Affect analysis of text using fuzzy semantic typing - IEEE Xplore
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
483
Affect Analysis of Text Using Fuzzy Semantic Typing Pero Subasic, Member, IEEE, and Alison Huettner
Abstract—We propose a novel, convenient fusion of natural language processing and fuzzy logic techniques for analyzing the affect content in free text. Our main goals are fast analysis and visualization of affect content for decision making. The main linguistic resource for fuzzy semantic typing is the fuzzy-affect lexicon, from which other important resources—the fuzzy thesaurus and affect category groups—are generated. Free text is tagged with affect categories from the lexicon and the affect categories’ centralities and intensities are combined using techniques from fuzzy logic to produce affect sets—fuzzy sets representing the affect quality of a document. We show different aspects of affect analysis using news content and movie reviews. Our experiments show a good correspondence between affect sets and human judgments of affect content. We ascribe this to the representation of ambiguity in our fuzzy affect lexicon and the ability of fuzzy logic to deal successfully with the ambiguity of words in a natural language. Planned extensions of the system include personalized profiles for Web-based content dissemination, fuzzy retrieval, clustering, and classification. Index Terms—Computing with words, fuzzy logic, knowledge engineering, text mining, world wide web.
I. INTRODUCTION
T
HE huge amount of text stored on computer systems is getting larger every day. Moving beyond the basic assumption that a given piece of text should be easily located, the next generation of systems aims toward integrated, personalized services and decision support. In these areas, a quick analysis of particular qualities in the text and an intuitive presentation to the user become increasingly important. To match an individual user’s profile on the World Wide Web, for example, it is necessary to introduce a human dimension into text understanding and representation. Expectations for modeling the purely subjective, human dimensions of text and data understanding are high, on both the users’ and providers’ sides. Here we experiment with qualitative analysis of affect-related information in free text. Affect-related information includes words describing emotions; like fear, anger, love, joy, and sorrow; feelings like warmth and excitement; attitudes like helpful, friendly, hostile; and other related categories such as temperament, humor, frame of mind, mood, spirits, morale, and disposition (including words like sweet, farce, wary, sanguine, depressed, eagerness, selfish). Our reasons for selecting this particular domain are threefold. First, affect-related information is pervasive in electronic documents: in news stories (on politics, competitive sports, etc.), economic reports (corporate acquisitions, investor reactions), Manuscript received September 11, 2000; revised May 8, 2001. The authors are with the Clairvoyance Corporation, Pittsburgh, PA 15232 USA (e-mail: [email protected]). Publisher Item Identifier S 1063-6706(01)06537-7.
customer-oriented Internet sites (eBay.com, amazon.com), newsgroups (misc.consumer), corporate customer service, email (customer complaints, questions, opinions), artistic and cultural material (movie and art reviews), etc. Second, affect information is critical to human communication: recent work shows the importance of emotions in decision-making, perception and learning [8]. And finally, we believe that conclusions with respect to affect are extensible to other types of subjective information, such as flavors, styles, motivations, and perceptions, in general. Potential applications for qualitative text mining technology are completely open ended. This paper is an extended version of an earlier paper [9]. Here, we present our work in more detail and present more application examples. Analyzing affect in a text presents us with two obvious sources of ambiguity and imprecision: the first being emotions themselves and the second, words in a natural language [10]. Rather than attempting to constrain and limit this ambiguity, we have taken the opposite approach. We explicitly represent and process ambiguity by introducing fuzzy logic into the picture. Specifically, we integrate basic techniques from fuzzy logic and from computing with words [13] with techniques from natural language processing (NLP). This work is also very related to the recent work on representing and manipulating perceptions [14]. Since the central technique we use from NLP is semantic typing [7], we refer to this approach as fuzzy semantic typing. The fuzzy semantic typing approach is general in scope and can be applied to many different kinds of analysis. We illustrate its use in analyzing affect. At the most basic level, it involves: 1) isolating a vocabulary of words belonging to a meta-linguistic domain (here, affect or emotion); 2) using multiple categorizations and scalar metrics to represent the meaning of each word in that domain; 3) computing profiles for texts based on the categorizations and scores of their component domain words; 4) manipulating the profiles to visualize the texts. We take a multi-faceted approach to representing and manipulating qualitative information. We begin with an affect lexicon, which characterizes a large vocabulary of affect words in terms of a small set of basic categories, such as love, hate, happiness, and anger, each to some numerical degree. The categories from the affect lexicon constitute semantic tags, which are associated with words within a broad semantic domain. In the past, semantic tagging has generally been used for thematic role assignment [3] or for word sense disambiguation (WSD) [4]. Our approach is similar to the standard WSD approach in that the lexicon entry for an ambiguous word represents all of its possible meanings. However, where WSD requires selecting a single meaning for the word in context, our
1063–6706/01$10.00 © 2001 IEEE
484
system simply assigns them all, exploiting rather than reducing the word’s ambiguity. For any relevant word that appears in a text, we include all of its possible meanings and connotations in our analysis of that text; and, we depend on the associated numerical weightings and the cumulative effect of related vocabulary to create a realistic picture of the text’s affective content. Semantic treatments of lexical ambiguity are more typically components of machine translation than of information retrieval (IR) or filtering systems, although there is some evidence that ambiguity resolution can improve performance in IR [6]. Our approach is unusual in integrating lexico-semantic tags into a general-purpose text management system, capable of IR, filtering, categorization, and potentially other text management functions. We have dealt with ambiguity by allowing a single lexicon entry (domain word) to belong to multiple semantic categories. Imprecision is handled, not only via multiple category assignments, but also by allowing degrees of relatedness (centrality) between lexicon entries and their various categories. Gleeful, for example, is assigned to both happiness and excitement, but is given a higher centrality score in the happiness category. In addition to centralities, lexicon entries are also assigned numerical intensities, which represent the strength of the affect level described by that word. Thus, for example, abhorrent and distasteful have roughly the same centrality on the repulsion scale, but abhorrent receives a higher intensity. After the affect words in a document are tagged, the fuzzy logic part of the system handles them by using fuzzy combination operators, set extension operators, and a fuzzy thesaurus to analyze fuzzy sets representing affects. Fuzzy techniques provide an excellent framework for computational management of the ambiguity and imprecision that are pervasive in the words of a natural language. There are additional reasons why fuzzy logic is a good choice for text management. First, eliminating ambiguity and imprecision from texts is unnatural and leads to misconceptions about the underlying meaning. When properly understood and managed, ambiguity and imprecision lead to enhanced, more concrete and precise representations than other (e.g., statistical) methods used for text analysis. This is especially true in the case of qualitative analysis, when we are interested in what essential features are present in some content. And second, since their emphasis is qualitative, fuzzy techniques are more appealing to humans, who tend to think qualitatively; this makes them good candidates for any human-friendly application. The appeal of fuzzy analysis will become apparent when we show visualizations of affect sets later in this document. Besides the fuzzy affect lexicon, we generate additional resources for enhanced functionality. A fuzzy thesaurus is generated from the affect lexicon and used to expand affect sets; affect category groups are generated by clustering the fuzzy thesaurus to enable easier visualization, navigation, and browsing for the user. We provide a detailed explanation of these resources and the ways in which we generate them, with several practical examples, in Section II. A primary representation vehicle in our system is a set of fuzzy semantic categories (affect categories) followed by their centralities and/or intensities, called the affect set. An affect set with attached centralities is always treated as a pure fuzzy set,
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
and all fuzzy techniques applicable to fuzzy sets are applied to such affect sets. The handling of affect sets with intensities is different and more statistical, since, intensities represent less ambiguous, more quantitative features of the text. Affect sets and the fuzzy semantic typing technique are presented in Section III. Finally, visualization is a very important issue in our system. It demonstrates the real power of fuzzy semantic typing by presenting a concise, to-the-point, qualitative representation of affect in texts. Such visualizations constitute an excellent tool for decision making. We show some interesting visualization samples of affect sets for movies and news articles in Section IV. In Section V, we illustrate computational applications of the approach: retrieval of phrases whose affect content is similar to that of a given phrase and filtering of phrases based on a pre-set intensity threshold. In Section VI, we discuss ideas for further development of the system and in Section VII we summarize the paper and its conclusions. II. LINGUISTIC RESOURCES This section describes the linguistic resources of the fuzzy typing system: the affect lexicon, the fuzzy thesaurus and the affect category groups. A. Affect Lexicon The affect lexicon is a compendium of lexical entries for affect words, with their corresponding parts of speech, affect categories, centralities, and intensities. Affect words were gathered from several sources. We began with an existing affect wordlist collected from newspaper articles by Mark Kantrowitz of Justsystem Pittsburgh Research Center. We converted the list to a new format and supplemented it rather haphazardly from an on-line thesaurus. We are currently experimenting with a more systematic, semi-automatic collection method based on WordNet [2], which we hope will prove both scalable and transferable. An affect word is any word having an affect-related meaning or connotation: e.g., abhor, abusive, amity, apprehend, arrogance, etc. Any given affect word may have multiple entries in the affect lexicon, differing by part of speech value and/or category. The expressions of interest are a superset of simple "emotion" words: they include emotions (happiness), feelings (desire), attitudes (resentful), temperament (good-natured), humor (hilarious), frame of mind (cheerful), mood (sulk), spirits (morale), and disposition (sunny). We represent an affect word’s meaning by associating the word with one or more affect categories, from our initial inventory of 83 "atomic" affects. A relatively straightforward affect word, like terror will be associated with a single affect category. An affect word with a more complex meaning will be assigned to multiple categories—for example, infatuation is assigned to both love and insanity. An ambiguous word is simply assigned to all the categories necessary to capture its various meanings: e.g., mad has three entries, associating it with insanity in its first sense and with irritation and anger in its second. All lexicon entries are root forms; forms in the text are part-of-speech tagged and stemmed before lookup.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
Entries in the affect lexicon are of this form
as in Lexical entry is a single entry for a word that has an affectual connotation or denotes an affect directly. At present, our fuzzy affect lexicon contains 3876 lexical entries, about half of what we plan. Part of speech tag. Since ambiguity sometimes depends on a word’s part of speech—and since NLP allows us to differentiate parts of speech in documents— we have included POS information for lexicon entries. For example, the word alert has different category assignments associated with different POS values
That is, the adjective alert means quick to perceive and act—a kind of intelligence—while the verb alert means to call to a state of readiness—a kind of warning. A word’s POS can affect its centrality or intensity values as well as its category assignment. For example, lexicon entries with POS, categories, and centrality degrees for the word craze include
That is, the verb craze belongs to affect category insanity with a degree of 0.8; the singular noun craze belongs to the same category with a degree of 0.5. This reflects the fact that the verb craze means to make insane or as if insane—very central to the insanity category!—while the noun craze means an exaggerated and often transient enthusiasm—i.e., it belongs to insanity only in a less central, more metaphorical sense. Affect category. Many of our categories have strayed somewhat from the strictly affect domain: for example, deprivation, health, and intelligence are only marginally affects, and death, destruction and justice are not affects at all. Such categories have been created in cases where (a) some significant portion of an affect word’s meaning cannot be captured using pure affect categories; and (b) the same meaning component recurred again and again in the vocabulary we were trying to handle. For example, a word like corpse certainly entails some affect, and can plausibly be assigned to categories sadness and horror; at the same time, a part of its meaning is obviously being missed by those categorizations. Moreover, words like assassination, cyanide, execute, funeral, genocide, and homicidal share this missing meaning component. On this first pass, we have gone ahead and created extra, not-strictly-affect categories to handle such words; in the future, when we review and revise the category inventory, we may rethink this decision. At present, there are 83 affect categories. Each affect category has an explicit opposite, with three exceptions: death, irritation and crime. Affect words are spread unevenly across affect categories, with the least frequent categories being health, sickness and facilitation (only 0.35% of entries), and most frequent being
485
conflict and violence (3.7% of entries). The complete list of affect categories with their opposites is given in Appendix A. Centrality. Centrality degrees range from 0 to 1 by increments of 0.1. A word that belongs to several affect categories will generally have different centralities from category to category, as in this example
That is, the element of weakness is fairly central to the word emasculate (a rating of 0.7); the notion of a specific lack is also present but less central (rating of 0.4); and an additional element of violence is possible but not really necessary (rating of 0.3). In assigning centrality, typical questions the lexicon developer should answer for each entry and affect category include: To what extent is affect word X related to category C? To what extent does affect word X co-occur with category C? To what extent can affect word X be replaced with category C in the text, without changing the meaning? Since centralities indicate the presence of certain qualities (represented by appropriate affect categories) in a given affect word, centrality computations are handled as computations of fuzzy membership degrees. Intensity. In addition to centralities, lexicon entries are assigned numerical intensities, which represent the strength of the affect level described by that entry. Intensity degrees, like centrality degrees, range from 0 to 1 by increments of 0.1. Here are some examples (the second number represents the intensity)
All of these words have some element or connotation of repulsion. A word like abhor expresses very intense repulsion (as well as being very central to the concept of repulsion); contempt, aversion, and displeasure are progressively less intense on the repulsion scale. A word like fat—which is not at all central to the repulsion concept, as expressed by its low centrality of 0.2, but which has some slight overtones of repulsion to many Americans—is an objective description, hence, hardly an affect word at all. This is reflected in its low intensity score of 0.1. In general, scores below 0.4 on both scales tend to be the most subjective and notional, since it is easier to rate prominent qualities than backgrounded ones. A word that belongs to several affect categories will generally have different intensities from category to category, as in this example
That is, avenge is a high-intensity conflict word, but only a moderate-intensity word with respect to violence; its intensity rating for justice is somewhere in between.
486
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Assigning category labels and membership degrees to lexicon entries is a very subjective process. During the present proof-ofconcept phase, the assignments have been made by a single linguist. They are obviously influenced by the linguist’s own experience, reading background, and (since affects are in question) personal/emotional background and prejudices. Though subjective, the process is not completely arbitrary—the assignments are general enough in the main to yield useful results. Ideally, however, we would like to involve additional linguists, to review and refine the inventory of atomic categories and to ensure some consensus on the representation of difficult items. In a finished system, repeated iterations and use of additional profiles or personal lexicons will allow the individual user to fine-tune membership degrees and accommodate his or her own subjective criteria. B. Fuzzy Thesaurus The fuzzy thesaurus establishes relationships between pairs of affect categories, based on the centralities of items assigned to both categories in the lexicon. It contains entries of the form
Since it is difficult to modify the affect intensity set consistently to reflect the changes in the affect centrality set, we leave it to the user to accommodate the intensities of the added categories for his/her particular purposes. Expansion increases the number of categories and the level of detail, as shown in Section IV. Note, that since the fuzzy thesaurus is generated from the affect lexicon, it must be recomputed and re-generated whenever the affect lexicon changes. For efficiency, only those entries directly affected by a change are recomputed. C. Affect Category Groups Affect category groups are generated automatically by clustering the fuzzy thesaurus. In this process, affect categories with high degrees of similarity (as defined in the fuzzy thesaurus) are grouped together. For example, we find that love, attraction, happiness, desire and pleasure form one affect category group, while repulsion, horror, inferiority and pain form another. If the automatically-created groups are not as intuitively natural as these examples, the user can edit them. Affect category group ACG is a set of affect categories such that
as in
(3)
arranged in a matrix. When the relationship degree is equal to 0, no entry is recorded in the fuzzy thesaurus. When the relationship degree is equal to 1.0 we say that we have discovered affectual synonyms, as in
Non-synonymous pairs having entries in the matrix are related to some specified degree. The fuzzy thesaurus is generated by the system from the affect lexicon. It is generated using max-min composition [12] (1) are affect categories whose relationship degree we want to compute and represent the centralities of affect categories with reare taken directly from spect to affect . the affect lexicon. The fuzzy thesaurus is primarily used for expansion of affect sets. For example, a single affect category humor with a centrality of 0.7 is expanded using the similarity class for humor from the fuzzy thesaurus. Using max-min composition, we expand this affect intensity profile as follows
where
humor
humor
humor where
excitement
excitement
intelligence
intelligence
represents the composition operator.
(2)
is a user-set threshold. It is worth where noting that ACG is not a similarity relation in the sense of Zadeh [12], since it lacks the transitivity property. The lack of transitivity is a direct consequence of the fact that the ACG is generated from the fuzzy thesaurus, which is in turn generated from the affect lexicon. Transitivity can be enforced by computing the missing relationship degrees from the existing ones. However, we prefer to keep the original relationship degrees intact, in order to ensure the proper interpretation of the original affect category assignments reflected in the affect lexicon. Affect category groups can be used for more efficient groupings of affect categories in visualization charts. An example is shown in Section IV. Since the affect category groups are computed from the fuzzy thesaurus, each time the fuzzy thesaurus is changed, the affect category groups must be recomputed. This is a computationally inexpensive operation, since the number of affect categories involved is typically small—in our prototype, there are only 83 categories. III. FUZZY SEMANTIC TYPING Fuzzy semantic typing is the process in which domain words from a document are identified. The words are assigned meta-information from the typing lexicon in the form of semantic categories and associated degrees; and the categories and degrees are combined to yield the overall representation of the document’s content. In this section, we describe in detail the steps in this process for affect analysis. A. Affect Sets A central construct in our affect analysis is the affect set. It comprises the set of unique affect categories from a given text,
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
487
have something to throw if the audience attacked him). This document is tagged with
Fig. 1. Generation of the document affect set, a fuzzy set representing affective content of a document.
with attached centralities and intensities. The following sections discuss the generation of an affect set for a general document.
B. Tagging of Free Text The process for tagging a document with an affect set is shown in Fig. 1. It includes the following steps. 1) Normalization and Tagging: 1) The document is parsed and tokens (individual words) are generated one at a time. 2) Each token is normalized using normalization rules for English language, shown as Grammar in Fig. 1. 3) The normalized tokens are looked up in the affect lexicon. If a token has one or several lexicon entries, we retrieve all affect categories with their associated centrality and intensity scores. Using this algorithm, we generate the initial affect set for each document. As an example, consider a simple document consisting of this sentence: His first film, Un Chien Andalou (1928), co-directed by Salvador Dali, caused an uproar (he filled his pockets with stones, he wrote in his autobiography, so he would
on the basis of its two affect words, uproar and attacked. Note, that since the word attacked belongs to both the affect categories violence and conflict, both categories are included as document tags. 2) Combination of Centralities and Intensities and Document Affect Set: The following algorithm describes how to reduce the initial affect set by combining the centralities and intensities of recurring categories. 1) For each affect category that appears in the tagging set: a) Compute the maximal centrality (fuzzy union) of all centralities attached to that affect category in the tagged document. The result is the centrality of that category for the document as a whole. b) Compute the average intensity of all intensities attached to that affect category in the tagged document. The result is the intensity of that category for the document as a whole. 2) Counts of affect categories are combined with intensities using simple averaging to yield the overall intensity score for the document. As an example, consider the following document: Luis Bunuel’s The Exterminating Angel (1962) is a macabre comedy, a mordant view of human nature that suggests we harbor savage instincts and unspeakable secrets. Take a group of prosperous dinner guests and pen them up long enough, he suggests, and they’ll turn on one another like rats in an overpopulation study. Bunuel begins with small, alarming portents. The cook and the servants suddenly put on their coats and escape, just as the dinner guests are arriving. The hostess is furious; she planned an after-dinner entertainment involving a bear and two sheep. Now it will have to be canceled. It is typical of Bunuel that such surrealistic touches are dropped in without comment. The dinner party is a success. The guests whisper slanders about each other, their eyes playing across the faces of their fellow guests with greed, lust and envy. After dinner, they stroll into the drawing room, where we glimpse a woman’s purse, filled with chicken feathers and rooster claws. The output produced after fuzzy semantic tagging is given in Table I. We combine recurring affect categories into a set of unique tags, with centralities and intensities that accurately reflect the
488
TABLE I ENTRIES AND ASSOCIATED AFFECT CATEGORIES WITH CENTRALITIES AND INTENSITIES FOR THE MOVIE REVIEW OF EXTERMINATING ANGEL
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
That is to say, the maximal purity of the quality in the document already implies vaguer or more diluted degrees of that quality and, therefore, is appropriate as the combined centrality/purity for that category. The appropriate operation here is thus fuzzy union. On the other hand, the more times an affect category is present in the document and the higher the intensities of its instances, the higher will be the combined intensity/strength attached to it. We, therefore, compute the intensity attached to an affect category as the average of all intensities attached to instances of that category. We believe this model is closer to how humans perceive intensities of words as they read. For example, if we encounter only one instance of a “strong” word (for example, greed in the desire category), together with many “weak” words (for example, wish, want, prefer), the “weak” words will influence the intensity proportionally and reduce the overall effect of the “strong” word. This is why we compute the average rather than some other function, such as the maximum—a maximum would imply that the intensity of the “strongest” word is the intensity for the whole article, which is not what we experience while reading. Another plausible approach would be to weight intensities depending on their position in the text, giving heavier weight to words near the beginning of the document, or belonging to highly exposed parts of the document, such as the title or abstract. After computing centralities using fuzzy union and arranging elements so that the elements with higher membership degrees (centralities) are at the front of the fuzzy set, we have violence, humor, warning, anger, success, slander, greed horror, aversion
absurdity, excitement, desire
pleasure, promise, surfeit
repulsion, fear
lack, death, slyness, intelligence, deception, insanity clarity, innocence, inferiority pain, disloyalty, failure, creation, surprise
overall document content. For this purpose, we discard the original affect words and the POS information, and combine the intensities and centralities of the remaining affect categories. Intensities and centralities are handled differently, since they represent different types of information. Centrality indicates the purity of a quality represented by an affect category. Intensity indicates the strength of that quality. Thus, the number of occurrences of a particular affect category in the document does not affect its centrality, but does affect its intensity. Centrality, as the purity of a quality, depends on the maximal centrality over all instances of that affect category in a particular document.
(4)
This representation of the fuzzy set of affect categories enables us easily to spot predominant qualities of affect categories in the document. The meaning of this affect category set is that the document has a high degree of violence, humor, warning, anger, success, slander, greed, horror, aversion, absurdity, excitement, desire, pleasure, promise and surfeit; a medium degree of repulsion, fear, lack, death, slyness, intelligence, deception, insanity, clarity, innocence and inferiority; and a low degree of pain, disloyalty, failure, creation and surprise. To compute the overall intensity we use a simple average over all affect category instances and their respective intensities (5) where overall intensity of a document ; total number of affect category instances in the document ; intensity of an affect category instance . For the example document, overall intensity is 0.6.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
489
Fig. 2. Centralities with positive connotations up are shown separately from those whose connotations are negative. Affect categories are arranged into similar groups around the circle.
Fig. 3. When all centralities from Fig. 2 are expanded using the fuzzy thesaurus, we get a greater level of detail. Note that additional affect categories exist in the new chart. Centrality values are omitted for greater legibility (scale 0–1).
Overall intensity is used to detect documents with offensive content. For example, high overall intensity (0.7) in combination with the specific centrality profile distaste violence pain may indicate offensive and undesirable content.
IV. AFFECT SET VISUALIZATION An interesting and important area related to the fuzzy typing work is visualization of the results. We have developed a simple affect tagging and affect profile browsing application called the
490
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Fig. 4. News report on train crash in London, October 1999. Centralities describe quality, and intensities, quantity (count and strength) of the affects in a document. As might be expected, the affects fear, harm, pain and surprise are most central. The most intense affects are conflict, confusion, disadvantage and pain.
Fig. 5. Profile of a recent cult movie Matrix generated from a movie review. Opposite affect categories are placed on opposite sides of the circle (the left side shows negative affects, the right side, positive). With a well-developed negative side, this movie is deservedly rated “R” in the US.
Affect Inspector. We show a basic browsing setup from that application in Appendix B. Each affect category’s centralities and intensities can be represented as a point on the perimeter of a unit circle. Centralities and intensities can then be visualized, as shown in Figs. 2–8. In order to demonstrate various ways these charts can be organized, we show different charts for different information objects. In Fig. 2 we show centralities with positive affects separated from centralities with mainly negative affects, for the movie review of Exterminating Angel. In this way, the positive versus the negative side of the document can be easily analyzed. Moreover, groups of similar affect categories are shown close to each other. This facilitates a quick overview of aspects denoted by those groups. For example, in Fig. 2, clarity, creation and intelligence are not as well developed as success, desire, pleasure,
humor and excitement. Groups of similar affect categories are generated using the technique discussed in Section II-C. When all affect categories from Fig. 2 are expanded using the fuzzy thesaurus, we obtain the chart in Fig. 3. The chart contains both positive and negative affects, with a higher level of detail, since new affect categories have been added to the chart through expansion. In Fig. 4 we show the affect structure of a news report concerning a train crash in London. We show both centralities and intensities for qualities typical of news on accidents. Opposing affect categories can be placed on opposite sides of the circle with respect to the center point. This is illustrated in Fig. 5, for centralities in a Matrix review. With this arrangement, we can easily spot which part of the circle is best developed and understand its affective impact.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
491
Fig. 6. Centralities (up) and intensities in T. S. Eliot’s famous “Rhapsody on a Windy Night.” Even the surrealistic content of this poem can be succesfully analyzed. The poem contains mixed affects of weakness, disadvantage, insanity and strength, clarity and openness. For intensities, clarity, weakness, and inferiority seem to prevail.
Affect categories can be generated for sets of movies. We analyzed sets in the romance, action, science fiction, comedy, and family genres. Each set contains about 15 movies and some movies belong to multiple sets (for example, 12 Monkeys belongs both to the action and science fiction genres). The results are shown in Table II, along with profiles for news about accidents. The generated profiles confirm our expectations about affect categories for different movie genres: romance movies showed high levels of happiness, innocence and justice; action movies scored high on surprise, strength and slyness; family movies emphasized sensitivity, morality, responsibility and nurturance; and so on. The composite profiles also reveal some less obvious features: high levels of immorality and inferiority in the romance movie set, and high levels of inferiority and destruction in the comedies, for example. Although the results in many cases represent typical characteristics of movie genres, we believe that some aspects of the results may reflect only the selected movies within the genre. Next, we show that poetry can be analyzed using this approach. Fig. 6 shows centralities and intensities for a famous poem with a very accurate affect profile. Another application of this technique is personalization. We generated affect profiles for different users based on their movie preferences and the results clearly reflect different personal tastes. The illustration in Fig. 7 shows the centrality affect profiles obtained after merging profiles for each person’s favorite movies. In merging centrality affect profiles, we use the same approach as when merging affect profiles of individual words or sentences, that is, the union (maximum) operator. V. COMPUTING WITH AFFECT PROFILES To illustrate the computational potential of the fuzzy semantic typing framework, we present here an example of retrieval based on similarities drawn from affect categories. We illustrate the technique on retrieval of similar phrases, but it extends equally
TABLE II CENTRALITY RESULTS BY TYPE
492
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
TABLE III PHRASE RETRIEVAL
well to retrieval of larger portions of texts, such as sentences or paragraphs. Here is the outline of the experiment from which we produced Table III. 1) We found the affect profile of the sentence fear, anger, grief and pain filled the room. 2) We retrieved all phrases from a precompiled list of 34 phrases, from four different documents that had affect profiles most similar to the seed phrase. As a similarity measure between the affect profile of the seed phrase and the affect profile of a candidate phrase , we use [15] (6) represents the sum of the intersections (minwhere imum values) of the affect sets’ centralities for the respective represents the sum of the unions affect categories and (maximum values) of the affect sets’ centralities for the respective affect categories. We found phrases as presented in Table III, ordered by decreasing scores for similarity to the seed phrase. Each phrase is shown with its respective affect centrality and intensity profiles, similarity degrees, and average intensities.
This retrieval-by-similarity technique can be complemented with the filtering-on-intensity technique. For example, we can set the intensity threshold to a certain category’s intensity, and filter out all expressions whose intensity for that category is greater than the threshold intensity. Alternatively, we can set the threshold on overall intensity and filter out all documents with intensities higher than the given threshold. Such similarity computations can be conveniently combined with statistical methods. For short phrases like those shown here, the similarity measure we used is very convenient since it reflects the overlapping of qualitative features (affect categories) without taking into account statistical features like number of words or categories. In certain cases, it may be reasonable to use a vector space model with cosine similarity to find similar phrases. One must be careful, however, as that approach will not always return the desired qualitative profile. To illustrate this point, let us imagine that we have submitted the query love, pain to a text corpus and that we are searching for sentences that have both qualities—i.e., sentences that match both love and pain. However, vector space retrieval would return sentences containing many instances of love with a high score, even if there is no mention of pain in them. Our similarity computation, on the other hand, gives preference
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
493
Fig. 7. Affect profiles generated from personal movie preferences. Each person submitted from 10 to 81 favorite movies. Individual movie profiles are merged using the maximum operator on centralities to yield the personal profiles in this figure.
to documents having higher-level maximal centralities for both categories, regardless of their total count in the sentence. This example clearly illustrates the main differences in the quantitative vs. the qualitative approach to love and pain. In the quantitative approach, we are concerned with finding information that is statistically similar to our query. In the qualitative approach, we are more concerned with a particular qualitative profile of the target information. VI. FURTHER DEVELOPMENT The fuzzy semantic typing approach deals well with ambiguity and imprecision in free text. It can be effectively combined with a set of visualization tools for easy, accurate analysis of the affect content in a document. These results are promising, but we feel that we have just begun exploration in uncharted territory. Our plans in the immediate future include: 1) Enrichment of the existing affect typing mechanism. To the extent that expressions in free text and their constituent words have syntactically and semantically different roles, we are making an approximation by using fuzzy union to compute centralities. Although individual affect words are treated equally, their roles are frequently uneven. For example, modifiers (very, more or less, not), comparative and superlative adjectives, adverbial phrases, noun phrases, and complex phrases (with and/or connectives, etc.) all represent different classes of expressions with possibly different roles in a phrase or sentence. A finer-grained analysis
would, therefore, lead to different centrality combination operators for each of these classes. The same holds true for the hierarchical centrality computation, but to a lesser extent, since text hierarchies are more regular and do not have as richly varying a structure. Still, for many documents, sentences and paragraphs can be assigned weights that reflect their relative importance in centrality computations. For example, since report-style prose is typically “front-loaded,” a possible approach would be to increase the centrality of categories appearing in the title, summary, or lead paragraph of such texts. 2) Support for management of linguistic resources. We would like to begin experimenting with personal (user-initiated) updating of general-purpose affect lexicons. This would include the modification of centralities and intensities attached to affect categories, the addition or removal of affect words, the definition of complex affect categories in terms of basic affect categories, and the tuning of the fuzzy thesaurus to reflect changes in the affect lexicon. It would be especially interesting to experiment with different affect lexicons on the sending and receiving sides of a communication. In such an experiment, a message composed with a help of personal affect lexicon would be interpreted using personal lexicon . 3) Generalization of the fuzzy typing framework. Fuzzy typing may be adapted to many different application areas by developing appropriate resources: busi-
494
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Fig. 8. Basic browsing setup in TreeViewer. The tree view at upper right shows a browsable text hierarchy; the text pane at lower right displays analyzed text. Affect centrality and affect intensity sets appear on the left. Circles represent average values.
ness, food/cooking, fashion, architectural styles, cultural events and artistic material, psychology-related material, wine and beer, perfumes. 4) Analysis of different points of view. Using an affect lexicon, we can analyze affects in a text. Using a lexicon with expressions and types that describe intentions (e.g., would like to, will, is considering, is thinking about) would give us an intention fingerprint for a document. Using both lexicons at the same time, to find a text’s affectual and intentional
fingerprints, could reveal interesting juxtapositions in the data. 5) Integration with existing text management techniques. Although not currently included in our system, quasistatistical information can be generated to complement qualitative analysis. This would involve a simple extension of common statistical approaches like statistical indexing, term-weighting, IDF-TF scoring and cosine distance [5], on a full feature space (all terms and phrases), or on
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
495
TABLE IV THE COMPLETE LIST OF AFFECT CATEGORIES AND RESPECTIVE OPPOSITE AFFECT CATEGORIES
a significantly reduced feature space consisting of a limited number of affect categories. Fuzzy typing would then be complemented with retrieval techniques applied to affect profiles: e.g., querying and retrieval, clustering and classification. We believe the best way to approach this issue is to start with basic similarity assumptions as given in [11], and further investigate techniques from [12] and [15] for computing similarities between fuzzy objects. Finally, we hope to devise an algorithm that effectively combines statistical and quasistatistical similarity scores with fuzzy similarity scores to compute the overall similarity of two texts.
6) Extension to additional languages. While the lexicon for our prototype system was created manually, we are currently experimenting with a more systematic, semi-automatic collection method based on WordNet [2]. If this technique proves successful, it could be applied to foreign language semantic nets such as EuroWordNet [1], to accommodate a variety of other languages. VII. CONCLUSION We describe a novel approach to text analysis that combines semantic typing techniques from natural language processing
496
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
with fuzzy techniques, under the common framework of fuzzy semantic typing. Fuzzy semantic typing is an innovative way to capture metalinguistic facts about a text while allowing for linguistic ambiguity and vagueness. From our analysis of generated profiles for news reports and movie reviews, we believe that the metalinguistic representation can be usefully applied in retrieval, clustering, and classification. The approach is applicable to an indefinite number of domains and lends itself to customization for a particular user or task. We look forward to continuing our research in these directions. APPENDIX A See Table IV. APPENDIX B AFFECT INSPECTOR APPLICATION In Fig. 8 we show an example screen from our affect inspector application, illustrating the display of a profile’s centralities and intensities at different levels of the text hierarchy. The essential part of the browsing system is the TreeViewer. It is a Java/XML application that contains a text hierarchy in the form of a tree. Each node in the tree is clickable, and clicking on it displays the affect profile associated with that node in two panes on the left—the centrality profile in the upper pane, the intensity profile in the lower. The circles represent the average values for centralities (upper left) and intensities (lower left). The text associated with the node is displayed in the lower right pane. Browsing is possible at any level of the hierarchy from top to bottom: corpus, document, paragraph, sentence, affect category. Affect profiles to the left are click-sensitive maps: by clicking on certain points, one can search on individual affects, or change the form of the current display to see opposite category placements, linear placements, or placements ordered by decreasing scores. Such arrangements facilitate easy browsing, verification of affect profiles, and comparison of affect profiles for different text elements, both vertically (affect-sentence-paragraph-document-corpus) and laterally (two or more documents, paragraphs or sentences). REFERENCES [1] University of Amsterdam, Dep. Computional Linguistics EuroWordNet, . (2000, May). [Online]. Available: http://www.hum.uva.nl/~ewn/ [2] C. Fellbaum, Ed., WordNet: An Electronic Lexical Database: MIT Press, 1998. [3] C. Fillmore and B. T. S. Atkins, “FrameNet and lexicographic relevance,” in Proc. First Int. Conf. on Language Resources and Evaluation , 1998, pp. 417–420. [4] T. Fontanelle, “Semantic tagging: A survey,” in Papers in Computational Lexicography, COMPLEX 99, 1999, pp. 39–56.
[5] D. A. Grossman and O. Frieder, Information Retrieval Algorithms and Heuristics. Norwell, MA: Kluwer, 1998. [6] R. Krovetz and W. B. Croft, “Lexical ambiguity and information retrieval,” ACM Trans. Inform. Syst., vol. 10, no. 2, pp. 115–141, 1992. [7] G. Miller and C. Walter, “Contextual correlates of semantic similarity,” Language and Cognitive Processes, vol. 6, pp. 1–28, 1991. [8] R. W. Picard, Affective Computing. Cambridge, MA: MIT Press, 1997. [9] P. Subasic and A. Huettner, “Affect analysis of text using fuzzy semantic typing,” presented at the Proc. of FUZZ-IEEE 2000, The 9th International Conference on Fuzzy Systems, San Antonio, Taxas, 2000. [10] M. Sugeno, “On organization of imprecision based on word classification,” in 2nd Fuzzy System Symposium, 1986, pp. 148–153. Japanese. [11] A. Tversky, “Features of similarity,” Psychological Rev. 84, pp. 327–352, 1977. [12] L. A. Zadeh, “Similarity relations and fuzzy orderings,” in Inform. Sci. 3: Elsevier Sci., 1977, pp. 177–200. , “Fuzzy logic computing with words,” IEEE Trans. Fuzzy Syst., [13] pp. 103–111, 1996. [14] , “A new AI: Toward computational theory of perceptions,” AAAI Magazine, vol. 22, no. 1, pp. 73–84, Spring 2001. [15] R. Zwick, E. Carlstein, and D. V. Budescu, “Measures of similarity among fuzzy concepts: A comparative analysis,” International Journal of Approximate Reasoning, vol. 1, pp. 221–242, 1987.
=
Pero Subasic (M’97) obtained the Dipl. Eng. and M.S. degrees from the School of Electrical Engineering, Begrade University, Yugoslavia, and the Dr. Eng. from Yamagata University, Japan, in 1989, 1993, and 1996, respectively He has been with the Institute Mihajlo Pupin, Belgrade, Yugoslavia; Belgrade University, Yugoslavia; Tohoku University, Japan; Yamagata University, Japan; and, the Tokyo Insitute of Technology, Japan, as a Faculty Member or Researcher. He is currently with Clairvoyance, Corporation, Pittsburgh, PA, where he has worked in text mining, navigation and visualization, data analysis, preprocessing, including semantic Typing and Resource Management. He is the Principal Inventor of the Fuzzy Semantic Typing framework. He has published over 40 research papers and reports in international journals, monographs and conferences. He is author of a book on fuzzy systems and neural networks. Dr. Subasic is a member of Japan Society for Fuzzy Systems (ACM SIGCHI, SOFT) and Yugoslav Society for Soft Computing and Intelligent Systems (SOCOIS).
Alison Huettner received the Ph.D. in linguistics from the University of Massachusetts, Amherst, MA, in 1989. In 1998, she joined Clairvoyance Corporation, Pittsburgh, PA as a Project Manager working on natural language processing (NLP). She refined the CLARIT NLP resources and experimented with extensions, including expanded lexical equivalences, semantic typing, specialized affect handling, and a prototype question-answering system; currently, she is working with e-commerce applications. Prior to joining Clairvoyance, she was a knowledge engineer with Carnegie Group, Inc., and worked on a fact-extraction system and a commercial machine translation system. She is also one of the inventors of the associated patent for an integrated authoring and translation system.
483
Affect Analysis of Text Using Fuzzy Semantic Typing Pero Subasic, Member, IEEE, and Alison Huettner
Abstract—We propose a novel, convenient fusion of natural language processing and fuzzy logic techniques for analyzing the affect content in free text. Our main goals are fast analysis and visualization of affect content for decision making. The main linguistic resource for fuzzy semantic typing is the fuzzy-affect lexicon, from which other important resources—the fuzzy thesaurus and affect category groups—are generated. Free text is tagged with affect categories from the lexicon and the affect categories’ centralities and intensities are combined using techniques from fuzzy logic to produce affect sets—fuzzy sets representing the affect quality of a document. We show different aspects of affect analysis using news content and movie reviews. Our experiments show a good correspondence between affect sets and human judgments of affect content. We ascribe this to the representation of ambiguity in our fuzzy affect lexicon and the ability of fuzzy logic to deal successfully with the ambiguity of words in a natural language. Planned extensions of the system include personalized profiles for Web-based content dissemination, fuzzy retrieval, clustering, and classification. Index Terms—Computing with words, fuzzy logic, knowledge engineering, text mining, world wide web.
I. INTRODUCTION
T
HE huge amount of text stored on computer systems is getting larger every day. Moving beyond the basic assumption that a given piece of text should be easily located, the next generation of systems aims toward integrated, personalized services and decision support. In these areas, a quick analysis of particular qualities in the text and an intuitive presentation to the user become increasingly important. To match an individual user’s profile on the World Wide Web, for example, it is necessary to introduce a human dimension into text understanding and representation. Expectations for modeling the purely subjective, human dimensions of text and data understanding are high, on both the users’ and providers’ sides. Here we experiment with qualitative analysis of affect-related information in free text. Affect-related information includes words describing emotions; like fear, anger, love, joy, and sorrow; feelings like warmth and excitement; attitudes like helpful, friendly, hostile; and other related categories such as temperament, humor, frame of mind, mood, spirits, morale, and disposition (including words like sweet, farce, wary, sanguine, depressed, eagerness, selfish). Our reasons for selecting this particular domain are threefold. First, affect-related information is pervasive in electronic documents: in news stories (on politics, competitive sports, etc.), economic reports (corporate acquisitions, investor reactions), Manuscript received September 11, 2000; revised May 8, 2001. The authors are with the Clairvoyance Corporation, Pittsburgh, PA 15232 USA (e-mail: [email protected]). Publisher Item Identifier S 1063-6706(01)06537-7.
customer-oriented Internet sites (eBay.com, amazon.com), newsgroups (misc.consumer), corporate customer service, email (customer complaints, questions, opinions), artistic and cultural material (movie and art reviews), etc. Second, affect information is critical to human communication: recent work shows the importance of emotions in decision-making, perception and learning [8]. And finally, we believe that conclusions with respect to affect are extensible to other types of subjective information, such as flavors, styles, motivations, and perceptions, in general. Potential applications for qualitative text mining technology are completely open ended. This paper is an extended version of an earlier paper [9]. Here, we present our work in more detail and present more application examples. Analyzing affect in a text presents us with two obvious sources of ambiguity and imprecision: the first being emotions themselves and the second, words in a natural language [10]. Rather than attempting to constrain and limit this ambiguity, we have taken the opposite approach. We explicitly represent and process ambiguity by introducing fuzzy logic into the picture. Specifically, we integrate basic techniques from fuzzy logic and from computing with words [13] with techniques from natural language processing (NLP). This work is also very related to the recent work on representing and manipulating perceptions [14]. Since the central technique we use from NLP is semantic typing [7], we refer to this approach as fuzzy semantic typing. The fuzzy semantic typing approach is general in scope and can be applied to many different kinds of analysis. We illustrate its use in analyzing affect. At the most basic level, it involves: 1) isolating a vocabulary of words belonging to a meta-linguistic domain (here, affect or emotion); 2) using multiple categorizations and scalar metrics to represent the meaning of each word in that domain; 3) computing profiles for texts based on the categorizations and scores of their component domain words; 4) manipulating the profiles to visualize the texts. We take a multi-faceted approach to representing and manipulating qualitative information. We begin with an affect lexicon, which characterizes a large vocabulary of affect words in terms of a small set of basic categories, such as love, hate, happiness, and anger, each to some numerical degree. The categories from the affect lexicon constitute semantic tags, which are associated with words within a broad semantic domain. In the past, semantic tagging has generally been used for thematic role assignment [3] or for word sense disambiguation (WSD) [4]. Our approach is similar to the standard WSD approach in that the lexicon entry for an ambiguous word represents all of its possible meanings. However, where WSD requires selecting a single meaning for the word in context, our
1063–6706/01$10.00 © 2001 IEEE
484
system simply assigns them all, exploiting rather than reducing the word’s ambiguity. For any relevant word that appears in a text, we include all of its possible meanings and connotations in our analysis of that text; and, we depend on the associated numerical weightings and the cumulative effect of related vocabulary to create a realistic picture of the text’s affective content. Semantic treatments of lexical ambiguity are more typically components of machine translation than of information retrieval (IR) or filtering systems, although there is some evidence that ambiguity resolution can improve performance in IR [6]. Our approach is unusual in integrating lexico-semantic tags into a general-purpose text management system, capable of IR, filtering, categorization, and potentially other text management functions. We have dealt with ambiguity by allowing a single lexicon entry (domain word) to belong to multiple semantic categories. Imprecision is handled, not only via multiple category assignments, but also by allowing degrees of relatedness (centrality) between lexicon entries and their various categories. Gleeful, for example, is assigned to both happiness and excitement, but is given a higher centrality score in the happiness category. In addition to centralities, lexicon entries are also assigned numerical intensities, which represent the strength of the affect level described by that word. Thus, for example, abhorrent and distasteful have roughly the same centrality on the repulsion scale, but abhorrent receives a higher intensity. After the affect words in a document are tagged, the fuzzy logic part of the system handles them by using fuzzy combination operators, set extension operators, and a fuzzy thesaurus to analyze fuzzy sets representing affects. Fuzzy techniques provide an excellent framework for computational management of the ambiguity and imprecision that are pervasive in the words of a natural language. There are additional reasons why fuzzy logic is a good choice for text management. First, eliminating ambiguity and imprecision from texts is unnatural and leads to misconceptions about the underlying meaning. When properly understood and managed, ambiguity and imprecision lead to enhanced, more concrete and precise representations than other (e.g., statistical) methods used for text analysis. This is especially true in the case of qualitative analysis, when we are interested in what essential features are present in some content. And second, since their emphasis is qualitative, fuzzy techniques are more appealing to humans, who tend to think qualitatively; this makes them good candidates for any human-friendly application. The appeal of fuzzy analysis will become apparent when we show visualizations of affect sets later in this document. Besides the fuzzy affect lexicon, we generate additional resources for enhanced functionality. A fuzzy thesaurus is generated from the affect lexicon and used to expand affect sets; affect category groups are generated by clustering the fuzzy thesaurus to enable easier visualization, navigation, and browsing for the user. We provide a detailed explanation of these resources and the ways in which we generate them, with several practical examples, in Section II. A primary representation vehicle in our system is a set of fuzzy semantic categories (affect categories) followed by their centralities and/or intensities, called the affect set. An affect set with attached centralities is always treated as a pure fuzzy set,
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
and all fuzzy techniques applicable to fuzzy sets are applied to such affect sets. The handling of affect sets with intensities is different and more statistical, since, intensities represent less ambiguous, more quantitative features of the text. Affect sets and the fuzzy semantic typing technique are presented in Section III. Finally, visualization is a very important issue in our system. It demonstrates the real power of fuzzy semantic typing by presenting a concise, to-the-point, qualitative representation of affect in texts. Such visualizations constitute an excellent tool for decision making. We show some interesting visualization samples of affect sets for movies and news articles in Section IV. In Section V, we illustrate computational applications of the approach: retrieval of phrases whose affect content is similar to that of a given phrase and filtering of phrases based on a pre-set intensity threshold. In Section VI, we discuss ideas for further development of the system and in Section VII we summarize the paper and its conclusions. II. LINGUISTIC RESOURCES This section describes the linguistic resources of the fuzzy typing system: the affect lexicon, the fuzzy thesaurus and the affect category groups. A. Affect Lexicon The affect lexicon is a compendium of lexical entries for affect words, with their corresponding parts of speech, affect categories, centralities, and intensities. Affect words were gathered from several sources. We began with an existing affect wordlist collected from newspaper articles by Mark Kantrowitz of Justsystem Pittsburgh Research Center. We converted the list to a new format and supplemented it rather haphazardly from an on-line thesaurus. We are currently experimenting with a more systematic, semi-automatic collection method based on WordNet [2], which we hope will prove both scalable and transferable. An affect word is any word having an affect-related meaning or connotation: e.g., abhor, abusive, amity, apprehend, arrogance, etc. Any given affect word may have multiple entries in the affect lexicon, differing by part of speech value and/or category. The expressions of interest are a superset of simple "emotion" words: they include emotions (happiness), feelings (desire), attitudes (resentful), temperament (good-natured), humor (hilarious), frame of mind (cheerful), mood (sulk), spirits (morale), and disposition (sunny). We represent an affect word’s meaning by associating the word with one or more affect categories, from our initial inventory of 83 "atomic" affects. A relatively straightforward affect word, like terror will be associated with a single affect category. An affect word with a more complex meaning will be assigned to multiple categories—for example, infatuation is assigned to both love and insanity. An ambiguous word is simply assigned to all the categories necessary to capture its various meanings: e.g., mad has three entries, associating it with insanity in its first sense and with irritation and anger in its second. All lexicon entries are root forms; forms in the text are part-of-speech tagged and stemmed before lookup.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
Entries in the affect lexicon are of this form
as in Lexical entry is a single entry for a word that has an affectual connotation or denotes an affect directly. At present, our fuzzy affect lexicon contains 3876 lexical entries, about half of what we plan. Part of speech tag. Since ambiguity sometimes depends on a word’s part of speech—and since NLP allows us to differentiate parts of speech in documents— we have included POS information for lexicon entries. For example, the word alert has different category assignments associated with different POS values
That is, the adjective alert means quick to perceive and act—a kind of intelligence—while the verb alert means to call to a state of readiness—a kind of warning. A word’s POS can affect its centrality or intensity values as well as its category assignment. For example, lexicon entries with POS, categories, and centrality degrees for the word craze include
That is, the verb craze belongs to affect category insanity with a degree of 0.8; the singular noun craze belongs to the same category with a degree of 0.5. This reflects the fact that the verb craze means to make insane or as if insane—very central to the insanity category!—while the noun craze means an exaggerated and often transient enthusiasm—i.e., it belongs to insanity only in a less central, more metaphorical sense. Affect category. Many of our categories have strayed somewhat from the strictly affect domain: for example, deprivation, health, and intelligence are only marginally affects, and death, destruction and justice are not affects at all. Such categories have been created in cases where (a) some significant portion of an affect word’s meaning cannot be captured using pure affect categories; and (b) the same meaning component recurred again and again in the vocabulary we were trying to handle. For example, a word like corpse certainly entails some affect, and can plausibly be assigned to categories sadness and horror; at the same time, a part of its meaning is obviously being missed by those categorizations. Moreover, words like assassination, cyanide, execute, funeral, genocide, and homicidal share this missing meaning component. On this first pass, we have gone ahead and created extra, not-strictly-affect categories to handle such words; in the future, when we review and revise the category inventory, we may rethink this decision. At present, there are 83 affect categories. Each affect category has an explicit opposite, with three exceptions: death, irritation and crime. Affect words are spread unevenly across affect categories, with the least frequent categories being health, sickness and facilitation (only 0.35% of entries), and most frequent being
485
conflict and violence (3.7% of entries). The complete list of affect categories with their opposites is given in Appendix A. Centrality. Centrality degrees range from 0 to 1 by increments of 0.1. A word that belongs to several affect categories will generally have different centralities from category to category, as in this example
That is, the element of weakness is fairly central to the word emasculate (a rating of 0.7); the notion of a specific lack is also present but less central (rating of 0.4); and an additional element of violence is possible but not really necessary (rating of 0.3). In assigning centrality, typical questions the lexicon developer should answer for each entry and affect category include: To what extent is affect word X related to category C? To what extent does affect word X co-occur with category C? To what extent can affect word X be replaced with category C in the text, without changing the meaning? Since centralities indicate the presence of certain qualities (represented by appropriate affect categories) in a given affect word, centrality computations are handled as computations of fuzzy membership degrees. Intensity. In addition to centralities, lexicon entries are assigned numerical intensities, which represent the strength of the affect level described by that entry. Intensity degrees, like centrality degrees, range from 0 to 1 by increments of 0.1. Here are some examples (the second number represents the intensity)
All of these words have some element or connotation of repulsion. A word like abhor expresses very intense repulsion (as well as being very central to the concept of repulsion); contempt, aversion, and displeasure are progressively less intense on the repulsion scale. A word like fat—which is not at all central to the repulsion concept, as expressed by its low centrality of 0.2, but which has some slight overtones of repulsion to many Americans—is an objective description, hence, hardly an affect word at all. This is reflected in its low intensity score of 0.1. In general, scores below 0.4 on both scales tend to be the most subjective and notional, since it is easier to rate prominent qualities than backgrounded ones. A word that belongs to several affect categories will generally have different intensities from category to category, as in this example
That is, avenge is a high-intensity conflict word, but only a moderate-intensity word with respect to violence; its intensity rating for justice is somewhere in between.
486
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Assigning category labels and membership degrees to lexicon entries is a very subjective process. During the present proof-ofconcept phase, the assignments have been made by a single linguist. They are obviously influenced by the linguist’s own experience, reading background, and (since affects are in question) personal/emotional background and prejudices. Though subjective, the process is not completely arbitrary—the assignments are general enough in the main to yield useful results. Ideally, however, we would like to involve additional linguists, to review and refine the inventory of atomic categories and to ensure some consensus on the representation of difficult items. In a finished system, repeated iterations and use of additional profiles or personal lexicons will allow the individual user to fine-tune membership degrees and accommodate his or her own subjective criteria. B. Fuzzy Thesaurus The fuzzy thesaurus establishes relationships between pairs of affect categories, based on the centralities of items assigned to both categories in the lexicon. It contains entries of the form
Since it is difficult to modify the affect intensity set consistently to reflect the changes in the affect centrality set, we leave it to the user to accommodate the intensities of the added categories for his/her particular purposes. Expansion increases the number of categories and the level of detail, as shown in Section IV. Note, that since the fuzzy thesaurus is generated from the affect lexicon, it must be recomputed and re-generated whenever the affect lexicon changes. For efficiency, only those entries directly affected by a change are recomputed. C. Affect Category Groups Affect category groups are generated automatically by clustering the fuzzy thesaurus. In this process, affect categories with high degrees of similarity (as defined in the fuzzy thesaurus) are grouped together. For example, we find that love, attraction, happiness, desire and pleasure form one affect category group, while repulsion, horror, inferiority and pain form another. If the automatically-created groups are not as intuitively natural as these examples, the user can edit them. Affect category group ACG is a set of affect categories such that
as in
(3)
arranged in a matrix. When the relationship degree is equal to 0, no entry is recorded in the fuzzy thesaurus. When the relationship degree is equal to 1.0 we say that we have discovered affectual synonyms, as in
Non-synonymous pairs having entries in the matrix are related to some specified degree. The fuzzy thesaurus is generated by the system from the affect lexicon. It is generated using max-min composition [12] (1) are affect categories whose relationship degree we want to compute and represent the centralities of affect categories with reare taken directly from spect to affect . the affect lexicon. The fuzzy thesaurus is primarily used for expansion of affect sets. For example, a single affect category humor with a centrality of 0.7 is expanded using the similarity class for humor from the fuzzy thesaurus. Using max-min composition, we expand this affect intensity profile as follows
where
humor
humor
humor where
excitement
excitement
intelligence
intelligence
represents the composition operator.
(2)
is a user-set threshold. It is worth where noting that ACG is not a similarity relation in the sense of Zadeh [12], since it lacks the transitivity property. The lack of transitivity is a direct consequence of the fact that the ACG is generated from the fuzzy thesaurus, which is in turn generated from the affect lexicon. Transitivity can be enforced by computing the missing relationship degrees from the existing ones. However, we prefer to keep the original relationship degrees intact, in order to ensure the proper interpretation of the original affect category assignments reflected in the affect lexicon. Affect category groups can be used for more efficient groupings of affect categories in visualization charts. An example is shown in Section IV. Since the affect category groups are computed from the fuzzy thesaurus, each time the fuzzy thesaurus is changed, the affect category groups must be recomputed. This is a computationally inexpensive operation, since the number of affect categories involved is typically small—in our prototype, there are only 83 categories. III. FUZZY SEMANTIC TYPING Fuzzy semantic typing is the process in which domain words from a document are identified. The words are assigned meta-information from the typing lexicon in the form of semantic categories and associated degrees; and the categories and degrees are combined to yield the overall representation of the document’s content. In this section, we describe in detail the steps in this process for affect analysis. A. Affect Sets A central construct in our affect analysis is the affect set. It comprises the set of unique affect categories from a given text,
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
487
have something to throw if the audience attacked him). This document is tagged with
Fig. 1. Generation of the document affect set, a fuzzy set representing affective content of a document.
with attached centralities and intensities. The following sections discuss the generation of an affect set for a general document.
B. Tagging of Free Text The process for tagging a document with an affect set is shown in Fig. 1. It includes the following steps. 1) Normalization and Tagging: 1) The document is parsed and tokens (individual words) are generated one at a time. 2) Each token is normalized using normalization rules for English language, shown as Grammar in Fig. 1. 3) The normalized tokens are looked up in the affect lexicon. If a token has one or several lexicon entries, we retrieve all affect categories with their associated centrality and intensity scores. Using this algorithm, we generate the initial affect set for each document. As an example, consider a simple document consisting of this sentence: His first film, Un Chien Andalou (1928), co-directed by Salvador Dali, caused an uproar (he filled his pockets with stones, he wrote in his autobiography, so he would
on the basis of its two affect words, uproar and attacked. Note, that since the word attacked belongs to both the affect categories violence and conflict, both categories are included as document tags. 2) Combination of Centralities and Intensities and Document Affect Set: The following algorithm describes how to reduce the initial affect set by combining the centralities and intensities of recurring categories. 1) For each affect category that appears in the tagging set: a) Compute the maximal centrality (fuzzy union) of all centralities attached to that affect category in the tagged document. The result is the centrality of that category for the document as a whole. b) Compute the average intensity of all intensities attached to that affect category in the tagged document. The result is the intensity of that category for the document as a whole. 2) Counts of affect categories are combined with intensities using simple averaging to yield the overall intensity score for the document. As an example, consider the following document: Luis Bunuel’s The Exterminating Angel (1962) is a macabre comedy, a mordant view of human nature that suggests we harbor savage instincts and unspeakable secrets. Take a group of prosperous dinner guests and pen them up long enough, he suggests, and they’ll turn on one another like rats in an overpopulation study. Bunuel begins with small, alarming portents. The cook and the servants suddenly put on their coats and escape, just as the dinner guests are arriving. The hostess is furious; she planned an after-dinner entertainment involving a bear and two sheep. Now it will have to be canceled. It is typical of Bunuel that such surrealistic touches are dropped in without comment. The dinner party is a success. The guests whisper slanders about each other, their eyes playing across the faces of their fellow guests with greed, lust and envy. After dinner, they stroll into the drawing room, where we glimpse a woman’s purse, filled with chicken feathers and rooster claws. The output produced after fuzzy semantic tagging is given in Table I. We combine recurring affect categories into a set of unique tags, with centralities and intensities that accurately reflect the
488
TABLE I ENTRIES AND ASSOCIATED AFFECT CATEGORIES WITH CENTRALITIES AND INTENSITIES FOR THE MOVIE REVIEW OF EXTERMINATING ANGEL
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
That is to say, the maximal purity of the quality in the document already implies vaguer or more diluted degrees of that quality and, therefore, is appropriate as the combined centrality/purity for that category. The appropriate operation here is thus fuzzy union. On the other hand, the more times an affect category is present in the document and the higher the intensities of its instances, the higher will be the combined intensity/strength attached to it. We, therefore, compute the intensity attached to an affect category as the average of all intensities attached to instances of that category. We believe this model is closer to how humans perceive intensities of words as they read. For example, if we encounter only one instance of a “strong” word (for example, greed in the desire category), together with many “weak” words (for example, wish, want, prefer), the “weak” words will influence the intensity proportionally and reduce the overall effect of the “strong” word. This is why we compute the average rather than some other function, such as the maximum—a maximum would imply that the intensity of the “strongest” word is the intensity for the whole article, which is not what we experience while reading. Another plausible approach would be to weight intensities depending on their position in the text, giving heavier weight to words near the beginning of the document, or belonging to highly exposed parts of the document, such as the title or abstract. After computing centralities using fuzzy union and arranging elements so that the elements with higher membership degrees (centralities) are at the front of the fuzzy set, we have violence, humor, warning, anger, success, slander, greed horror, aversion
absurdity, excitement, desire
pleasure, promise, surfeit
repulsion, fear
lack, death, slyness, intelligence, deception, insanity clarity, innocence, inferiority pain, disloyalty, failure, creation, surprise
overall document content. For this purpose, we discard the original affect words and the POS information, and combine the intensities and centralities of the remaining affect categories. Intensities and centralities are handled differently, since they represent different types of information. Centrality indicates the purity of a quality represented by an affect category. Intensity indicates the strength of that quality. Thus, the number of occurrences of a particular affect category in the document does not affect its centrality, but does affect its intensity. Centrality, as the purity of a quality, depends on the maximal centrality over all instances of that affect category in a particular document.
(4)
This representation of the fuzzy set of affect categories enables us easily to spot predominant qualities of affect categories in the document. The meaning of this affect category set is that the document has a high degree of violence, humor, warning, anger, success, slander, greed, horror, aversion, absurdity, excitement, desire, pleasure, promise and surfeit; a medium degree of repulsion, fear, lack, death, slyness, intelligence, deception, insanity, clarity, innocence and inferiority; and a low degree of pain, disloyalty, failure, creation and surprise. To compute the overall intensity we use a simple average over all affect category instances and their respective intensities (5) where overall intensity of a document ; total number of affect category instances in the document ; intensity of an affect category instance . For the example document, overall intensity is 0.6.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
489
Fig. 2. Centralities with positive connotations up are shown separately from those whose connotations are negative. Affect categories are arranged into similar groups around the circle.
Fig. 3. When all centralities from Fig. 2 are expanded using the fuzzy thesaurus, we get a greater level of detail. Note that additional affect categories exist in the new chart. Centrality values are omitted for greater legibility (scale 0–1).
Overall intensity is used to detect documents with offensive content. For example, high overall intensity (0.7) in combination with the specific centrality profile distaste violence pain may indicate offensive and undesirable content.
IV. AFFECT SET VISUALIZATION An interesting and important area related to the fuzzy typing work is visualization of the results. We have developed a simple affect tagging and affect profile browsing application called the
490
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Fig. 4. News report on train crash in London, October 1999. Centralities describe quality, and intensities, quantity (count and strength) of the affects in a document. As might be expected, the affects fear, harm, pain and surprise are most central. The most intense affects are conflict, confusion, disadvantage and pain.
Fig. 5. Profile of a recent cult movie Matrix generated from a movie review. Opposite affect categories are placed on opposite sides of the circle (the left side shows negative affects, the right side, positive). With a well-developed negative side, this movie is deservedly rated “R” in the US.
Affect Inspector. We show a basic browsing setup from that application in Appendix B. Each affect category’s centralities and intensities can be represented as a point on the perimeter of a unit circle. Centralities and intensities can then be visualized, as shown in Figs. 2–8. In order to demonstrate various ways these charts can be organized, we show different charts for different information objects. In Fig. 2 we show centralities with positive affects separated from centralities with mainly negative affects, for the movie review of Exterminating Angel. In this way, the positive versus the negative side of the document can be easily analyzed. Moreover, groups of similar affect categories are shown close to each other. This facilitates a quick overview of aspects denoted by those groups. For example, in Fig. 2, clarity, creation and intelligence are not as well developed as success, desire, pleasure,
humor and excitement. Groups of similar affect categories are generated using the technique discussed in Section II-C. When all affect categories from Fig. 2 are expanded using the fuzzy thesaurus, we obtain the chart in Fig. 3. The chart contains both positive and negative affects, with a higher level of detail, since new affect categories have been added to the chart through expansion. In Fig. 4 we show the affect structure of a news report concerning a train crash in London. We show both centralities and intensities for qualities typical of news on accidents. Opposing affect categories can be placed on opposite sides of the circle with respect to the center point. This is illustrated in Fig. 5, for centralities in a Matrix review. With this arrangement, we can easily spot which part of the circle is best developed and understand its affective impact.
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
491
Fig. 6. Centralities (up) and intensities in T. S. Eliot’s famous “Rhapsody on a Windy Night.” Even the surrealistic content of this poem can be succesfully analyzed. The poem contains mixed affects of weakness, disadvantage, insanity and strength, clarity and openness. For intensities, clarity, weakness, and inferiority seem to prevail.
Affect categories can be generated for sets of movies. We analyzed sets in the romance, action, science fiction, comedy, and family genres. Each set contains about 15 movies and some movies belong to multiple sets (for example, 12 Monkeys belongs both to the action and science fiction genres). The results are shown in Table II, along with profiles for news about accidents. The generated profiles confirm our expectations about affect categories for different movie genres: romance movies showed high levels of happiness, innocence and justice; action movies scored high on surprise, strength and slyness; family movies emphasized sensitivity, morality, responsibility and nurturance; and so on. The composite profiles also reveal some less obvious features: high levels of immorality and inferiority in the romance movie set, and high levels of inferiority and destruction in the comedies, for example. Although the results in many cases represent typical characteristics of movie genres, we believe that some aspects of the results may reflect only the selected movies within the genre. Next, we show that poetry can be analyzed using this approach. Fig. 6 shows centralities and intensities for a famous poem with a very accurate affect profile. Another application of this technique is personalization. We generated affect profiles for different users based on their movie preferences and the results clearly reflect different personal tastes. The illustration in Fig. 7 shows the centrality affect profiles obtained after merging profiles for each person’s favorite movies. In merging centrality affect profiles, we use the same approach as when merging affect profiles of individual words or sentences, that is, the union (maximum) operator. V. COMPUTING WITH AFFECT PROFILES To illustrate the computational potential of the fuzzy semantic typing framework, we present here an example of retrieval based on similarities drawn from affect categories. We illustrate the technique on retrieval of similar phrases, but it extends equally
TABLE II CENTRALITY RESULTS BY TYPE
492
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
TABLE III PHRASE RETRIEVAL
well to retrieval of larger portions of texts, such as sentences or paragraphs. Here is the outline of the experiment from which we produced Table III. 1) We found the affect profile of the sentence fear, anger, grief and pain filled the room. 2) We retrieved all phrases from a precompiled list of 34 phrases, from four different documents that had affect profiles most similar to the seed phrase. As a similarity measure between the affect profile of the seed phrase and the affect profile of a candidate phrase , we use [15] (6) represents the sum of the intersections (minwhere imum values) of the affect sets’ centralities for the respective represents the sum of the unions affect categories and (maximum values) of the affect sets’ centralities for the respective affect categories. We found phrases as presented in Table III, ordered by decreasing scores for similarity to the seed phrase. Each phrase is shown with its respective affect centrality and intensity profiles, similarity degrees, and average intensities.
This retrieval-by-similarity technique can be complemented with the filtering-on-intensity technique. For example, we can set the intensity threshold to a certain category’s intensity, and filter out all expressions whose intensity for that category is greater than the threshold intensity. Alternatively, we can set the threshold on overall intensity and filter out all documents with intensities higher than the given threshold. Such similarity computations can be conveniently combined with statistical methods. For short phrases like those shown here, the similarity measure we used is very convenient since it reflects the overlapping of qualitative features (affect categories) without taking into account statistical features like number of words or categories. In certain cases, it may be reasonable to use a vector space model with cosine similarity to find similar phrases. One must be careful, however, as that approach will not always return the desired qualitative profile. To illustrate this point, let us imagine that we have submitted the query love, pain to a text corpus and that we are searching for sentences that have both qualities—i.e., sentences that match both love and pain. However, vector space retrieval would return sentences containing many instances of love with a high score, even if there is no mention of pain in them. Our similarity computation, on the other hand, gives preference
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
493
Fig. 7. Affect profiles generated from personal movie preferences. Each person submitted from 10 to 81 favorite movies. Individual movie profiles are merged using the maximum operator on centralities to yield the personal profiles in this figure.
to documents having higher-level maximal centralities for both categories, regardless of their total count in the sentence. This example clearly illustrates the main differences in the quantitative vs. the qualitative approach to love and pain. In the quantitative approach, we are concerned with finding information that is statistically similar to our query. In the qualitative approach, we are more concerned with a particular qualitative profile of the target information. VI. FURTHER DEVELOPMENT The fuzzy semantic typing approach deals well with ambiguity and imprecision in free text. It can be effectively combined with a set of visualization tools for easy, accurate analysis of the affect content in a document. These results are promising, but we feel that we have just begun exploration in uncharted territory. Our plans in the immediate future include: 1) Enrichment of the existing affect typing mechanism. To the extent that expressions in free text and their constituent words have syntactically and semantically different roles, we are making an approximation by using fuzzy union to compute centralities. Although individual affect words are treated equally, their roles are frequently uneven. For example, modifiers (very, more or less, not), comparative and superlative adjectives, adverbial phrases, noun phrases, and complex phrases (with and/or connectives, etc.) all represent different classes of expressions with possibly different roles in a phrase or sentence. A finer-grained analysis
would, therefore, lead to different centrality combination operators for each of these classes. The same holds true for the hierarchical centrality computation, but to a lesser extent, since text hierarchies are more regular and do not have as richly varying a structure. Still, for many documents, sentences and paragraphs can be assigned weights that reflect their relative importance in centrality computations. For example, since report-style prose is typically “front-loaded,” a possible approach would be to increase the centrality of categories appearing in the title, summary, or lead paragraph of such texts. 2) Support for management of linguistic resources. We would like to begin experimenting with personal (user-initiated) updating of general-purpose affect lexicons. This would include the modification of centralities and intensities attached to affect categories, the addition or removal of affect words, the definition of complex affect categories in terms of basic affect categories, and the tuning of the fuzzy thesaurus to reflect changes in the affect lexicon. It would be especially interesting to experiment with different affect lexicons on the sending and receiving sides of a communication. In such an experiment, a message composed with a help of personal affect lexicon would be interpreted using personal lexicon . 3) Generalization of the fuzzy typing framework. Fuzzy typing may be adapted to many different application areas by developing appropriate resources: busi-
494
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
Fig. 8. Basic browsing setup in TreeViewer. The tree view at upper right shows a browsable text hierarchy; the text pane at lower right displays analyzed text. Affect centrality and affect intensity sets appear on the left. Circles represent average values.
ness, food/cooking, fashion, architectural styles, cultural events and artistic material, psychology-related material, wine and beer, perfumes. 4) Analysis of different points of view. Using an affect lexicon, we can analyze affects in a text. Using a lexicon with expressions and types that describe intentions (e.g., would like to, will, is considering, is thinking about) would give us an intention fingerprint for a document. Using both lexicons at the same time, to find a text’s affectual and intentional
fingerprints, could reveal interesting juxtapositions in the data. 5) Integration with existing text management techniques. Although not currently included in our system, quasistatistical information can be generated to complement qualitative analysis. This would involve a simple extension of common statistical approaches like statistical indexing, term-weighting, IDF-TF scoring and cosine distance [5], on a full feature space (all terms and phrases), or on
SUBASIC AND HUETTNER: AFFECT ANALYSIS OF TEXT USING FUZZY SEMANTIC TYPING
495
TABLE IV THE COMPLETE LIST OF AFFECT CATEGORIES AND RESPECTIVE OPPOSITE AFFECT CATEGORIES
a significantly reduced feature space consisting of a limited number of affect categories. Fuzzy typing would then be complemented with retrieval techniques applied to affect profiles: e.g., querying and retrieval, clustering and classification. We believe the best way to approach this issue is to start with basic similarity assumptions as given in [11], and further investigate techniques from [12] and [15] for computing similarities between fuzzy objects. Finally, we hope to devise an algorithm that effectively combines statistical and quasistatistical similarity scores with fuzzy similarity scores to compute the overall similarity of two texts.
6) Extension to additional languages. While the lexicon for our prototype system was created manually, we are currently experimenting with a more systematic, semi-automatic collection method based on WordNet [2]. If this technique proves successful, it could be applied to foreign language semantic nets such as EuroWordNet [1], to accommodate a variety of other languages. VII. CONCLUSION We describe a novel approach to text analysis that combines semantic typing techniques from natural language processing
496
IEEE TRANSACTIONS ON FUZZY SYSTEMS, VOL. 9, NO. 4, AUGUST 2001
with fuzzy techniques, under the common framework of fuzzy semantic typing. Fuzzy semantic typing is an innovative way to capture metalinguistic facts about a text while allowing for linguistic ambiguity and vagueness. From our analysis of generated profiles for news reports and movie reviews, we believe that the metalinguistic representation can be usefully applied in retrieval, clustering, and classification. The approach is applicable to an indefinite number of domains and lends itself to customization for a particular user or task. We look forward to continuing our research in these directions. APPENDIX A See Table IV. APPENDIX B AFFECT INSPECTOR APPLICATION In Fig. 8 we show an example screen from our affect inspector application, illustrating the display of a profile’s centralities and intensities at different levels of the text hierarchy. The essential part of the browsing system is the TreeViewer. It is a Java/XML application that contains a text hierarchy in the form of a tree. Each node in the tree is clickable, and clicking on it displays the affect profile associated with that node in two panes on the left—the centrality profile in the upper pane, the intensity profile in the lower. The circles represent the average values for centralities (upper left) and intensities (lower left). The text associated with the node is displayed in the lower right pane. Browsing is possible at any level of the hierarchy from top to bottom: corpus, document, paragraph, sentence, affect category. Affect profiles to the left are click-sensitive maps: by clicking on certain points, one can search on individual affects, or change the form of the current display to see opposite category placements, linear placements, or placements ordered by decreasing scores. Such arrangements facilitate easy browsing, verification of affect profiles, and comparison of affect profiles for different text elements, both vertically (affect-sentence-paragraph-document-corpus) and laterally (two or more documents, paragraphs or sentences). REFERENCES [1] University of Amsterdam, Dep. Computional Linguistics EuroWordNet, . (2000, May). [Online]. Available: http://www.hum.uva.nl/~ewn/ [2] C. Fellbaum, Ed., WordNet: An Electronic Lexical Database: MIT Press, 1998. [3] C. Fillmore and B. T. S. Atkins, “FrameNet and lexicographic relevance,” in Proc. First Int. Conf. on Language Resources and Evaluation , 1998, pp. 417–420. [4] T. Fontanelle, “Semantic tagging: A survey,” in Papers in Computational Lexicography, COMPLEX 99, 1999, pp. 39–56.
[5] D. A. Grossman and O. Frieder, Information Retrieval Algorithms and Heuristics. Norwell, MA: Kluwer, 1998. [6] R. Krovetz and W. B. Croft, “Lexical ambiguity and information retrieval,” ACM Trans. Inform. Syst., vol. 10, no. 2, pp. 115–141, 1992. [7] G. Miller and C. Walter, “Contextual correlates of semantic similarity,” Language and Cognitive Processes, vol. 6, pp. 1–28, 1991. [8] R. W. Picard, Affective Computing. Cambridge, MA: MIT Press, 1997. [9] P. Subasic and A. Huettner, “Affect analysis of text using fuzzy semantic typing,” presented at the Proc. of FUZZ-IEEE 2000, The 9th International Conference on Fuzzy Systems, San Antonio, Taxas, 2000. [10] M. Sugeno, “On organization of imprecision based on word classification,” in 2nd Fuzzy System Symposium, 1986, pp. 148–153. Japanese. [11] A. Tversky, “Features of similarity,” Psychological Rev. 84, pp. 327–352, 1977. [12] L. A. Zadeh, “Similarity relations and fuzzy orderings,” in Inform. Sci. 3: Elsevier Sci., 1977, pp. 177–200. , “Fuzzy logic computing with words,” IEEE Trans. Fuzzy Syst., [13] pp. 103–111, 1996. [14] , “A new AI: Toward computational theory of perceptions,” AAAI Magazine, vol. 22, no. 1, pp. 73–84, Spring 2001. [15] R. Zwick, E. Carlstein, and D. V. Budescu, “Measures of similarity among fuzzy concepts: A comparative analysis,” International Journal of Approximate Reasoning, vol. 1, pp. 221–242, 1987.
=
Pero Subasic (M’97) obtained the Dipl. Eng. and M.S. degrees from the School of Electrical Engineering, Begrade University, Yugoslavia, and the Dr. Eng. from Yamagata University, Japan, in 1989, 1993, and 1996, respectively He has been with the Institute Mihajlo Pupin, Belgrade, Yugoslavia; Belgrade University, Yugoslavia; Tohoku University, Japan; Yamagata University, Japan; and, the Tokyo Insitute of Technology, Japan, as a Faculty Member or Researcher. He is currently with Clairvoyance, Corporation, Pittsburgh, PA, where he has worked in text mining, navigation and visualization, data analysis, preprocessing, including semantic Typing and Resource Management. He is the Principal Inventor of the Fuzzy Semantic Typing framework. He has published over 40 research papers and reports in international journals, monographs and conferences. He is author of a book on fuzzy systems and neural networks. Dr. Subasic is a member of Japan Society for Fuzzy Systems (ACM SIGCHI, SOFT) and Yugoslav Society for Soft Computing and Intelligent Systems (SOCOIS).
Alison Huettner received the Ph.D. in linguistics from the University of Massachusetts, Amherst, MA, in 1989. In 1998, she joined Clairvoyance Corporation, Pittsburgh, PA as a Project Manager working on natural language processing (NLP). She refined the CLARIT NLP resources and experimented with extensions, including expanded lexical equivalences, semantic typing, specialized affect handling, and a prototype question-answering system; currently, she is working with e-commerce applications. Prior to joining Clairvoyance, she was a knowledge engineer with Carnegie Group, Inc., and worked on a fact-extraction system and a commercial machine translation system. She is also one of the inventors of the associated patent for an integrated authoring and translation system.
