Constructing and maintaining knowledge organization tools: a symbolic approach

2.1 Statistical and clustering approaches

Unsurprisingly, answers for the term collection task came from the IR field (Salton and McGill, 1983) where different versions of the IDF index (Inverse document frequency) have been formulated to select index terms from documents. Adapted to the thesaurus construction problem, the assumption is that frequently co‐occurring words within a text window (sentence, paragraph, whole text) point to some semantic cohesiveness. A considerable amount of work in the computational linguistics field aiming to identify “collocations” is based on similar assumptions (Smadja, 1993). Following this tradition, Church and Hanks (1990) used a clustering technique to produce classes of words found in the vicinity of one another. However, the co‐occurrence approach to unit extraction is not error free. Some frequently co‐occurring units may be syntactically invalid. Much depends on the size of the window within which the collocations are sought. Once the “index” terms are extracted, the next stage is how to form “classes” of semantically coherent terms, which could be useful for thesaurus construction. A possible answer is to cluster frequently co‐occurring terms in order to form classes but like the extraction of the units themselves, the statistical approach cannot make explicit the relations between the units in the same class nor can it consider word order. Often, the units extracted were lone words whereas terms, especially in technical domains are multiword units (MWU). Lin (2002) proposed a clustering method for identifying different senses of lone words. This approach is useful for identifying groups of synonymous words, which share a generic/specific relation with the parent concept. Precision is measured by comparing a cluster (a word sense) to a WordNet2 sense of the word. However, as most statistically oriented methods, his method needs very high word frequencies. As evidenced by previous authors (Lancaster, 1998; Schneider and Börlund, 2004), methods based on statistical co‐occurrence of text units are bound to produce semantically non‐motivated classes, which will require a lot of human intervention before they can be of use for thesaurus construction. For this reason, the results from statistical co‐occurrence methods are called “first‐order word associations” (Grefenstette, 1994). On the other hand, classes produced by statistical methods have been applied to information retrieval to enhance recall with a certain degree of success. Indeed, a user interested in a word may also be interested in those frequently appearing with it.

An alternative approach to thesaurus construction and maintenance through bibliometrics has been suggested by Rees‐Potter (1989), followed up by Schneider and Börlund (2004). Bibliometrics is the quantitative study of scientific publications. As such, it relies on several statistical measures (distance and similarity measures) and data plotting techniques (multidimensional scaling). The most widely used bibliometric methods are the co‐citation (Small, 1999; White and Mccain, 1989) and the co‐word (Callon et al., 1983) analyses. The basic hypothesis on which these methods are based is that frequently co‐cited documents or frequently co‐occurring keywords identify closely related information units (references, journals, authors). These methods employ different clustering techniques to map out clusters of co‐cited documents or co‐occurring terms onto a 2D space. The layout of the clusters is often interpreted in terms of core and peripheral research fronts. The co‐cited documents are regarded as “concept symbols” to which the co‐citing documents point. Usually, the results of bibliometric studies are used for science policy making to rank authors, laboratories or journals using tools such as the Citation Index. A lot of debate has risen over the use of co‐citation studies for science evaluation policy. Many researchers have questioned the validity of the co‐citation hypothesis. It is not the object of this paper to air these views but an interested reader can find a detailed account in Schneider and Börlund (2004).

What we are interested in here is the claim that bibliometric methods can be applied to thesaurus construction and maintenance. This hypothesis is based on the assumption that since co‐citation analysis links groups of related works, this can lead to grouping documents of related contents. Accordingly, Rees‐Potter (1989) sought to identify conceptual changes in two domains, sociology and economy. She used co‐citation and citation context analyses to identify candidate thesaurus terms from a corpus. Citation context analysis consists in examining the textual surroundings of a citation in order to extract terms, which could shed some light on the context of citation. However, as this was done manually, it was time consuming and was not followed up by any implementation. Schneider and Börlund (2004) followed up this idea by using different tools to automate the stages involved in co‐citation analysis. After performing a co‐citation analysis on a corpus of texts in the field of periodontics, the authors semi‐automatically performed citation‐context analysis on a sample of citing documents in order to extract candidate noun phrases (NPs) around a “concept symbol” (a cited document). This enabled them to obtain a “concept symbol word profile”, i.e. a list of frequently occurring words or phrases attached to a cited document. The authors then sought for conceptual changes over time by comparing these “concept symbol word profiles” over three time periods. Furthermore, the concept symbol word profiles were mapped onto a 2D space using clustering and visualization techniques (co‐word analysis, multidimensional scaling and network analysis). The co‐word analysis measures the strength of association of each pair of words or terms as the frequency of their co‐occurrence divided by the frequencies of their separate occurrences in a given context (usually a document). This measure is normalized in several ways (Inclusion index, cosine, Jaccard index). Here the documents are replaced by the citation contexts from where the concept symbol profiles were extracted. Multi‐dimensional scaling (MDS) is a data plotting technique, which generates a visual display of similarity matrices. Like most data analysis techniques, MDS has the major problem of plotting data from n dimensional spaces onto a 2D or 3D space. This is usually done at the price of some distortion or information loss. Moreover such techniques cannot scale up to handle very large matrices. An algorithm based on MDS was used by Schneider and Börlund (2004) to plot several maps of co‐cited documents and corresponding co‐words found in their citation contexts. The authors then looked for visual changes in these maps across different time periods. It was not clear from the study whether the candidate terms extracted actually shared any semantic relations as no examples of such terms were provided. Usually, the maps produced by bibliometric methods are topographic representations of research topics or actors in a field and cannot label explicitly the semantic relations between each concept pair. More empirical evidence will be needed in support of the claim that bibliometric methods can be used for thesaurus construction and maintenance.

2.2 Linguistic‐oriented approach

We call the methods found here “linguistic‐oriented” because the main foundation for selecting terms and relations are based on linguistic criteria. However, some of the methods make use of statistical measures to select terms or to form classes of terms. The linguistically oriented approach to thesaurus construction can be broken down into three specific approaches:

1. 1.

   Functional similarity (distributional approach).

2. 2.

   Internal syntactic evidence.

3. 3.

   Corpus‐based semantic relation markers and external semantic resources.

The internal syntactic evidence approach relies on the grammatical function of component words within a term. Several studies have been done on organizing terms by “head‐modifier” relations (Ruge, 1992; Woods, 1997; Grefenstette, 1997). The head is the noun focus and is the last noun in a compound structure (thesaurus construction system) or the noun preceding the preposition in a syntagmatic structure (system for thesaurus construction). The idea of the ”modifier‐noun” function distinction is to form classes of nouns (heads) that share same modifier or alternatively classes of modifiers around the same head noun. The former would portray sets of related concepts that share similar properties (modifiers) while the second would portray sets of related properties around the same concepts (heads). Usually, a similarity coefficient is calculated using the mutual information index (Church and Hanks, 1990) in order to select the sets of heads (resp. modifiers), which will be put into the same class. In this way, Grefenstette (1997) produced classes for the word “research”, depending on the grammatical function of the word: types of research (market research, recent research, scientific research) and research things (research project, research program, research center) (Sanderson and Croft, 1999). In the first case where research is the head noun, the three terms tell us something about “research” but without making explicit the exact semantic relation. As with the distributional approach, internal evidence applied in this way identifies loosely‐related sets of terms which may share several semantic relations: “research project” and “research program” maybe considered as synonyms whereas a “related” relation holds between the former two and “research center” (place where research project and programs are carried out). The above two approaches are useful in information retrieval tasks such as query expansion but require more filtering before they can be used for thesaurus construction. Also, the output of these approaches is not a hierarchical structure but rather horizontal one (synonyms, related terms). They need to be completed by other linguistic cues like “subsumption” (lexical inclusion) in order to identify also hierarchical relations.

Aside from internal evidence gained from the term's internal structure itself, other means of acquiring explicit semantic relations from texts have been investigated. These means can be categorized either as endogeneous (corpus‐based) or exogeneous (use of external semantic resource). The distributional and internal evidence approaches fall under endogeneous approach. A third type of endogeneous approach to semantic relations acquisition relies on a series of surface relational markers identified in each language. The underlying hypothesis is that semantic relations can be expressed via a variety of surface lexical and syntactic patterns. According to Condamines (2002, pp. 144‐5), this idea can be traced back to Lyons (1978) who used the term formulae to refer to terms linked by a hyperonymic relation, and also to Cruse (1986) who spoke of diagnostic frames. Hearst (1992) identified a list of hypernym/hyponym markers, which have come to be widely used for corpus‐based semantic relations acquisition. Examples are the sequence “such NP1 as NP2, NP3 and/or NP4” where NP is a noun phrase. Morin and Jacquemin (2004) adapted these hypernym/hyponym (generic/ specific) markers to the French language and used them to acquire relations from corpora with the aim to assist thesaurus construction. However, for their method to function, it needs a bootstrap of manually defined pairs of semantically related words. Then combining three separate tools: Prométhée (Morin, 1998), Acabit (Daille, 1996) and Fastr (Jacquemin, 2001), the method incrementally learns other relational contexts in the corpus. Prométhée uses a bootstrap of handcrafted lexico‐syntactic patterns to acquire specific semantic relations between terms in a corpus. This tool extracts generic/specific term candidates incrementally. Acabit is a term extractor based on symbolic and statistical features. The symbolic features are used to select likely morpho‐syntactic patterns of terms (N‐N, N‐of‐N, …) while statistical measures are used to sort the selected candidates by order of likelihood. Fastr is a term variant generator. Starting from a lemmatized list of terms, Fastr finds their morpho‐syntactic and semantic variants in a corpus. An inherent limitation of the relational markers as defined by Hearst is that they are intra‐sentential, i.e. the related terms have to occur in the same sentence. This leads to the loss of semantically related terms, which do not occur in the same sentence. As a remedy to this limitation, Morin and Jacquemin (2004) perform an additional expansion of hypernym links by mapping related one‐word terms onto multiword terms. For instance, given a link between “fruit” and “apple”, similar links between the multi‐word terms “fruit juice” and “apple juice” are extracted. This relationship is established on the basis that:

1. 1.

   The two terms share the same head (juice).

2. 2.

   The substituted words have the same grammatical function (modifiers).

3. 3.

   The substituted words are semantically close.

According to the authors, the rate of error among the links generated in this way is rather low because this was done on very specialized domain where many of the multi‐word terms were non polysemous. However, they admitted that this method couldn't acquire all possible links between a single term and its hyponyms because multi‐word terms are less frequent in a corpus. Their methodology is considered as a “context‐based assistance to the extension of a thesaurus” (Morin and Jacquemin, 2004). Their study constitutes, to our knowledge, the most elaborate attempt to use linguistic knowledge instead of statistical one to build or update a thesaurus.

Our methodology shares some common features with Morin and Jacquemin's study. First is the fact that we also extract terms and their variants (morphological, syntactic and semantic), which constitute the “index units”. We also establish several linguistic relations between these terms using internal evidence (the term structure itself) and contextual evidence (relational markers). Complementary external resources can be used where available to increase the number of semantic links acquired. Internal evidence can be viewed also as a way of projecting relations between terms that are not necessarily in the same sentence.

Where our method differs from the work of these authors is in the way in which the terms are organized: we build classes of terms based on a clustering algorithm which not only groups synonymous terms, generic/specific terms but also associated terms. This is done iteratively such that specific semantic relations can be integrated at different iterations of the clustering algorithm. Thus these classes represent, for a given term, its closest neighbours in terms of semantic relations. A thesaurus builder can then decide how to place these terms and their links in the target hierarchy.

3.1 Corpus constitution

Illustration of the methodology will be done on a collection of scientific titles and abstracts of papers published in 16 information retrieval journals between 1997‐2003, called IRCorpus henceforth. IRCorpus comprises 455,000 words which were downloaded from the PASCAL bibliographic database maintained by the French institute for scientific and technical information (INIST/CNRS). The corpus was collected on the basis that they were summaries of research published by leading IR journals over eight years and as such should reflect important terminology in the IR field. The list of source journals is given in the Appendix.

3.2 Linguistic component

This component relies on shallow NLP to extract terms and relates them through various relational devices. Terms are linguistic units (words or phrases) which taken out of their contexts, refer to existing concepts or objects in a given field. In other words, terms are choice linguistic units, rich in information content because experts in a field to name the objects or concepts of that particular field use them. A lot of research has been done on automatic term extraction by computational terminologists (see Jacquemin and Bourigault, 2003, for a review). However, focus has now shifted towards automatic terminology structuring for various applications (ontology learning and population mostly) as shown by the recent issue of the journal Terminology (Ibekwe‐SanJuan et al., 2005).

3.2.1 Term extraction. It is an accepted fact that most terms appear as noun phrases (NPs) although some verbs and prepositional phrases can be terms. We currently extract only terminological NPs, which are multiword units (information retrieval system; peer review of technology‐experts). Term extraction is performed using the LTPOS tagger and LTChunker developed by the University of Edinburgh. LTPOS is a probabilistic part‐of‐speech tagger based on hidden Markov models. It has been trained on a large corpus and achieves an acceptable performance. Since LTChunker only identifies simplex NPs without prepositional attachments, we wrote contextual rules to identify complex terminological NPs. The term extraction process is exhaustive in that frequency is not used as a filter. Only badly formed candidates from the morpho‐syntactic basis are dropped from further processing. This ensures that we will be working on all the candidate terms from the corpus and not on a subset, as is usually the case with approaches based on statistical measures. Selection will be done through linguistic means. Terms, which do not share any semantic relations with others, cannot be further processed. Likewise, terms, which are too loosely bound to the terminological network, may disappear from the final clusters obtained. Term extraction on the IRcorpus yielded 44,665‐candidate terms.

3.2.2 Acquiring semantic relations between terms. The issue of establishing relations between text units is not a recent one. In the computational terminology community, a paradigm has emerged which is known as the “variation paradigm”. As more and more corpus‐based evidence become available, terms are no longer considered as “fixed” units but as dynamic units, which vary under several linguistic influences. Variation denotes the fact that a term may appear under different forms and that locating these different forms is essential for many information‐oriented tasks like query expansion, question‐answering, information retrieval, indexing, terminology acquisition and structuring, ontology population. Variations are not rare phenomena; they affect about 35 percent of terms in a domain. Variations can take place at different linguistic levels: morphological, syntactic or semantic, thus making their identification impossible without integrating NLP techniques. Capturing these variations enhances the representation of the actual state of a domain's terminology. Also identifying and structuring terms along variation lines enhances the understanding of the conceptual relations between a domain's concepts and hence represents an initial step towards knowledge organization.

Term structuring through variations has been explored for various applications like building domain lexical resources from corpora (Daille, 2003), automatic thesaurus building (Morin and Jacquemin, 2004), information retrieval, question answering (Dowdall et al., 2003), science and technology watch (Ibekwe‐SanJuan and SanJuan, 2004), the list not being closed.

Here we combine several linguistic approaches to semantic relations acquisition: internal evidence (morpho‐syntactic variations), contextual evidence (lexico‐syntactic patterns), and xternal resource (domain thesaurus, ontology or general purpose semantic resource). The first two stem from endogeneous corpus‐based approach while the third is an exogeneous approach to semantic relations acquisition. We thus propose a comprehensive approach in order to complement the shortcomings of each approach and increase the number of relations mined between corpus terms. This constitutes an enhancement to our previous studies where only the first approach was used (internal evidence).

1. 1.

   Internal evidence or morpho‐syntactic variations. This relies on the structure of the terms themselves to establish relations between them. The relations identified here rely on two morpho‐syntactic operations: expansions and substitutions. Expansions denote the fact that term1 is subsumed in term 2 either by the addition of modifier or headwords. For instance, between “academic library” and “hellenic academic library”, the addition of modifier words (regardless of their number) tends to create a generic/specific relation (the longer variant is more specific) while in “British Library” and “British Library restoration”, the addition of a new head word leads to a horizontal relation thus creating loosely “related terms”. If a chain of modifier expansions is found in the corpus, this can lead to building a hierarchy from the most generic term to the most specific.

   Substitution operates on terms of equal length in which one and only one element is substituted, either in a modifier or in a head position. This idea was exploited in Grefenstette (1997). Substitutions engender horizontal relations between terms. Therefore, the resulting conceptual relation is a more general “relatedness”. Modifier substitutions can denote members of the same concept family with alternative qualifications, siblings in an IS_A hierarchy. The conceptual shift engendered by head substitutions, on the other hand, links different IS_A hierarchies at the same level of specificity.

   The semantic links acquired through expansions and substitutions are effected through pure lexical association, i.e. the fact that two terms share some common elements. This enables us to capture quite a considerable number of links between corpus terms. Figure 1 shows the type of hierarchies that can be obtained through internal evidence.

   Simple lines denote modifier substitutions while simple arrows denote specific links generated by modifier expansions. Bold arrows denote head expansion links while bold lines denote head substitutions. The terms preceded by a star may not be direct hyponyms of the generic term since they contain many modifier words. There may be intermediary terms between them and the topmost generic term but which were absent in the corpus.

   Lexical association enables us to obtain a polyhierarchical structure where a term may belong to two hierarchies, as is the case for “inductive learning method” and “prominent machine learning method”. While the association link between “inductive method” and “learning method” seems understandable (both can be attached to a hierarchy of “methods” thus making them co‐hyponyms), the link between “machine learning” and “machine type” (head substitution) is less obvious and less interesting. Given their definition, substitutions are the most prolific and create sets of loosely related terms whose significance is not always clear for the domain specialist. For this reason, they will be subjected to further filtering before clustering in order to retain only semantically motivated substitution variants (section 2.2.3). 26,128 terms were involved in relations through internal evidence, thus more than 58 percent of the total number of extracted terms (44,665).

2. 2.

   Contextual evidence or lexico‐syntactic relational markers. While internal evidence undeniably enables the acquisition of a certain number of semantic links such as generic/specific, the approach is inherently limited in that it cannot capture conceptually related terms, which do not share any lexical element. For instance, AltaVista and search engine will not be related, nor will car and vehicle be, whereas both pairs obviously share a generic/specific relation. In other words, internal evidence cannot detect semantic variants when the relation is materialized by other linguistic devices aside from lexical association (common words between the two terms). To ensure a maximum capture of semantic relations between domain terms, we need a complementary approach, which relies on contextual relational markers. These markers have been studied by Hearst (1998) among others. The generic lexico‐syntactic patterns amongst which generic/specific relations can occur are represented by the following regular expressions:

   H1: such NP₀ as NP₁ (<cc>)+ NP₂ …, (<cc>)* NP_n

   H2: NP₀ such as NP₁ (<cc>)+ NP₂ …, (<cc>)* NP_n

   H3: NP₀ (<cc> | NP₁) like (NP₂ | <cc>)+ NP_n

   H4: NP₀ (<cc> | NP₁) (particularly | especially | in particular) (NP₂ | <cc>)* NP_n

   H5: NP₀ (<cc> | NP₁) including (<cc> | (NP₂ | <cc>*)) NP_n

   H6: NP₀ ([;|:]) NP₁ [parenthesis] (,) NP₂* (<cc>) NP_n

   H7: NP₀ (<cc> | NP₁) [,] (<&>) other NP₂ …, (<cc>)* NP_n

   H8: NP₁ (, | namely) NP₂ …, (<cc>)* NP_n

   Where NP can either be a simplex or a complex noun phrase with a prepositional (PP) attachment; “ < cc>” is a coordinating conjunction element such as “and, or, comma”; “*” is Kleene's star; “+” means that there must be at least one occurrence of the preceding element. Brackets encase optional syntactic structures (NP) or lexical elements (cc). The words in brackets are optional. The lexical elements in angle brackets are not terminal categories. When such relational markers are found in the text, the surrounding NPs are marked and extracted with the labeled semantic relation. For instance, pattern H2 would apply to the sentence below.

   (1) challenging requirements in [HYPER] non‐traditional applications such as [HYPO] geographic information systems ([HYPO] GISs), [HYPO] computer‐aided design ([HYPO] CAD), and [HYPO] multimedia databases.

   This yields five pairs of terms in a generic/specific relation. The specific terms are in turn related as co‐hyponyms. Graphically, we obtain Figure 2 where arrows indicate the hierarchy engendered by the generic/specific relations. Symmetrical relations denoting co‐hyponyms are shown by simple straight lines. Furthermore, since “GIS” is an acronym of “geographic information system” and “CAD” an acronym of “computer‐aided design”, the two pairs are semantic equivalents. Bold lines represent this.

   Certain lexico‐syntactic patterns signal the presence of synonyms relations between terms. Authors writing up a text make use of different naming mechanisms in order to introduce alternative terms for the same concept. Amongst the most cited patterns found in the literature (Suárez and Cabré, 2002), we tested the following:

   S1: NP₁ (also | equally | now) called (now, also equally) NP₂

   S2: NP₁ (also | equally | now) known as NP₂

   S3: NP₁ (also | equally | now) named NP₂

   S4: NP₁ (also | equally | now) viewed as NP₂

   S5: NP₁ (also | equally | now) termed as NP₂

   S6: NP₁ termed (for | as) NP₂

   S7: NP₁ referred to as NP₂ (<cc> | NP₃)

   S8: NP₁ ( < have> (also | equally | now) |<be>) defined as NP₂

   S9: NP₁ <be> no other than NP₂

   The infinitive verbs appear in inflected mode (tense, number). We observed that these patterns are subject to variations, for instance, the presence of an optional punctuation mark (the comma) appearing before the cue words (called, named, …). Also the cue words can appear in parenthesis with one of the synonym terms. It is interesting to note that while the adjective form “named” functions as a synonymy marker, its adverbial form (namely) is used in a generic/specific relational pattern (H8). Given the following sentence, pattern S2 will mark and extract the terms “mathematical operation” and “convolution” as synonyms.

   (2) This combination is performed by a [SYN] mathematical operation known as [SYN] convolution.

   We manually validated the contexts extracted by the two types of relation markers and measured their precision and recall. Generic/specific markers enabled us to extract 571 contexts relating, 162 pairs of terms. A context with all candidate pairs was accepted if the relation was judged semantically sound, if not both the context and the candidate terms were rejected. 77 percent out of the 571 contexts contained relevant relations while 23 percent were discarded.

   Synonymy markers enabled us to extract 107 contexts containing 146 pairs of synonym terms. As synonymy is often a binary construction (term A is also called term B), it hardly involves enumeration devices unlike the generic relation markers; hence the number of synonym pairs is close to the number of contexts. Out of the 107 contexts extracted by synonym markers, 92 percent represented valid synonymy relations while 8 percent were discarded.

   Recall consisted in checking if any of the relation contexts was missed by the lexico‐syntactic patterns. Given the size of the corpus (455,000 words), we could only carry out this task on a small portion. The first 200 out of the 3,355 abstracts were manually checked for other generic/specific or synonym relations, which were not extracted by our lexico‐syntactic patterns. The result was compared against the list of generic/specific terms acquired by the patterns on this portion of the corpus.

   The rate of recall was found to be high (86 percent) for generic/specific relations whereas it was slightly lower for synonym patterns (77 percent).

3. 3.

   External semantic resource. This third approach to semantic relations acquisition completes the other two (internal evidence and contextual relation markers) by enabling the acquisition of relations, which were not signaled by endogeneous markers. This can be done via a domain thesaurus, ontology or a general language resource such as WordNet (Fellbaum, 1998). WordNet is a general language semantic database with three categorial hierarchies: noun, verb and adjective. WordNet contains generic/specific, synonym, meronym and “relatedness” as well as other semantic relations. Words are organized in “synsets” in WordNet, a synset being a class of words representing the same sense. Typically, a polysemous word may have up to six synsets. Using an external resource is also a means of filtering the substitution relations acquired by internal evidence. In the IRcorpus, substitutions were filtered in order to obtain two subtypes: strong and weak substitutions. Strong substitutions are those whose substituted words belonged to the same synset. Weak substitutions are all the other lexical substitutions involving terms of length ≥3. The idea was to filter out binary substitutions, as these were very prolific and create a lot of noise in later processes. Some examples of strong weak substitutions are given in Table I.

   We were able to filter out the majority of the substitutions acquired through internal evidence while highly improving the precision. The starred examples illustrate the gap between what can be accepted as general language synonyms (way/language) and what can be considered as synonyms in a specialized domain when using a general vocabulary external resource such as WordNet. Here, multiple language and multiple way are surely not IR terms and not synonymous in this field.

Currently, two means of acquiring semantic relations have been implemented in the TermWatch system: internal evidence and external evidence using WordNet. Acquiring semantic relations through relational markers (contextual evidence) is yet to be implemented.

3.3 Clustering component

Clustering is a data analysis technique used to generate automatically groups of similar objects. Objects in a cluster should have a greater degree of proximity than those in different clusters. Similarity is usually defined as a co‐occurrence function in a text window (sentence, paragraph, whole document). There exists in the literature, different ways of classifying data analysis techniques. A distinct family of methods is factor analysis comprising principal component analysis (PCA) and multi‐dimensional scaling (MDS). However, as we have pointed out earlier, factor analysis involves some data loss and distortion during the plotting. Also, it does not scale well to very large and sparse matrices. Another family of data analysis methods is clustering. There are several clustering algorithms, which divide into two families: hierarchical or partitioning algorithms. Hierarchical clustering algorithms are either agglomerative (they start from atomic units and merge them iteratively into nested clusters) or are divisive (they start by putting all objects into one cluster and then progressively divide it into several clusters). Partitioning algorithms (k‐means) on the other hand have to find a partition where each object is affected to the cluster with the closest “centroid” or center. Partitioning algorithms and factor analysis both rely on distance measures. As such, the analyzed units have to lend themselves to the definition of a distance like χ² or the Euclidian distance. Hierarchical algorithms on the other hand use different mechanisms to interpret the similarity measures between objects in a matrix: single link, complete link or average link. In single link clustering (SLC), an object need only have one link for it is integrated into a cluster. This leads to the well known and undesirable “chain effect”, i.e. very long clusters whose internal links may be weak in some places. By opposition, complete link clustering imposes several links between an object and members of a cluster. This leads to more cohesive clusters but also to forming many singletons (clusters of lone items).

Graph theoretic algorithms are better suited to the nature of our sparse data. The graph theoretical algorithm that has the strongest mathematical properties is clearly SLC because it entails forming connected components from the graph of similarities. However, SLC has the major drawback of the chain effect, thus making it only adapted for special datasets where the desired clusters have this property. Our clustering technique follows the SLC principle but avoids the chain effect by interpreting similarity values in a relative manner, i.e. two vertices are merged if the similarity values are higher than the ones surrounding the vertices. Thus at a given iteration, several sets of vertices can be merged at different similarity values. Similarity values are not considered as an ordered set. In other words, we consider the local maximums of the similarity function whereas the classical SLC considers similarity as an ordered set and at a given iteration, will cluster units only above a given threshold. Thus at a given iteration, classical SLC will merge only sets of items with the same similarity value. Some implementations of SLC limit the chain effect by imposing a maximum size on clusters (Callon et al., 1983). However this comes at a cost: the output of the clustering is no longer unique. A change of cluster size or of the order in which units are clustered affects the results.

Because we cannot define a distance on our datasets (terms) and the semantic relations they share (there is no way to define triangular inequality), because frequency information is not used in our clustering scheme, we required an algorithm which performs well on very sparse data and can differentiate between several symbolic relations. This excluded algorithms based on distance measures (K‐means, factor analysis). Similarity in our method is not measured on a co‐occurrence matrix. Rather clustering is performed on a graph of semantically related terms, i.e. the semantic relations identified in the preceding stage. This is a radically different approach to current data analysis techniques. We describe below the functioning of the method.

3.3.1 Choosing the relations for clustering. The semantic relations identified by the linguistic component can be seen as dimensions from which interaction between domain concepts can be viewed. Depending on the application targeted, the user can choose the relations to be used for clustering and assign a role to them during the clustering process (see section 3.3.2). For instance, for domain terminology structuring where the focus is on obtaining semantically tight clusters, it will be more relevant to give priority to filtered substitutions and modifier expansions within the variants acquired through internal evidence, together with the semantic relations acquired through contextual and external evidence. In other words, applications like thesaurus and ontology building need a careful selection of only those variants, which are semantically close. Contrarily, when targeting applications like science and technology watch or query expansion, the focus is on “recall” or associations rather than on semantic precision. Hence, the tendency will be to include less semantically motivated relations and form loose semantic clusters whose contents are suggestive. For science and technology watch, this could trigger off more explorations to understand the nature of the link by returning to source documents (Ibekwe‐SanJuan and SanJuan, 2004). For query expansion, expanding cluster contents to related terms could be a way of suggesting other query terms to the user. The choice and role of each relation is done in an interactive interface. In between iterations, the user can change the role of any relation to better suit his/her purpose. We used the relations acquired by internal evidence and assigned them a role according to the semantic proximity induced between terms. Modifier expansions are likely to indicate generic/specific links; WordNet‐filtered substitutions are likely to denote synonyms while head expansions point to “related terms”. We thus selected strong substitutions (strong‐Sub) and left expansion (left‐Exp) as priority relations for gathering terms into connected components. We call these priority relations “COMP” relations. The remaining ones: weak substitutions (weak‐Sub), right expansion (right‐Exp) and insertions (ins) are second level relations called “CLAS relations”.

3.3.2 Clustering semantic variants. The variation relations used are represented as a graph. We recall briefly the functioning of the algorithm. CPCL (Classification by Preferential Clustered Link) is a hierarchical two‐step extractor of clusters from a graph of term variants (Ibekwe‐SanJuan, 1998).

One notable attribute of this algorithm is that the clustering begins not at the atomic level (term level) but at the component level. Components are obtained by grouping terms sharing COMP relations. A dissimilarity function d is calculated for every pair of components as the proportion of CLAS relations between them. More formally, d is an application in [0,1] defined for every pair (i, j) of components as follows:

1. 1.

   d(i,j) =1 if for every r in {1, … ,k}, N_r(i,j) =0 and d(i,j) =0 if i=j;

2. 2.

   d(i,j)=where R₁ … R_k are CLAS relations and N_r(i,j) is the number of links in R_r between i and j.

The results obtained on the IRcorpus are taken at the first iteration; no threshold for minimal dissimilarity index was set. The system built 674 clusters with 1,595 components and a total of 5,632 terms. 15 clusters had more than 30 terms and 517 clusters had less than ten terms. The following section analyzes the structure obtained.

4.1 Exploiting cluster content for thesaurus construction or maintenance

We now illustrate how the semantic variants in clusters can assist domain knowledge organization tasks like thesaurus building and maintenance. An example of the organization that can be obtained from a cluster's content is given for cluster 141, labeled “system transaction logs”. This cluster contains 13 terms. Table II shows all the terms in this cluster, structured according to the semantic relation induced by each linguistic operation studied.

Currently, all the relations are induced by internal evidence (lexical association) and WordNet‐filtered substitutions. They only require a shallow NLP processing. The proposed structure in Table II is not built automatically yet, but formatting cluster contents to suit thesaurus presentation can easily do this. It suffices to rebuild the hierarchy of relations basing on the formal properties of the relations used for clustering. Modifier relations involving terms of different length imply anti‐symmetrical links. This can be used to obtain the hierarchy of broader and narrower terms. Head relations involving terms of different length as well as substitution relations imply association (horizontal links). WordNet substitutions imply synonymy, thus “used for/use” relations. However, the visual display of a thesaural structure is not a difficult issue, as this will depend very much on the type of tool used by the thesaurus builder.