,” “artist,” “composer” and many other descriptions of the person or non-human resource responsible for the intellectual content of a resource with just “creator” (as Dublin Core does) results in a substantial loss of precision when the description is created and can cause misunderstanding when the descriptions are reused.^[6]

The negative impacts of growing scope and scale on resource description can sometimes be avoided if the ultimate scope and scale of the organizing system is contemplated when it is being created. It would not be smart for a business with customers in six US states to create an address field in its customer database that only handled those six states; a more extensible design would allow for any state or province and include a country code. In general, however, just as there are problems in adapting a simple vocabulary as scope and scale increase, designing and applying resource descriptions that will work for a large and continuously growing collection might seem like too much work when the collection at hand is small.

The challenges that arise with large description vocabularies are transformed when resource descriptions are created and assigned by computer algorithms. A large dataset might contain many thousands of descriptions for each resource, but clearly the computer does not have cognitive difficulty generating or using them. However, computer models with this many features can be hard for people to understand and trust.

Resource Description to Support Selection

Defining selection as the process by which resources are identified, evaluated, and then added to a collection in an organizing system emphasizes resource descriptions created by someone other than the person who is using them. We can distinguish several different ways in which resource description supports selection:

We also consider the perspective of the person creating the resource description and his or her primary purpose, which is often to encourage the selection of the resource by someone else. Product marketing is about devising names and descriptions to make a resource distinctive and attractive compared to alternatives. For many years prunes were promoted as a dietary supplement that people (especially old ones) need to “maintain regularity.” But after the California Prune Board (the world’s biggest supplier) re-branded them as “dried plums” and started marketing them as a snack food (and simultaneously renaming itself as the California Dried Plum Board) sales increased significantly.^[8]

Many countries require that imported goods are labeled with their country or origin. Consumers often use this property in resource descriptions as an indicator of high quality, as they might with Swiss watches, French or Italian fashions, or Canadian bacon. Alternatively, consumers might want to buy domestic or locally-sourced goods out of economic patriotism or to comply with procurement regulations. Not surprisingly, when consumers view origin in a positive light, this information is conspicuous and easy to read. In contrast, when consumers view origin less positively, perhaps as a warning of low-quality goods, the supplier is likely to make the origin information as inconspicuous as legally possible, or might even misrepresent the goods as domestic ones.^[9]

This misrepresentation is also ubiquitous in online dating, though the amount of misrepresentation must be balanced with the goals of the relationship and chances of the deception being discovered.^[10]

Resource Description to Support Organizing

We have defined organizing as specifying the principles for describing and arranging resources to create the capabilities upon which interactions are based. This definition treats the creation of resource descriptions and their use to organize resources for interactions as separate and sequential activities. This is easiest to see when people assign keywords and classifications to documents, or when sensors produce data, and these resource descriptions are later used to enable document retrieval or data analysis. A department store clerk might sort dress shirts on a display table using labels that describe their brands, sizes, and other properties. Rules governing the collection, integration, and analysis of personal information are also resource descriptions that influence the organization of information resources.

However, even if resource description and resource organization are logically separable, at times they are intertwined. When you arrange your own clothes, you don’t use explicit resource descriptions and instead rely on implicit ones about easily perceived properties like color, shape, and material of composition. When algorithms rather than people analyze texts to identify descriptive features for applications like information retrieval, spam classification, and sentiment analysis, resource descriptions and resource organization co-evolve, often continuously as the algorithm adapts and learns with each new resource it describes. This tight connection between resource description and resource organization is also exploited in organizing systems that use usage records from session logs, browsing, or downloading activities as interaction resources, tying them to payments for using the resources or analyzing them to influence the selection and organizing of resources in future personalized interactions. (See “The Concept of “Interaction Resource””)

Resource Description to Support Interactions

Most discussions of the purposes of resource descriptions and metadata emphasize the interactions that are based on resource descriptions that have been intentionally and explicitly assigned. The Functional Requirements for Bibliographic Records(FRBR), defined by library scientists, specifies the four interactions of Finding, Identifying, Selecting, and Obtaining resources, but these apply generically to organizing systems, not just those in libraries.^[11]

Elaine Svenonius proposed that a fifth task called Navigation be added to the FRBR list, and in 2016 that happened but it was renamed as “Explore”:^[16]

Resource Description to Support Maintenance

Many types of resource descriptions that support selection (“Resource Description to Support Selection”) are also useful over time to support maintenance of specific resource and the collection to which they belong. In particular, technical information about resource formats and technology (software, computers, or other) needed to use the resources, and information needed to ensure resource integrity is often called “preservation metadata” in a maintenance context.^[19]

Resource descriptions that are more exclusively associated with maintenance activities include version information and effectivity, or useful life information. Equipment maintenance schedules are typically related to the number of miles driven (indicated by a car’s odometer), number of hours operated (stored by many engines), number of pages printed, or other easily recorded information about resource use or interactions. With smart resources now capable of capturing, analyzing, and communicating more data about real-time performance, more sophisticated prediction and scheduling of maintenance work is now possible. It is also easier to identify resources that are not being used as much as expected, which might imply that they are no longer needed and can thus be safely archived or discarded.

Resource Description for Sensemaking and Science

Up to now in “Determining the Purposes”, we have discussed how resource descriptions are used to perform well-defined tasks within an existing organizing system. However, there is a broader and less well-defined purpose of resource description that is older and more fundamental: the use of resource descriptions as the raw material for making sense of the world.

For thousands of years, even before the invention of written language, people have systematically collected things, information about those things, and observations of all kinds to understand how their world works. Paleolithic humans made cave paintings depicting the results of hunts and animal migrations; ancient Egyptians recorded the annual floods of the Nile River in stone carvings; and Babylonian, Egyptian, Chinese, and Mesoamerican astronomers organized lunar, solar, and planetary observations as calendars starting about five thousand years ago.

These diverse efforts to impose meaning on experience by recording, analyzing, organizing, and reorganizing observations can be collectively described as sensemaking. (See the sidebar, Sensemaking and Organizing.)

Some aspects of sensemaking are hard-wired by evolution, which has given our brains powerful mechanisms that automatically simplify and organize the perceptual data we obtain from the world (see the sidebar Gestalt Principles). But this automatic sensemaking is dominated and amplified by intentional sensemaking.

Intentional sensemaking takes place when systematic statistical, experimental, and scientific methods are consciously followed to extract and organize knowledge from collections of samples, observations, or measurements. It is critical to recognize here that the contents of these collections represent choices made about what to collect, because most things and most phenomena have a great many descriptions or properties that could be recorded about them.

After things or data have been collected, statistical methods summarize the values of properties in a collection or dataset and the relationships among them. Making sense of a single collection or dataset by determining the properties that contrast and classify the instances is the start toward the more important goal of understanding the larger set or population from which the initial collection is just a sample. There is no better example of this than the periodic table of elements developed by Mendeleev in 1869, who organized known elements on the basis of their common chemical properties and then successfully predicted some properties of yet undiscovered ones.

Computational models developed from the initial dataset can predict future observations. Classification models assign a new instance to a category (e.g., spam or not spam message, Madison or Hamilton as author, outdoor or indoor scene); regression models predict a specific value of some measurement (given a description of a new movie, how much money will it make?); ordinal regression models predict values for non-metric measures (how much will you like the movie?). Experimental methods for hypothesis testing help develop and refine models of any type by systematically varying the conditions under which observations are made to discover how the results change in different situations.

A fundamental challenge in sensemaking and modeling is finding a balance between the competing goals of understanding a particular collection or dataset and being able to apply that understanding to new instances. Models can differ in the number of resource descriptions they use as parameters, and it is easy and tempting to overfit a model by using more parameters that capture random variations in observations. Overfitting produces spurious accuracy in reproducing the original observations, but it makes models less generalizable.

The highest level of sensemaking is the creation of scientific models or theories that propose interpretable and causal mechanisms for the observations. And just as automatic sensemaking creates simple explanations, scientists generally prefer simpler theories, a heuristic known as Occam’s Razor, or the law of parsimony. Even though complex theories can sometimes be more accurate, simpler theories produce more testable predictions, making it easier to verify or refine the theory. Occam’s famous principle, expressed eight centuries ago, is to prefer models that make the fewest assumptions, often measured in terms of the number of parameters or variables needed to make a prediction.^[21][22][23]

Intrinsic Static Properties

Intrinsic or implicit properties are inherent in the resource and can often be directly perceived or experienced. Static properties do not change their values over time. The species of an animal, the material of composition of a wooden chair, and the diameter of a wheel are all static properties that do not change their values over time. Static properties like color or shape are often used to describe and organize physical resources.

Intrinsic physical properties are usually just part of resource descriptions. In many cases, physical properties describe only the surface layer of a resource, revealing little about what something is or its original intended purpose, what it means, or when and why it was created. The author of a song and the context of its creation are other examples of intrinsic and static resource properties that are not directly perceivable.

Intrinsic descriptions are often extracted or calculated by computational processes. For example, a computer program might calculate the frequency and distribution of words in some particular document. Similarly, visual signatures or audio fingerprints are intrinsic descriptions (“Describing Non-text Resources”).

Some relationships among resources are intrinsic and static, like the parent-child relationship or the sibling relationship between two children with the same parents. Part-whole or compositional relationships for resources with parts are also intrinsic static properties often used in resource descriptions. However, it is better to avoid treating resource relationships as properties, and instead express them as relations. Describing Relationships and Structures discusses part-whole and other semantic relationships in great detail.

Extrinsic Static Properties

Extrinsic or explicit properties are assigned to a resource rather than being inherent in it. The name or identifier of a resource is often arbitrary but once assigned does not usually change. Arranging resources according to the alphabetical or numerical order of their descriptive identifiers is a common organizing principle. Classification numbers and subject headings assigned to bibliographic resources are extrinsic static properties, as are the serial numbers stamped on or attached to manufactured products.

For information resources that have a digital form, the properties of their printed or rendered versions might not be intrinsic. Some text formats completely separate content from presentation, and as a result, style sheets can radically change the appearance of a printed document or web page without altering the primary resource in any way. For example, were a different style applied to this paragraph to highlight it in bold or cast in 24-point font, its content would remain the same.

Intrinsic Dynamic Properties

Intrinsic dynamic properties change over time. Developmental personal characteristics like a person’s height and weight, skill proficiency, or intellectual capacity, for example. Because these properties are not static, they are usually employed only to organize resources whose membership in the collection is of limited duration. Sports programs or leagues that segregate participants by age or years of experience are using intrinsic dynamic properties to describe and organize the resources.

Extrinsic Dynamic Properties

Extrinsic dynamic properties are in many ways arbitrary and can change because they are based on usage, behavior, or context. The current owner or location of a resource, its frequency of access, the joint frequency of access with other resources, its current popularity or cultural salience, or its competitive advantage over alternative resources are typical extrinsic and dynamic properties that are used in resource descriptions. A topical book described as a best seller one year might be found in the discount sales bin a few years later. A student’s grade point average is an extrinsic dynamic property.

Extrinsic dynamic properties are useful features for data scientists making prediction or classification models. Your current location, the thing you just bought, and the place you bought it can be viewed as manifestations of unobservable preferences and values. Fingerprints found on a doorknob at a crime scene are an extrinsic dynamic property associated with the door, and clever detectives would analyze them along with other interaction resources they discovered with the goal of identifying the person for whom the fingerprints are intrinsic static properties.

Many relationships between resources are extrinsic and dynamic properties, like that of best friend.

Contextual properties are those related to the situation or context in which a resource is described. Dey defines context as “any information that characterizes a situation related to the interactions between users, applications, and the surrounding environment.” ^[25] This open-ended definition implies a large number of contextual properties that might be used in a description; crisper definitions of context might be “location + activity” or “who, when, where, why.” Since context changes, context-based descriptors might be appropriate when assigned but can have limited persistence and effectivity (“Resources over Time”); the description of a document as “receipt of a recent purchase” will not be useful for very long.

Citations of one information resource by another are extrinsic static descriptions when they are in print form, but when they are published in digital libraries it is usually the case that “cited by” is a dynamic resource description. Similarly, any particular link from one web page to another is an extrinsic static description, but because many web pages themselves are highly dynamic, we can also consider links as dynamic as well. Citations and web links are discussed in more detail in Describing Relationships and Structures.

Resources are often described with cultural properties that derive from conventional language or culture, often by analogy, because they can be highly evocative and memorable. ^[27]

Sometimes a cultural description outlives its salience, losing its power to evoke anything other than puzzlement about what it might mean.^[28]

For the Lego boys, current with the latest Star Wars movies, “light saber” was just the obvious description for a long, neon tube with a handle. However, someone unfamiliar with the Star Wars franchise might not understand“light saber,” and would describe the piece some other way.

Principles of Good Description

In The Intellectual Foundation of Information Organization, Svenonius proposes a set of principles or “directives for design” of a description language.^[29] Her principles, framed in the narrow context of bibliographic descriptions, generally apply to the broad range of resource types we consider in this book.

Any set of general design principles faces two challenges.

Who Uses the Descriptions?

Focus on the user of the descriptions. This is a core idea that we cannot overemphasize because it is implicit in every step of the process of resource description. All of the design principles in the previous section share the idea that the design of the description vocabulary should focus on the user of the descriptions. Are the resources being organized personal ones, for personal and mostly private purposes? In that case, the description properties and terms can be highly personal or idiosyncratic and still follow the design principles.

Similarly, when resource users share relevant knowledge, or are in a context where they can communicate and negotiate, if necessary, to identify the resources, their resource descriptions can afford to be less precise and rigorous than they might otherwise need to be. This helps explain the curious descriptions in the Lego story with which we began this chapter. The boys playing with the blocks were talking to each other with the Legos in front of them. If they had not been able to see the blocks the others were talking about, or if they had to describe their toys to someone who had never played with Legos before, their descriptions would have been quite different.

More often, however, resource descriptions can not assume this degree of shared context and must be designed for user categories rather than individual users: library users searching for books, business employees or customers using part and product catalogs, scientists analyzing the datasets from experiments or simulations. In each of these situations resource descriptions will need to be understood by people who did not create them, so the design of the description vocabulary needs to be more deliberate and systematic to ensure that its terms are unambiguous and sufficient to ensure reliable context-free interpretation. A single individual seldom has the breadth of domain knowledge and experience with users needed to devise a description vocabulary that can satisfy diverse users with diverse purposes. Instead, many people working together typically develop the required description vocabulary. We call the results institutional vocabularies, to contrast them with individual or cultural ones. (We will discuss this contrast more fully in Categorization: Describing Resource Classes and Types)

Some resource descriptions are designed for use by machines, which seemingly reduces the importance of design principles that consider user preferences or common uses. However, even if resources are described and organized by algorithms, when people need to explain the classifications and predictions that the algorithms produce, resource descriptions that are comprehensible and easily communicated are preferable to statistically optimal ones. Moreover, standardization and integration principles become more important for inter-machine communication to enable efficient processing, reuse of data and software, and increased interoperability among organizing systems.^[30]

Controlled Vocabularies and Content Rules

As we defined in “Use Controlled Vocabularies”, a controlled vocabulary is a fixed or closed set of description terms in some domain with precise definitions that is used instead of the vocabulary that people would otherwise use. For example, instead of the popular terms for descriptions of diseases or symptoms, medical researchers and teaching hospitals can use the National Library of Medicine’s Medical Subject Headings(MeSH) controlled vocabulary.^[31]

We can distinguish a progression of vocabulary control: a glossary is a set of allowed terms; a thesaurus is a set of terms arranged in a hierarchy and annotated to indicate terms that are preferred, broader than, or narrower than other terms; an ontology expresses the conceptual relationships among the terms in a formal logic-based language so they can be processed by computers. We will say more about ontologies in Describing Relationships and Structures.

Content rules are similar to controlled vocabularies because they also limit the possible values that can be used in descriptions. Instead of specifying a fixed set of values, content rules typically restrict descriptions by requiring them to be of a particular data type (integer, Boolean, Date, and so on). Possible values are constrained by logical expressions (e.g., a value must be between 0 and 99) or regular expressions (e.g., must be a string of length 5 that must begin with a number). Content rules like these are used to ensure valid descriptions when people enter them in web forms or other applications.

Vocabulary Control as Dimensionality Reduction

In most cases, a controlled vocabulary is a subset of the natural or uncontrolled vocabulary, but sometimes it is a new set of invented terms. This might sound odd until we consider that the goal of a controlled vocabulary is to reduce the number of descriptive terms assignable to a resource. Stated this way the problem is one of dimensionality reduction, transforming a high-dimensional space into a lower-dimensional one. Reducing the number of components in a multidimensional description can be accomplished by many different statistical techniques that go by names like “feature extraction,” “principle components analysis,” “orthogonal decomposition,” “latent semantic analysis,” “multidimensional scaling,” and “factor analysis.” ^[32]

These techniques might sound imposing and they are computationally complex, but they all have the same simple concept at their core, that the features or properties that describe some resource are often highly correlated. For example, a document that contains the word “car” is more likely to contain the words “driver” and “traffic” than a document that does not. Similar correlations exist among the visual features used to describe images and the acoustic features that describe music. Dimensionality reduction techniques analyze the correlations between resource descriptions to transform a large set of descriptions into a much smaller set of uncorrelated ones. In a way this implements the principle of Sufficiency and Necessity we mentioned in “Principles of Good Description” because it eliminates description dimensions or properties that do not contribute much to distinguishing the resources.

Here is an oversimplified example that illustrates the idea. Suppose we have a collection of resources, and every resource described as “big” is also described as “red,” and every “small” resource is also “green.” This perfect correlation between color and size means that either of these properties is sufficient to distinguish “big red” things from “small green” ones, and we do not need clever algorithms to figure that out. But if we have thousands of properties and the correlations are only partial, we need the sophisticated statistical approaches to choose the optimal set of description properties and terms, and in some techniques the dimensions that remain are called “latent” or “synthetic” ones because they are statistically optimal but do not map directly to resource properties.

Resource Description by Professionals

Before the web made it possible for almost anyone to create, publish, and describe their own resources and to describe those created and published by others, resource description was generally done by professionals in institutional contexts. Professional indexers and catalogers described bibliographic and museum resources after having been trained to learn the concepts, controlled descriptive vocabularies, and the relevant standards. In information systems domains professional data and process analysts, technical writers, and others created similarly rigorous descriptions after receiving analogous training. We have called these types of resource descriptions institutional ones to highlight the contrast between those created according to standards and those created informally in ad hoc ways, especially by untrained or undisciplined individuals.^[35]

Resource Description by Authors or Creators

The author or creator of a resource can be presumed to understand the reasons why and the purposes for which the resource can be used. And, presumably, most authors want to be read, so they will describe their resources in ways that will appeal to and be useful to their intended users. However, these descriptions are unlikely to use the controlled vocabularies and standards that professional catalogers would use.

Resource Description by Users

Today’s web contains a staggering number of resources, most of which are primary information resources published as web content, but many others are resources that stand for “in the world” physical resources. Most of these resources are being described by their users rather than by professionals or by their authors. These “at large” users are most often creating descriptions for their own benefit when they assign tags or ratings to web resources, and they are unlikely to use standard or controlled descriptors when they do so.^[36] The resulting variability can be a problem if creating the description requires judgment on the tagger’s part. Most people can agree on the length of a particular music file but they may differ wildly when it comes to determining to which musical genre that file belongs. Fortunately most web users implicitly recognize that the potential value in these “Web 2.0” or “user-generated content” applications will be greater if they avoid egocentric descriptions. In addition, the statistics of large sample sizes inevitably leads to some agreement in descriptions on the most popular applications because idiosyncratic descriptions are dominated in the frequency distribution by the more conventional ones.^[37]

We are not suggesting that professional descriptions are always of high quality and utility, and socially produced ones are always of low quality and utility.^[38] Rather, it is important to understand the limitations and qualifications of descriptions produced in each way. Tagging lowers the barrier to entry for description, making organizing more accessible and creating descriptions that reflects a variety of viewpoints. However, when many tags are associated with a resource, it increases recall while decreasing precision. (See “Resource Description by Users”)

Automated and Computational Resource Description

A picture’s EXIF file created by a digital camera records properties associated with the camera and its settings, as well as some properties of the photo-taking context. (See Figure: Contrasting Descriptions for a Work of Art. for an example.) Creating this highly detailed description by hand would be nearly impossible. The downside, however, is that the automated description does not capture the meaning of the photo; an automated picture description captures the time and place, but not that it is a picture of a honeymoon vacation. The difference between automated and human description is called the semantic gap (“The Semantic Gap”).

Any resource that is smart enough to collect data about its state or environment is creating resource descriptions automatically (See “Active or Operant Resources”). Resources with computational capabilities can process the raw sensor data to identify important events and create more interpretable descriptions.

Some computational approaches create resource descriptions that are similar in purpose to those created by human describers. Text mining and summarization systems for customer comments about products can reduce thousands of comments to a list of the most important features.^[39] People shopping for books at Amazon.com get insights about a book’s content and distinctiveness from the statistically improbable phrases that it has identified by comparing all the books for which it has the complete text.^[40]

Computational descriptions can use any observable or latent variable (see the sidebar, Latent Feature Creation and Netflix Recommendations) except some that are prohibited by law, such as race, religion, national origin, and marital status, to prevent discrimination. In practice, however, this prohibition is easily circumvented because these properties can usually be predicted using other ones. For example, race can often be reliably predicted using residence address and surname.^[41]

Of course, all information retrieval systems compare a description of a user’s needs with descriptions of the resources that might satisfy them. IR systems differ in the resource properties they emphasize; word frequencies and distributions for documents in digital libraries, links and navigation behavior for web pages, acoustics for music, and so on. These different property descriptions determine the comparison algorithms and the way in which relevance or similarity of descriptions is determined. We say a lot more about this in “Describing Non-text Resources” and in Interactions with Resources.

Evaluating the Creation of Resource Descriptions

When professionals create resource descriptions in a centralized manner, which has long been the standard practice for many resources in libraries, there is a natural focus on quality at the point of creation to ensure that the appropriate controlled vocabularies and standards have been used. However, the need for resource description generalizes to resource domains outside of the traditional bibliographic one, and other quality considerations emerge in those contexts.

Resource descriptions in private sector firms are essential to running the business and in interacting efficiently with suppliers, partners, and customers. Compared to the public sector, there is much greater emphasis on the economics and strategy of resource description.^[43] What is the value of resource description? Who will bear the costs of producing them? Which of the competing industry standards will be followed? Some of these decisions are not free choices as much as they are constraints imposed as a condition of doing business with a dominant economic partner, which is sometimes a governmental entity.

For example, a firm like Wal-Mart with enormous market power can dictate terms and standards to its suppliers because the long-term benefits of a Wal-Mart contract usually make the initial accommodation worthwhile. Likewise, governments often require their suppliers to conform to open standards to avoid lock-in to proprietary technologies. ^[44]

In both the public and private sectors there is increased use of computational techniques for creating resource descriptions because the number of resources to be described is simply too great to allow for professional description. A great deal of work in text data mining, web page classification, semantic enrichment, and other similar research areas is already under way and is significantly lowering the cost of producing useful resource descriptions. Some museums have embraced approaches that automatically create user-oriented resource descriptions and new user interfaces for searching and browsing by transforming the professional descriptions in their internal collections management systems.^[45] Google’s ambitious project to digitize millions of books has been criticized for the quality of its algorithmically extracted resource descriptions, but we can expect that computer scientists will put the Google book corpus to good use as a research test bed to improve the techniques.^[46]

Web 2.0 applications that derive their value from the aggregation and interpretation of user-generated content can be viewed as voluntarily ceding their authority to describe and organize resources to their users, who then tag or rate them as they see fit. In this context the consistency of resource description, or the lack of it, becomes an important issue, and many sites are using technology or incentives to guide users to create better descriptions.

Evaluating the Use of Resource Descriptions

Regardless of, or in addition to, any quality criteria applied to the creation and selection of resource descriptions, at some point the resource descriptions meet their intended users. The most important quality criterion at that point is whether the resource descriptions satisfy their intended purposes in a usable way. In many ways, the answer is a disappointing no.

For example, in one of the earliest revisions to the original HTML specification, a <META> tag was added to allow creators of web resources to define a set of key terms to describe a website or web page. This well-motivated resource description was to be used by search engines to improve the relevance of retrieved pages. However, it soon became obvious that it was possible to “game” the META tag by adding popular terms even though they did not accurately describe the page. Today search engines ignore the <META> tag for ranking pages, but many other techniques that use false resource descriptions continue to plague web users. (See “Social and Web Curation”.)

The design of a description vocabulary circumscribes what can be said about a resource, so it is important to recognize that it implicitly determines what cannot be said as well, with unintended negative consequences for users. The resource description schema implemented in a physician’s patient management system defines certain types of recordable information about a patient’s visit—the date of the visit, any tests that were ordered, a diagnosis that was made, a referral to a specialist. The schema, and its associated workflow, impose constraints that affect the kinds of information medical professionals can record and the amount of space they can use for those descriptions. Moreover, such a schema might also eliminate vital unstructured space that paper records can provide, where doctors communicate their rationale for a diagnosis or decision without having to fit it into any particular box.

However, when resource descriptions are the data used to train models for prediction or classification, the focus of evaluation is not on the descriptions, which are often assumed to be accurate observations about the world. Instead, evaluation focuses on the model, and “model selection” is the task of choosing which of several competing models best fit the original data while also generalizing well to new data. In any event, any quality problems or selection biases with the original data will undermine the value of whatever model is selected.

The Importance of Iterative Evaluation

The inevitable conflicts between quality goals mean that there will be compromises among the quality criteria. Furthermore, increasing scale in an organizing system and the steady improvements of computational techniques for resource description imply that the nature of the compromise will change over time. As a result, a single evaluation of resource descriptions at one moment in time will not suffice.

This makes usage records, navigation history, and transactional data extremely important kinds of resource descriptions because they enable you to focus efforts on improving quality where they are most needed. Furthermore, for organizing systems with many types of resources and user communities, this information can enable the tailoring of the nature and extent of resource description to find the right balance between “rich and comprehensive” and “simple and efficient” approaches. Each combination of resource type and user community might have a different solution.

The idea that quality is a property of an end-to-end process is embodied in the “quality movement” and statistical process control for industrial processes but it applies equally well to resource description. The central idea is that quality cannot be tested in by inspecting the final products. Instead, quality is achieved through process control—measuring and removing the variability of every process needed to create the products.^[47] Explicit feedback from users or implicit feedback from the records of their resource interactions needs are essential as we iterate through the design process and revisit the decisions made there.