Systems vs. Methods: an Analysis of the A ordances of Formal Concept Analysis for Information Retrieval?

Abstract

We review previous work using Formal Concept Analysis (FCA) to build Information Retrieval (IR) applications seeking a wider adoption of the FCA paradigm in IR. We conclude that although a number of systems have been built with such paradigm (FCA in IR), the most e ective contribution would be to help establish IR on rmer grounds (FCA for IR). Since such an approach is only incipient, we contribute to the general discussion by discussing a ordances and challenges of FCA for IR.

Introduction

To the best of our knowledge only two information retrieval-incepted books have realised the potential of FCA for IR: [ 45 ] and [ 11 ]. On the one hand, Van Rijsbergen brie y notes down that the Boolean Retrieval Model is captured in terms of Galois connections between documents and features (terms) [45, p. 37], although he includes there the inverse index on terms and documents which may best be conceived in terms of a Galois adjunction[ 42 ]. On the other hand, Dominich makes a very cursory review of the state-of-the-art up until 2008 [ 11 ]. He notes down the work of [ 37 ] on faceted information retrieval and that of [ 6 ] on browsing Web retrieval results with concept lattices, and the disjunctive approximation to boolean retrieval of [ 32 ]. Curiously, the data-driven nature of FCA is downplayed in this work.

In the FCA camp, the broadest review is still [ 6 ] but [ 5, 23, 37, 38 ] have narrower foci. Notice that both [ 11, 37 ] review work in lattice-based IR systems prior to the groundbreaking [ 21 ], but pre-FCA emphasis is in designing the lattice instead of obtaining it from the relevance relation: the data driven quality of FCA is missing in this early work, e.g. [ 33 ].

We believe that part of the explanation for this divide may be that only the most simple, basic tasks in IR|and using the oldest IR models|have been successfully tackled with FCA techniques. After all, IR in some 60+ years has developed its own set of techniques, methods for research and testing and is practised by, probably, the most thriving community in ICT. It is only natural that FCA can only be considered as a subsidiary discipline to such endeavour. Or not?

In this paper we want to put forward the distinction between FCA in IR and FCA for IR, that is implementing IR systems with FCA vs. augmenting IR with the methods and ideas of FCA. We claim that most of the work so far has been in FCA in IR and the time is ripe to expound on a FCA for IR, that is a theory of the a ordances and challenges of using FCA to solve IR tasks, already started in [ 6 ]. Here we use a ordances in the sense of [ 34 ], to refer to \the actionable properties between the world and an actor", that is, the 'world' of FCA and the 'actor' that is an IR practitioner.

This paper is about raising awareness of these two conceptions of the role of FCA vis-a-vis IR. For this purpose, we introduce in Sec. 2 a prototypical information retrieval task to make explicit what types of problems an IR practitioner comes up with. In Sec. 3 we review to what extent FCA actually solves such problems by supplying a set of a ordances of FCA for IR. Finally, we discuss in Sec. 4 what are further challenges that FCA has to solve for a wider adoption in a number of data-intensive application domains, including IR. 2

A prototypical information retrieval task

To guide our exposition we will discuss the ad-hoc retrieval task, that is, the task where the IR system is expected to produce the documents relevant to an arbitrary user need as expressed in a one-o , user-initiated query [29, p. 3]. Although Web retrieval is perhaps the prevalent IR task at present, ad-hoc retrieval is the best studied one and it admits many di erent models. In the following, we expand the modelling of this task propounded in [ 42 ] as a script to discuss a ordances and challenges in using FCA for IR tasks. A model for batch ad-hoc tasks To x notation, we adapt the formal model put forward by Fuhr [ 17 ] reproduced in Fig. 2|although we interpret the signs there di erently|and we let Q, D, and R respectively stand for a set of information needs for a querying user, a set of information-bearing percepts and a psychological capability whereby a particular user is going to judge the relevance of the information percepts for her information needs.

Q R D

Q D ! Q

R ! D

Q D ! !

Q0 R0 D0 the relevance judgment representations adopt the form of a relevance relation, R D Q . Finally, let Q0, D0, R0 be the query representations, document representations and the relevance judgments in representation space respectively, so that R0 D0 Q0 .

Whereas Fuhr's model considers queries, documents and judgements to be inside the information retrieval system, we consider them both inside and outside, since they are more properly conceived as (multimedia) recordings of the psychological entities and processes considered above. They have an immanent existence independent of the system yet are related to them by their representations. However, representations arise when we try to approximate the information content of queries and documents inside an IR system, hence they are sometimes called surrogates or surrogate representations (for their records).

Although the model posits four maps between the above-introduced domains, for practical reasons it is common to concentrate on only two A query representation process, Q : Q ! Q0, mapping from queries to query representation suitable for processing in a particular information retrieval system.

A document representation process, D : D ! D0, mapping from documents to document representations.

Therefore, we limit ourselves to the domains, mappings and sets enclosed by the square in Fig. 2, the recording- and representation-related domains. Assessment The ideal IR system SD;Q(R) =< %R > would consist in a relevance function %R describing relevant documents where %R(qi) is the set of documents relevant to query qi as dictated by the ideal relevance relation R . But in the process of building an IR system we may incur modelling errors, approximation, etc., whence we accept that the actual relation implemented will be the approximated relevance R^ 6= R for the implemented IR system SD;Q(R^) =< % ^ > . Its retrieval function may only return %R^ (qi) the set of

R documents retrieved for the same query as dictated by the approximate relevance R^, %R : Q ! 2D %R^ : Q ! 2D (1) qi 7! %R(qi) = fdj 2 Djdj Rqig

qi 7! %R^ (qi) = fdj 2 Djdj R^qig :

The batch retrieval task can be subjected to the so-called \Cran eld model of Information Retrieval system evaluation" [ 31 ], where a set of document records, or collection, DT D, a set of sampled query records, topics, QT Q, and a set of relevance judgments involving documents and query records, RT DT QT are known. Assessing the quality of SD;Q(R^) means, essentially, comparing R and R^ : For a given query q, the system would retrieve documents %R^ (q) whereas the relevant documents are given by the prescribed relevance as %R(q) . Therefore the retrieved relevant documents for each query q 2 Q would be %R(q) \ %R^ (q), and we would have precision PR^ and recall RR^ |or any measure derived therefrom|as (2) A decomposition of the problem. We believe it is convenient to conceptually decompose the synthesis of SD;Q(R^) into the following problems[cfr. 6, x. 4]: Problem 1 (Representation). Given di erent spaces of queries Q and their representations Q0 nd a mapping Q between them. Do likewise for documents D, their representations D0 and a surjective mapping D between them. Problem 2 (Generalization). Given local information about the relevance relation R in the form of a training subset RT0 = DT0 Q0T , extend/generalise such information to R^0 D0 Q0.

Problem 3 (Surrogate implementation). Given domains of documents D and queries Q (whether they be descriptions or representations), a querying hypothesis and an estimated relevance relation R^, build an information retrieval system that faithfully implements the prescribed relevance4.

Once solved these problems we can build the retrieval set as %R^ (q) =

D1[%R^0 ( Q[q])] (3) where we have taken the precaution of making all of the functions apply over sets rather than singletons.

Problem 4 (Post-retrieval interaction). Given the answer set to a query %R^ (q) present it to the user in an e ective manner.

Note that in standard IR engineering practice the steps of retrieving document representations and then nding their original document are often aggregated by means of an inverted index. Also, (3) is often complemented with retrieval status value for each result, a number stating the degree of relevance of each retrieved document to the query. 3

A ordances of FCA for IR

This list is going to be informally structured as a sort of proof: rst we state what we consider the a ordances of FCA for IR and then we explain the reasoning behind our assertion.

A ordance 1 (Solving problem 3 in the conjunctive Boolean Model). FCA implements the (conjunctive) Boolean Keyword model. 4 We use here R^ as a variable ranging over possible relation values, not necessarily the optimal one.

Suppose that there exists a set of keywords T 5, queries are represented as keywords Q0 T , documents are represented as set of keywords D0 2T , and estimated relevance R^0 is de ned by means of the inclusion relation d0R^0q0 , d0 3 q0 . The retrieval function is easy to write %R^0 (ftg) = fd 2 D j q 2 dg, but what are we to expect when supplying several queries, that is, several keywords?

To implement conjunctive querying we produce the intersection of the result sets, that is, for B = fqigi2I we have

%1R^0 (fqigi2I ) = fd 2 D j 8i 2 I; qi 2 dg = \i2I fd 2 D j qi 2 dg : In that case, the more keywords a query has the less documents the retrieval function returns, that is, q10 q20 implies %(q10) %(q20). Then we realise that this retrieval function is the query polar %1R0 (q) of the Galois Connection in Fig. 3.(a) %1R^0 : 2Q ! 2D 1R^0 : 2D0 ! 2Q0 %1R^0 (B) = fd0i 2 D0j8q0 2 B; d0iR^0q0g %2R(B) = fd0i 2 D0j9q0 2 B; d0iR^q0g %2R : 2Q ! 2D 2R^0 : 2D0 ! 2Q0 1R^0 (A) = q0 2 Q0 j 8d0 2 A; d0R0q0

2R^0 (A) = q0 2 Q0 j d0R0q0 ) d0 2 A (a) Galois connection (b) Galois adjunction

This is one of the contributions of [ 21 ], the rst paper to use FCA in IR, that is, to build a Galois connections that implements an IR system, to the best of our knowledge. Most of the work in FCA in IR uses this model [ 3, 9, 13, 37 ], with the notable exception of the work starting with [ 32 ], who de ne relevance in a way that leads to the disjunctive model of Fig. 3.(b). In this case, %2R^0 (B) = fdi 2 Dj9q 2 B; diRqg ; but, since there are some tricks to representing this in a concept lattice [ 44 ], the authors of [ 32 ] develop a browsing model of their own. A ordance 2. FCA implements query term expansion

In fact, the Galois connection has \another half", the document polar. Let A D0 be a set of documents. Then the set of queries for which all those documents are relevant is 1R^0 (A) = fq0 2 Q0 j 8d0 2 A; d0R0q0g : Actually retrieval sets come in pairs called formal concepts 6. In our example, a formal 5 This is sometimes called the bag-of-keywords model of documents. 6 In [ 21 ] they were originally called \complete pairs". concept (A; B) is a pair of a set of documents A and a set of queries B so that all the documents in A are relevant to all the queries in B, and dually,

A = %1R^0 (B); B = 1R^0 (A) . These pairs come from the properties of the polars in the Galois connection, as described in Fig. 4: the composition of the polars are extensive, idempotent operators, that is, closure operators.

2D r

T

Note that for a set of queries B 2 Q; R1^0 (B) B hence querying through formal concepts expands the query sets in a data-dependent manner. This was noted cursorily in [ 21 ] but is thoroughly explained in [6, Chap. 3] whose authors have contributed the most to this line of work.

A ordance 3. FCA provides for integrated browsing and querying.

As previously noted, query submission in a concept lattice-based IR system is just an application of the query polar, which obtains the concept whose extent is the retrieval set, and whose intent is the extended query. This acts as a querying mechanism.

On the other hand, formal concepts have a natural order based in the inclusion order of extents or the dual inclusion order of intents, (A1; B1) (A2; B2) , A1 A2 , B1 B2 . Furthermore, the Fundamental Theorem of Concept Lattices asserts that this order between concepts is a complete lattice [20, p. 20], representable as an order diagram.

Godin et al. [ 21, 22 ] put forth the idea that lower and upper neighbours as well as parallel concepts de ne a topology for browsing in a (concept) lattice (see Fig. 5 ). Consider a concept in focus C, { Below it lie its lower covers, those concepts with more stringent (higher cardinality) query sets. { Above it lie its upper covers, those which have less stringent (lower cardinality) query sets. { To each side of the concept in focus stand those sibling concepts sharing parents (and descendants) with it. They have incompatible query sets (inconsistent with the focus concept intent).

Although previous work had noted the interest of lattices for navigation, to the extent of our knowledge, Godin et al. were the rst to tie the modi cation of queries (and therefore retrieval sets) to navigation in a systematic manner. For in-detail reviews of this a ordance in the context of Personal Information Systems see [ 12, 13 ].

A ordance 4. FCA provides visualization schemes for the document-query lattice at di erent scales.

The scales we refer to in this a ordance are those related to the visual and informational complexity of the lattice. Complexity scales are in other contexts termed the micro-, meso-, macro- and mega-scales.

The local neighbourhood of a formal concept illustrated in Fig. 5 was posited in [ 21, 22 ] and developed in a number of works [ 12 ]. It is a micro-visualization device depicting the part of the lattice surrounding a particular concept in focus whether incorporating a sheye view [ 5, 21 ] or not [ 12 ].

On the other hand, the order diagram of the concept lattice acts as a mesoscale visualization technique. Similarly, visualizing only the concepts that lie below|or above|a focus concept produces visualization devices of comparable complexity and can be considered meso-scale visualizations. Here we consider the mapping of the downset of the focus as a tree as in [ 4 ]. Furthermore, the use of attribute and object projections on the whole lattice, reduced labelling and nested line diagrams [ 20, 39 ] are all tools that help us balance displayed information vs. visual complexity allowing us to display complex lattices at the mesoscale.

For those cases when these complexity-reducing strategies are not su cient, very little work has been done on observing lattices at the macro-scale|let alone the mega-scale|sacri cing concept and local structure readability for the quick glimpse of emerging features like height, width, overall shape, concept density, etc. For an illustration of such problems, see the lattices in [ 24 ]. Recently, [ 14 ] have proposed a technique to embed any concept lattice onto a boolean lattice of similar complexity which acts as a representation space disposing of a lot of information: its usability is, as yet, unassessed.

A ordance 5 (Solving problem 4). FCA provides retrieval-set navigation.

This niche application of FCA is perhaps the best-know to the IR community [25, x 10.7]. It is a natural consequence of treating the retrieval set as a subcorpus (of snippets, possibly) and using FCA to establish ordering relations between them as induced by their terms. Perhaps the rst to propose this use of concept lattices is [ 3 ], and it is thoroughly explained in [ 6 ], usability studies included. Systems implementing also a post-retrieval visualization of Web retrieval searches or Meta-searches can be found in [8{10, 26].

A ordance 6. FCA captures naturally occurring (immanent) term dependencies

If terms were independent, then concept lattices, at least from the perspective of terms, should be boolean: all possible combinations of terms would arise as intents, but this is never the case. Since the inception of the rst FCA in IR systems it was noticed that particular groupings of terms occur naturally in documents and this is re ected in the system of intents. Of course, this dovetails into the Automatic Expansion of Queries mentioned in A ordance 2: modelling term dependencies is how automatic query expansion is catered to. In terms of IR models, this means that the model implemented by FCA is actually in the empty square in Fig. 1. Carpineto and Romano [ 7 ] have investigated this issue heavily both from the point of view of IR and from that of FCA [see 6, x3.1 for a rather extensive review].

A ordance 7. FCA scaling implements faceted search & navigation.

Sometimes certain sets of attributes have di erent multiple possible values and/or special relationships between those values|such as hierarchies|and it is interesting for navigation purposes to see the collection of documents through the prism of those relations. This is called faceted information retrieval.

In FCA, discrete multi-valued attributes or otherwise-related attributes may be rendered in a data-dependent fashion by means of the process of scaling attributes [ 20 ]. But the e ectiveness of this process depends extraordinarily on the experience of the expert user doing the encoding of attributes.

Although faceted navigation is explicitly mentioned in [ 21 ], it seems that FaIR was the rst actual implementation using FCA [ 11, 37 ], albeit for a restricted application, thesaurus exploration. A review of faceted boolean IR can be found in [ 13 ]|as applied to Personal Information Systems|with an emphasis on usability, visualisation and navigation.

An alternative to scaling is logical concept analysis, LCA where any logical formula may be used to characterize intents [ 16 ], and it has been used to build a Personal Information Retrieval system for photos based in metadata [ 15 ]. Note that LCA is a proper generalization of FCA.

Although a number of other topics suggest themselves for this review|such as Semantic Filesystems [ 15, 30 ] or the duality of Information Pull & Push|to put them in context would demand more space than we have at our disposal. 4

Discussion: challenges of IR for FCA

Dealing with redundancy and noise in data. As in other sub elds of machine learning and pattern recognition, functions Q and D of Fig. 2 can be thought of as functions that reduce unnecesary redundancy and noise.

For instance, when dealing with text we should be aware that natural language is widely-acknowledged to be extremely redundant : many words, expressions, constructions, etc. convey the same ideas and essentially make the complexity of the system grow. Furthermore, if words are considered terms for IR, every single word encountered when tokenizing a text invokes all of the senses conventionally assigned to it in a language. Since it is these senses that are purported to mediate the actual relation between the terms and documents, serendipity may reinforce not just the originally intended sense but also some unrelated senses due to surrounding context. This is a manifestation of noise, e.g. undesired content. And these problems can only be compounded by the ubiquity of synonymy and polysemy in Natural Language.

On top of the excess complexity incurred by redundancy, it is well-known that FCA is very sensitive to the spureous absence or presence of crosses in the incidence relation between documents and terms: the addition or deletion of any such incidences may as much as double or halve the number of concepts in the lattice[ 20 ]. If FCA is to succeed in dealing with such problems it has to devise methods to cope with this kind of noise at the incidence relation level. Big data, supervised operation and training. The main challenge for FCA to be of any help to IR is scalability. Perhaps the maximum reported size for FCAinIR systems is some thousands of documents [ 39 ], while it is customary for present-day IR systems to have millions of documents. There is no easy way to overcome this inherent limitation for concept lattices: building them is just too costly in time and space[ 28 ].

One way to address the complexity of Big Data would be to assume the datadriven paradigm of Machine Learning or Pattern Recognitition [ 41 ]. However, FCA is an unsupervised machine learning technique: all of the information in the lattice stems from the information in the documents, the terms and the incidence relation between them. But the solution to Problem 2 seems to entail a supervised procedure whereby the training topic judgements can be used to improve unseen topics. At present, relevance in the boolean case is dictated a priori and there is no room for such supervision, only for post-retrieval assessment.

Unless this mismatch is addressed, machine learning-inspired techniques will still outperform FCA or address tasks which FCA simply cannot attempt. Catering to more complex IR models. The history of IR seems to be an account of progressively complex modelling of textual data. From the boolean bag-of-word models, conceived as boolean vectors, it is easy to take a conceptual jump towards softer weighting schemes in the Vector Space Model. From constant-dimension vectors in the Vector Space Model, it is an easy jump to probability-weighted formal series, that is (generative) language models. Similarly, from vector description in non-orthogonal systems of generators it is easy to conceive an orthonormal basis wherein to represent vectors, which is the essence of Latent Semantic Indexing, and so on. All such conceptual leaps are steps in a process of continual algebraization of the underlying models that entail better modelling or learning capabilities in IR.

Such a process has barely begun in FCA with the so-called generalizations of FCA, [ 1, 2, 43 ]. Nevertheless, coincidences can be seen in all such evolutions: it seems that the basis for any possible generalization of FCA is the theory of residuated semirings [ 36 ], while many of the models in IR have semiring-based costs (probabilities, log-probabilities, etc.)

In a similar tone, most of the implementations of FCA in IR deal with the conjunctive querying case, with the previously noted exception of [ 32 ], which implements a sort of disjunctive model. If FCA wants to embrace all possible \conceptualization modes" for queries, it needs to standardize and use habitually the whole gamut of Galois connections available [ 44 ].

A concluding note. . . On the one hand, the FCA community has an increasing collective expertise in the development of IR applications (FCA in IR) in di erent domains and tasks, but has achieved only limited impact in IR proper, for the reasons explained above among others.

On the other hand, FCA has strong theoretical foundations that can help IR understand better its own models and basic assumptions (FCA for IR). Yet FCA would very much pro t by the assessment-oriented approach to task-solving now prevalent in the eld of IR. It would seem FCA only needs to embrace the new generalization e orts outgrowing from the dynamic ourishing of FCA these past 15 years to do so.

At the risk of being too poetical, since IR is highly empirical (and in the quest for rmer theoretical grounding) and FCA highly theoretical yet completely data-driven (but still needs to come to terms with task-realities) there is still hope for a middle ground/sweet spot where someday the twain may meet.

References

[1] R. Belohlavek . Fuzzy Galois connections . Math. Logic Quarterly , Jan. 1999 .

[2] A. Burusco and R. Fuentes-Gonzalez . The study of the L-fuzzy concept lattice . Mathware & soft computing, Jan . 1994 .

[3] C. Carpineto and G. Romano. ULYSSES : A lattice-based multiple interaction strategy retrieval interface . In Human-Computer Interaction , 5th International Conference, EWHCI' 95 , pages 91 { 104 , Berlin, Heidelberg, 1995 . Human-Computer Interaction, 5th International Conference, EWHCI' 95 .

[4] C. Carpineto and G. Romano. Exploiting the potential of concept lattices for information retrieval with CREDO . Journal of Universal Computer Science , 10 : 8 , 2004 .

[5] C. Carpineto and G. Romano . Concept Data Analysis . Theory and Applications . John Wiley & Sons, Ltd, Chichester, UK , Sept. 2005 .

[6] C. Carpineto and G. Romano . Using concept lattices for text retrieval and mining . In Formal Concept Analysis , pages 161 { 179 . Springer Verlag, 2005 .

[7] C. Carpineto and G. Romano . A Survey of Automatic Query Expansion in Information Retrieval . Journal of the ACM , 44 ( 1 ):1{ 50 , Jan . 2012 .

[8] C. Carpineto , A. Della Pietra , S. Mizzaro , and G. Romano . Mobile clustering engine . In M. Lalmas, editor, Advances in Information Retrieval , pages 155 { 166 . Springer, 2006 .

[9] J. M. Cigarran , J. Gonzalo , A. Pen~as, and F. Verdejo . Browsing search results via Formal Concept Analysis: Automatic selection of attributes . In Concept Lattices, volume LNAI 2961 , pages 311 { 331 . Springer, 2004 .

[10] J. M. Cigarran , A . Pen~as, J. Gonzalo, and F. Verdejo . Automatic selection of noun phrases as document descriptors in an FCA-based information retrieval system . In Ganter and Godin [18] , pages 49 { 63 .

[11] S. Dominich . The Modern Algebra of Information Retrieval. The Information Retrieval Series . Springer, 2008 .

[12] J. Ducrou and P. Eklund. SearchSleuth: The conceptual neighbourhood of an web query . In Proceedings of the 5th International Conference on Concept Lattices and their Applications , (CLA07) , pages 253 { 263 , Montpellier , France, Oct. 2007 .

[13] J. Ducrou and P. W. Eklund . Faceted document navigation . In P. Hitzler and H. Scharfe, editors, Conceptual Structures in Practice. Chapman & Hall / CRC Press, 2009 .

[14] J. M. Fernandez-Calabozo , C. Pelaez-Moreno , and F. Valverde-Albacete . WebGeneKFCA: an On-line Conceptual Analysis Tool for Genomic Expression Data . In L. Szathmary and U. Priss, editors, Concept Lattices and Applications (CLA 2012 ), pages 345 { 350 , Oct . 2012 .

[15] S. Ferre . CAMELIS: Organizing and Browsing a Personal Photo Collection with a Logical Information System . Proceedings of the 5th International Conference on Concept Lattices and their Applications , (CLA07) , pages 112 { 123 , Sept . 2007 .

[16] S. Ferre and O. Ridoux . Introduction to logical information systems . Information Processing and Management , 40 : 383 { 419 , Jan . 2004 .

[17] N. Fuhr . Probabilistic models or information retrieval . The Computer Journal , 35 ( 3 ): 243 { 255 , 1992 .

[18] B. Ganter and R. Godin, editors. Proceedings of the 3nd International Conference on Formal Concept Analysis , ( ICFCA2005 ), number 3403 in LNAI, Berlin, Heidelberg, February 2005 . Springer.

[19] B. Ganter and L. Kwida, editors. Supplementary Proceedings of the 4th International Conference on Formal Concept Analysis , ( ICFCA2006 ), February 2006 . Verlag Allgemeine Wissenschaft.

[20] B. Ganter and R. Wille . Formal Concept Analysis: Mathematical Foundations . Springer, Berlin, Heidelberg, 1999 .

[21] R. Godin , E. Saunders, and J. Gecsei . Lattice model of browsable data spaces . Information Sciences , 40 : 89 { 116 , 1986 .

[22] R. Godin , J. Gecsel , and C. Pichet . Design of a browsing interface for information retrieval . In Proceedings of the 12th International Conference on Research and Development in Information Retrieval (ACM SIGIR '89) , pages 32 { 39 , Cambridge, MA, 1989 . ACM.

[23] R. Godin , G. Mineau, R. Missaoui , and H. Mili . Methodes de classi cation conceptuelle basees sur les treillis de Galois et applications . Revue d'intelligence arti cielle , 9 ( 2 ): 105 { 137 , 1995 .

[24] T. Hannan and A. Pogel . Spring-based lattice drawing highlighting conceptual similarity . Formal Concept Analysis , pages 264 { 279 , 2006 .

[25] M. Hearst . Search User Interfaces . Cambridge University Press, 2009 . ISBN 9780521113793.

[26] B. Koester . FooCA: Enhancing Google information research by means of Formal Concept Analysis . In Ganter and Kwida [19] , pages 1 { 17 .

[27] D. Kuropka . Modelle zur Reprasentation naturlichsprachlicher Dokumente. Ontologie-basiertes Information-Filtering und -Retrieval mit relationalen Datenbanken . Advances in Information Systems and Management Science . logos-verlag.de, 2004 .

[28] S. O. Kuznetsov and S. A. Obiedkov. Comparing performance of algorithms for generating concept lattices . Journal Of Experimental & Theoretical Arti cial Intelligence , 14 ( 2-3 ): 189 { 216 , Apr . 2002 .

[29] C. D. Manning , P. Raghavan , and H. Schutze. Introduction to Information Retrieval. Cambridge University Press, 2008 .

[30] B. Martin . Formal concept analysis and semantic le systems . In P. Eklund, editor, Concept Lattices , pages 88 { 95 . Springer Berling Heidelberg, 2004 .

[31] C. T. Meadow , B. R. Boyce , and D. H. Kraft . Text Information Retrieval Systems. Library and Information Sciences . Academic Press, San Diego and San Francisco, second edition, 2000 .

[32] N. Messai , M. D. Devignes , A. Napoli , and M. Smail-Tabbone. BR-Explorer: An FCA-based algorithm for Information Retrieval . In Proceedings of the 4th International Conference on Concept Lattices and their Applications , CLA '06 , 2006 .

[33] C. Mooers . A mathematical theory of language symbols in retrieval . In Proceedings of the International Conference on Scienti c Information , Washington, D.C., 1958 .

[34] D. A. Norman . A ordance, conventions, and design . Journal of the ACM , 6 ( 3 ): 38 { 43 , May 1999 .

[35] R. Pedraza-Jimenez , F. J. Valverde-Albacete , and A. Navia-Vazquez . A generalisation of fuzzy concept lattices for the analysis of web retrieval tasks . In Proceedings of the International Conference on Information Processing and Management of Uncertainty in Knowledge-Based Systems, (IPMU'06) , July 2006 .

[36] S. Pollandt. Fuzzy-Begri e. Formale Begri sanalyse unscharfer Daten . Springer, Heidelberg, 1997 .

[37] U. Priss . Lattice-based information retrieval . Knowledge Organization , 27 ( 3 ): 132 { 142 , 2000 .

[38] U. Priss . Formal concept analysis in information science . In B. Cronin, editor, Annual Review of Information Science and Technology (ARIST) , pages 521 { 543 . Information Today , Inc., Jan . 2006 .

[39] T. Rock and R. Wille . Ein TOSCANA-Erkundungssystem zur Literatursuche . In Stumme and Wille [40] , pages 239 { 253 .

[40] G. Stumme and R. Wille, editors. Begri iche Wissensverarbeitung; Methoden und Anwendungen , Heidelberg, 2000 . Springer.

[41] S. Theodoridis and K. Koutroumbas . Pattern Recognition. Academic Press, third edition, 2006 .

[42] F. J. Valverde-Albacete . Combining soft and hard techniques for the analysis of batch retrieval tasks . In E. Herrera-Viedma, G. Pasi , and F. Crestani, editors, Soft Computing for Information Retrieval on the Web. Models and Applications , volume 197 of Studies in Fuzziness and Soft Computing , pages 239 { 258 . Springer, 2006 . ISSN 1434 -9922 (print edition) . ISSN (electronic edition) 1860 - 0808 , ISBN- 10 3 - 540 -31588-8, ISBN- 13 978 - 3 - 540 -31588-9.

[43] F. J. Valverde-Albacete and C. Pelaez-Moreno . Towards a generalisation of formal concept analysis for data mining purposes . Concept Lattices. Proceedings of the International Conference on Formal Concept Analysis (ICFCA 06) , LNAI 3874 :161| 176 , Dec 2006 .

[44] F. J. Valverde-Albacete and C. Pelaez-Moreno . Extending conceptualisation modes for generalised Formal Concept Analysis . Information Sciences , 181 : 1888 {1909, May 2011 .

[45] C. van Rijsbergen . The Geometry of Information Retrieval . Cambridge University Press, Cambridge, New York,Melbourne, Madrid and Cape Town, 2004 .