1. Introduction

The PhiloNet project is a joint effort between ITC-irst and the Department of Philosophy of the University of Bologna that aims at providing students and scholars of philosophy with new computing tools. In particular, we are developing tools for computer-aided analysis of digitized (philosophical) texts capable of overcoming the limits of traditional methods, by offering greater flexibility, analytical power, and adaptation to the goals of different users. In the field of text analysis a number of already available software tools can produce very accurate lists of words, word concordances, word frequencies, collocations and so on (Biber et al., 1998). These computational techniques are all word-based, and they hardly provide support for more sophisticated analyses, like the targeting of concepts. It is conceivable, however, that exactly the latter kind of analytical tools are more suitable for a domain, such as philosophy, where the need for accessing, elucidating, comparing, etc. concepts across and within texts is an important part of the work of scholars. Thus, new computational concept-oriented resources are needed. More precisely, resources must be available that list concepts, enumerate their mutual relationships, and point at their realization at the word level. In this work we report on preliminary results obtained by developing techniques to create concept-based concordances rather than word-based ones, by using generic and domain-specific (philosophical) wordnets as tools.

2. Semantic concordances for text-analysis

We use the term "semantic concordance" to refer to the set of text passages in which a concept occurs. Concepts are expressed in texts through words and therefore, to build semantic concordances, two lexical semantic issues must be dealt with: polysemy (when a given word expresses different concepts) and synonymy (when different words express the same concept). If one searches for a concept on the base of a word "w" expressing that concept, polysemy affects precision (the proportion of retrieved words that are actually correct) because along with occurrences of "w" expressing the desired concept, those expressing the other concepts conveyed by "w" will also be retrieved. On the other hand, synonymy affects recall (the proportion of correct words actually retrieved in answer to a search request) because all the synonyms of "w", hence other possible occurrences of the target concept, are missed. Thus, to move from a word-based to a concept-based text-analysis it is necessary to disambiguate each occurrence of a word in a text, and to have a list of the words that can express the same concept. To this purpose, WordNet (Miller,1990) turns out to be a very useful tool. WordNet is a lexical database, created at Princeton University, in which nouns, verbs, adjectives and adverbs are organized into sets of synonyms (synsets), representing lexical concepts. Synsets are linked by means of various relations, including hypo/hypernymy, meronymy, and antonymy. As far as semantic concordances are concerned, WordNet synsets can be used to extend the search for a word to all its synonyms. Moreover, the relations between synsets can be exploited to search related concepts, something which is absolutely not available with normal word-based techniques. For instance, we can search for a concept and its hyponyms or hypernyms. WordNet can also be useful for addressing the polysemy problem, as it is currently used as a tool for word sense disambiguation (WSD), both automatic and manual. In manual WSD, WordNet plays the role of extensive sense inventory; a well-known effort of this kind is the SemCor project (Miller,1993) in which a subset of the Brown Corpus was tagged with WordNet senses for educational use. Our proposal consists in manually annotating philosophical texts with synsets selected from an extension of the generic WordNet called PhiloNet (which is specialized for the philosophical domain, see below), and having a user interface that given a synset/concept permits the accessing of all the passages in which the concept is used.

3. An example of semantic concordances for philosophical texts

In order to experiment with WordNet as a tool for creating semantic concordances, we analyzed the different philosophical concepts expressed by the word "reason" in Pearson's work "The Grammar of Science". The text was digitized by the University of Bologna following the TEI guidelines and using XML-based encoding schemes. The passages of the text relevant for our example were then manually semantically tagged by using the synsets of the specialized philosophical wordnet called PhiloNet. Three different types of concordances have been tested on Pearson's work: traditional word-based concordances, semantic concordances based on the generic Princeton WordNet, and semantic concordances based on the specialized PhiloNet. A simple word-based concordance produces 125 occurrences of the word "reason(s)", including verbs and nouns relating to different senses such as:
about these conceptions we reason, endeavouring to ascertain their
ask whether we have any reason for conceiving atoms as
human powers of perception and reason, lies the mystery and the
universe to be guided by reason. But reason, because it is a directive power
...... As can be seen, with a word-based concordance it is the scholar interested in the study of the philosophical concepts of "reason" who has to single out, among the 125 occurrences found, the ones that are relevant to her purpose. We expect semantic concordances based on WordNet to deliver more focused information. In WordNet, the word "reason" occurs in six synsets, but only the one reproduced here is relevant for our philosophical purposes:
Sense 3 reason, understanding, intellect -- (the capacity for rational thought or inference or discrimination)
hypernym => faculty, mental faculty, module This synset can be considered a philosophical concept, as it refers to the British empiricist tradition starting with Locke, in which reason is broadly viewed as the generic power or faculty of knowing. Having Pearson's text tagged with WordNet synsets and by using the philosophical synset above as a search key to find occurrences of the relevant philosophical concept of reason, we obtain more focused results compared to the ones obtained with the word-based concordance. In fact, 8 occurrences of the verb "to reason" and 40 occurrences of the noun "reason", which were retrieved in the word-based concordance seen above, are now dropped because expressing non-philosophical senses. The synset-based search produces 22 occurrences of the philosophical concept as delivered by the word "reason", along with previously undetected occurrences due to synonyms of "reason": 2 occurrences due to the word "understanding", and 4 due to "intellect", as for instance in:
the only proof that our intellect has been keen enough to
in his Essay concerning Human Understanding, where he writes Furthermore, we can exploit the relations between synsets available from WordNet to obtain information about concepts semantically related to the one we are interested in. In our example we find 27 occurrences of the hypernym of "reason, understanding, intellect", that is "faculty, mental faculty, module":
the power of the reasoning faculty possesses of resuming
product of the reflective faculty, analyzing the process of perception,
....... It seems to us that even this simple experiment demonstrates the advantages and power of the semantic concordance methodology based on the current, generic WordNet. However Princeton WordNet, aiming at capturing common linguistic knowledge, contains only one philosophically relevant concept for "reason". Therefore, in a WordNet-based concordance, the other occurrences of the word "reason" in different though philosophically relevant senses are missed. The results of the concept-based concordance of "reason" can be further refined by having access to a specialized philosophical wordnet. To this purpose we used PhiloNet, a wordnet specialized in the philosophical domain, currently under development in conjunction with the University of Bologna. PhiloNet is a component of MultiWordNet (Bentivogli et al., 2000), an aligned multilingual lexical database built at ITC-irst on the basis of Princeton WordNet. PhiloNet is developed in two stages. At the first stage, philosophical synsets already existing in WordNet are identified, labeled as philosophical, and imported into PhiloNet. At the second stage, new philosophical synsets are added to PhiloNet and linked to the generic WordNet. The criterion followed for the identification and for the creation of philosophical synsets is to represent only current philosophical concepts as found in philosophical handbooks. Moreover, PhiloNet allows context labels to be attached to synsets, words, or relations to represent special pieces of information that would be hardly captured by means of the usual lexical and semantic relationships. For instance, if a philosopher or philosophical school uses a idiosyncratic term to refer to a concept that corresponds to a common philosophical concept, a context label can be added to that word in the synset specifying the philosopher/s for which the synonymy holds. Coming back to our example, PhiloNet contains the synset "reason, understanding, intellect" which has been imported from the generic WordNet and other seven philosophical synsets:

• (1) "reason", a generic synset in which the concept of reason is broadly represented as the most distinctive characteristic of human beings guiding them.
• (2) "reason, dianoia"
• (3) "reason, evidence, good sense"
• (4) "reason, self-consciousness"
• (5) "reason, calculus, computation"
• (6) "reason, foundation, essence"
• (7) "reason, law, natural law, law of nature".

We can now perform more sophisticated analyses with respect to those possible on the basis of the generic WordNet. As a result, we can see that in Pearson's work synset (1) occurs 18 times, (6) occurs 28 times, (7) occurs 51 times (8 through the word "reason", 14 through "law", 22 through "natural law", 7 through "law of nature") while the other concepts yield no occurrences. Summing up the results of our experiments with the concept of "reason" in Pearson's work "The Grammar of Science", we can see that the three types of concordance provide different levels of precision and recall of the information retrieved:

• -Word concordance: 125 lines all due to the word "reason" and not distinguishing between relevant (77) and non relevant (48) occurrences. Precision is 61.6% and recall is 61.1%.
• -WordNet-based semantic concordance: 28 lines all relevant and referring to one philosophical concept (22 due to "reason", 2 to "understand", and 4 to "intellect"). Precision is 100% and recall is 22.2%.
• -PhiloNet-based semantic concordance: 126 lines all relevant and distinguishing four different philosophical concepts. In these 126 lines, 77 are due to the word "reason", 2 to "understanding", 4 to "intellect", 14 to "law", 22 to "natural law", and 7 to "law of nature". Both precision and recall amount to 100%.

4. Conclusion and prospects

In this paper we have introduced and discussed two levels of concept-based textual analysis: a "linguistic level", relying on a generic WordNet, and a "domain level", requiring a philosophical WordNet. These two levels should be carefully identified and delimited. In particular, it is important that the domain-based wordnet contains only concepts that are widely accepted in the field. Respecting these guidelines will guarantee that the tagging performed on a philosophical text can be used by students and scholars of different theoretical background. At still another level, very specialized wordnets relative to single books make sense, and can be used in the proposed methodology. However, in this case the theoretical background and aims of the scholar building the wordnet and tagging the text become crucial. In a way, the resulting tagged text should be seen more akin to an essay than to a reference work. On the other hand, pushing the metaphor a little further, linguistic and domain-based text analysis can be compared with general dictionaries and philosophy handbooks, respectively. In conclusion, the preliminary results of our research support the feasibility and importance of a concept-based text analysis.