Ghosts in the Machine: An AI Treatment of Medieval History

Introduction

The ghosts of the title are characters of · frenchmediaeval history. The project described here (the RESEDA* project) is an attempt to design a question answering system where the data basewhich is to provide the raw material for the answers con­ sists of biographical data, culled from a variety of contemporary and modern sources, concerning the humanist literary movement in France. Questions to the system are input in a strict formalization. Answers, which appear in the same format, are found by interrogation of the data base or by inference making on the contents of the data base. The pecu­ liar difficulties presented by the type of data dealt with and by the methods of inference needed are great enough to justify neglecting, or at least, postponing, the problem of natural language input.

Section I of this paper fills in the back­ ground by trying to explain some of the peculiari­ ties of the data, Section II to shew how these h~ ve been dealt with at the representation level by taking an example text and shewing how it has been coded. Section Ill demonstrates some of the infe­ rence techniques used by following in detail the process of finding an answer to a question about a piece of text presented in Section II, and the conclusion sketches further planes.

Section I : Problems presented by the data

The difficulties presented by the data break down into two main classes. First there are those which are a direct consequence of dealing withhis­ torical data. Secondly there are problems inhe­ rent in any attempt to deal with data which expresses not only facts but beliefs, attitudes and complex relationships.

* The RESEbA project is supported by the Centre National de la Recherche Scientifique and by the Delegation Generale a la Recherche Scientifique et Technique (contract 75.7.0456)

Let us first consider the problems of histo­ rical data. Most obviously, such information is incomplete. It is in the nature of things impossi­ ble to know every relevant fact about an historkal period.Thus it is often necessary to infer possible or probable facts from what we do know. This is a general characteristic of the system. In a much stronger sense than that normal in natural language processing, its inferences are only probable infe­ rences and its answers to questions, except in cases where some well-attested piece of informacion is asked for, are only probable answers.

Secondly the data is subjective. An authority reports what he believes to be the case, not what necessarily is the case. Sometimes this may lead him to report as fact something which is known, from other sources, to be actually false. And ye4 since that false statement may affect other evid­ ence, it must be represented in the data base. When this happens the false assertion is explicitly marked as such. A related problem arises withcoun­ ter-factual conditionals. Many pieces of evidence take the form ' If such-and-such had happened (al­ though it did not), the consequences would have been thus '.Evidence like this clearly gives im­ portant information and must be represented.

The subjectivity of the data gives rise to the further problem of contradictory data. It is quite possible to find one authority directly contradic­ ting another. When this happens, each of the contr­ adictory versions must be separately represented, and a special list kept of such contradictions. Then, when finding an answer to a question involves using a contradicted piece of evidence, the stra­ tegy which allowed the answer to be found must be repeated with each and all of the contradictory

versions of the same evidence.

All this can be summarized by sa~ing that the data is, in its very essence, uncertain, a fact which reinforces the aspect of question answering touched on earlier. The answers found are, often, plausible answers based on plausible inferences us­ ing uncertain data. This should not be interpreted as an excuse for producing random rubbish. The system is intended to provide specialists in the period with an interactive aid in their research. It must therefore perform at least as well as such a specialist working by hand in a mechanical fashion. This means that, whilst the system cannot be expected to have the intuition and general back­ ground of the historian, it must not give impossi-· ble or silly information.

The other difference between our basic mate­ rial and that of more normal AI systems is that, being biographical, it deals primarily with the attitudes, beliefs and ideology in political and religious matters of the characters who are our concern. This means that the representation of the data base must be adequate to express abstract ideas and processes and that the inference algo­ rithms also must take into account propositional attitudes. Thus they cannot depend on a tidy logic like that of the first order calculus. To illus­ trate this consider the following example : 'John dislikes apartheid and hence does not want

South African rugby teams to play in England'. Most people would agree that there is some fairly strict causal relation between the two halves of this sentence, marked by 'and hence'. But whatever this relation is, it cannot be interpreted as mate rial implication, since the converse

'John wants South African rugby teams to play in England and hence does not dislike apartheid' does not necessarily hold.

The issue becomes even more complex when it ~ considered that we are interested not only in find­ ing relationships between people and the ideas they hold, but also in relationships between people. New relationships are established initially because a question has been asked the answer to which depends on inferring the new relationship. But when a new relationship has been found, it may be added to the data base by asking a 'system' question, provi­ ding that its validity has first been checked by an historian. Thus RESEDA is envisioned as a contr nually growing data base which itself proposes pos­ sible additions to its knowledge.

It is worth noticing that the sort of problems just discussed are not the result of using histo­ rical data. Anyone who tried to design a system dealing with modern american politics would be fa­ ced with exactly similar problems. Thus, if our examples with names like 'Pope Benoit XIII' and controversial topics like 'schism in the church' strike the modern ear oddly, it is worth remember­ ing that 'Benoit XIII' could be replaced by 'Jimmy Carter' and 'schism' by 'the middle-east question' without essentially altering the basic problems discussed here.

Section II :representation of data

It should be said immediately that the level of 'understanding' aimed at here is not the 'deep' understanding of Charniak (Charniak 1977) and others. If someone takes a trip on a boat we are content to say simply that, without worrying about the exact description of the boat. This is not to deni­ grate systems using such very detailed knowledge. Simply their purpose is different.

On the other hand, some detailed knowledge which we do have is very specific to our data base. For example, 'civil war' refers only to the war between the Armagnacs and the Bourguignons, fought between 1392 and 1425 with the intention of gain­ ing central power in France.

With this in mind, we can now take an example piece of text and see how it would be represented.

Figure Ia gives the french text. Figure lb gi­ ves a somewhat free translation of the same text; the numbers in square brackets mark a rough corres­ pondance between parts of the text and the formal representation which follows later in figure 2,

'Les ambassadeurs de l'Universite etaient por­ teurs d'une lettre pour Benoit XIII, redigeepre­ cipitamment par Clamanges, redacteur habituel des lettres de l'Universite depuis 1394, le jour meme du depart (14 avril 1395). Cette lettrecon­ tenait presque une approbation des voies de con­ cile et de convention. Il n'est pas etonnant qu' une lettre redigee en toute hate ait reflete les positions personnelles de son auteur plutot que celles preconisees par l'Universite. Ils avaient cru devoir la garder par devers eux. Quand ils furent de retour, l'Universite elle4meme prit soin de corriger cette epitre en en retranchant tout ce qui ne tendait pas a l'eloge exclusif de la voie de cession -26 aout 1395 (d'apres Valois 1891-1902, III pp. 70-71, Ornato 1969, pp. 25-26)

Figure Ia

The University's ambassadors [1,2] carried a letter for Benoit XIII [3,4]which had been hastily drafted by Clamanges [5], the regular drafter of University letters after june 1394 [6] on the same day as their departure (14 avril 1395) [5]. The contents of this letter [7] almost constituted approval of the way of general co~n­ cil (to resolve the schism) and arbitration (be~ ween the popes). It is not surprising that a let­ ter drafted in great haste reflected its author's personal position rather than that advocated by the University [8]. They (the ambassadors) belie­ ved they should keep the letter to themselves [9]. When they returned the University took care to cor­ rect the letter [10,11], cutting out anything which did not lean towards approbation of abd~ tion (by both popes) [12] (26 august,l395) [11].

Figure lb

General discussion of the representation

It is impossible to describe our representa­ tion fully in a paper of this lenp,ht, even if we restrict ourselves to features exemplified in . the representation of the text given in figure I. How­ ever, some idea of its basic characteristics canbe given fairly briefly . Figure 2 shews our represen­ tation of the text given in figure 1. The reader is not expected to be able to follow the represen­ tation at first glance; the rest of this section will be devoted to explanation. It should be noted too, that for reasons of clarity, the syntax of the representation used here differs from that used in practice (for a full description of the latter, see Bozzolo et al 1976). There is, of course, a strict correspondance between the two.

The passage is represented by a set of 'planes' one for each episode of the text. (For 'planes' in this sense, see Quillian 1968 and Scragg 1975). A brief english description of the contents of each episode precedes the plane representing it.

The predicate is the first element of the plane, followed by a list of its arguments (enclo­ sed in the first set of braces), dating information

(in the second set of braces) and finally the bi­ bliographic source of the data.

<MODAL=lettre-officielle-1 > <SOURCE=Universite-de-paris Paris> <DEST=Benoit-XIII>} {(14-avril-1395) () } (Ornato)] *8. (False). That the University of Paris supported Benoit XIII in his policy of arbitration. [pour+AVOIR-ATTITUDE {<SUJ=Universite-de-paris Paris> <OBJ=Benoit-XIII Avignon> <ARG=voie-de-convention>} {(14-avril-1395) ()} (Valois)] 9. Benoit XIII was not given the official letter by the ambassadors. [neg+ETRE-AFFECTE {<SUJ=Benoit-XIII Avignon> <OBJ=lettre-officielle-1 > <SOURCE=Vedettes-1 >} { (14-avril-1395)(avant-26aout-1395) } (Valois)] 9 -Benoit's not being given the letter- was the re­ sult of Clamanges having falsely represented (in7) the policy of the University (given in 11) (corre­ lator- poi nter) 10. <CAUSE2 10 (COORD2 7 11) > On "'the 26th august, 1395, the University re-wrote the official letter to Benoit XIII. [rep+PRODUIRE {<SUJ=Universite-de-paris Paris> <OBJ=lettre-officielle-1 > <DEST=Benoit-XIII Avignon>} {(26-aout-1395) ()} (Valois)] The re-writing (10) was the result of Clamanges having falsely represented (in 7) the views of the University (11) (correlat or-pointer) 11. <CAUSE2 10 (COORD2 7 11 > On the 26th august, 1395, the University of Pa ris wanted to inform Benoit XIII, in Avignon, of their position as expressed in 12. [int+ment+DEPLACER {<SUJ=Universite-de-paris Paris> <OBJ=12> <DEST=Benoit-XIII Avignon>} {(26-aout-1395) ()} (Valois)] 12. Between 1395 and the beginning of 1408 t he Uni versity of Paris opposed the Avignon popes on the subject of their policy of arbitration. [contre+AVOIR-ATTITUDE {<SUJ=Universite-de-paris Paris> <OBJ=papes-d'avignon Avignon> <AR�=voie-de-convention~ {(1395) (debut-1408)} · (consensus) r

1. On the 14th april, 1395, an anonymous group of people, in Paris, held an official position in the University, at Paris. [soc+cons+ETRE-AFFECTE {<SUJ=Vedettes-1 Paris> <OBJ=universite Paris>} {(14-avril-1395) ()} (Valois)] 2. This same group were members of a deputation. <REL5 Vedettes-1 ambassade+membre> ~he relationship (being members of a deputation) specified in 2 is brought about by the episode ~eported in 3 (correZator- pointer) <CONFER 2 3> 3. The University of Paris sent a deputation to Beno1t XIII at Avignon which left Paris on 14th april 1395 and returned sometime before 25 august 1395. [DEPLACER {<SUJ=ambassade Paris> <OBJ=ambassade Paris> <SOURCE=Universite-de-paris Paris> <DEST=Beno1t-XIII Avignon>} {(14-avril-1395)(avant-25-aout-1395)} (Valois)] 4. The deputation took an official letter from the University of Paris to Benoit XIII. [DEPLACER {<SUJ=ambassade Paris> <OBJ=lettre-officielle-1> <SOURCE=Universite-de-paris Paris> <DEST=Beno1t-XIII Avignon>} {(14-avril-1395) ()} (Valois)] ~- Shortly before 14 april, 1395,Clamanges, in Paris, wrote this official letter from the Uni­ versity of Paris to Beno1t XIII at Avignon. [soc+ment+PRODUIRE {<SUJ=Clamanges Paris> <OBJ=lettre-officielle-1 > <SOURCE=Universite-de-paris Paris> <DEST=Benoit-XIII Avignon>} {(peu-avant-14-avril-1395) (14-avril-1395)} (Ornato)] 6. From june 1394 onwards Clamanges, as part of his official duties, drafted many letters on behalf of the University of Paris. [so~+mult+ment+PRODUIRE {<SUJ=Clamanges Paris> <OBJ=lettre-officielle> <SOURCE=Universite-deparis Paris>} {(juin-1394) ()} (Ornato)] 7. On the 14th april, 1395, Clamanges, in Paris, expressed the idea specified in plane 8 (in the official letter already mentioned) to Benoit XIII as bein~ the idea of the University. [soc+ment+DEPLACER {<SUJ=Clamanges Paris> <OBJ=8> Figure 2

The planes of figure 2 make up the data base for the system. The data base is organised logical­ ly into 'volumes', with each volume containing re­ ferences to all the planes representing episodes concerning one particular character (a 'star'). Since any particular episode may concern more than one star, a reference to the same plane may appear in several different volumes (cf. Scragg 1975).

The stars are the more important characters of the period, but are not necessarily individuals. For example, in the text given above, the follow­ ing stars appear :

1. Benoit XIII and Clamanges, who are both indivi­
2. duals and therefore 'normal' stars.
3. The University of Paris, which is considered to be a 'moral' person. That is, it may take deci­ sions, take part in juridical actions, own pro­ perty etc, just like an individual person.
4. The Popes of Avignon. These are a 'collection' of characters, defined extensionally.
5. The delegates from the University to Benoit XIIL The RELS (in plane 2) specifies that this group of people were all members of the same deputa­ tion to Benoit XIII. By its nature, the group is temporary. How long such a group lasts and the events in which the group is involved is shewn by attaching the RELS plane to the plane report­ ing the episode which brings the group into exis tence (done by the CONFER correlator-pointer).­ In the case of RELS, any relevant plane which has 'deputation' as its SUJ, gives further in­ formation about this particular group (for example, plane 4). ·
6. Vedettes-1 (Stars- I). There are several anonymous groups like this. One volume serves for all such groups, who are distinguished one from another by the index number after vedettes.

Each plane, apart from the RELATION plane al­ ready mentioned, also contains dating information. The two-element list towards the end of the plane gives the date at which the episode starts followed by the date when it finishes. Thus, if we take pl~ ne 3 as an example :

{(14-avril-1395)(avant-25-aout-1395)}

tells us that the deputation left Paris on the 14th april and returned before the 25th august, 1395. Dating is not always as simple as this. If the ac­ tion of the predicate continues over a period of time, more than one date may be given. For example, in 'it is known that he was Canon at Lille in 1398, and he is addressed by the same title both in 1390 and in 1391' it is importanttorecordall.three dates.

The last item in each plane is the bibliogra­ phic source from which the episode is taken. In plane 3, our authority is Valois'treatise cited in figure I.

Place information is very important, and is treated by attaching to each of the chief arguments of the predicate a location slot, which is filled when the physical location of the instantiation of that argument is known. Thus, to use plane 3 as an example again, <SUJ=ambassade Paris> tellsusthat the deputation was in Paris at the beginning of the episode <OBJ=ambassade Paris> that it was also in Paris at the end of the episode. (This interpreta­ tion is partially determined by the predicate used, as will be discussed later). The SOURCE o'f the epi­ sode (University-of-paris) is also in Paris, and the DESTination, Benoit-XIII,in Avignon. Not all arguments to the predicate may have their location specified : only SUJ, OBJ, SOURCE, DESTinationand ARGument. As with dating, it may be sometimes neces sary to specify more than one location for a giver! argument.

When the description proper of an episode is considered, the most characteristic is that itcon­ sists of a predicate with arguments. Of the argu­ ments (called by us correlator), six correspond more or less to conventional cases: SUBJECT,OBJEC~ DESTINATION, MODALITY and ARGUMENT. Others, like CAUSE2, COORD2 and CONFER in the example text, are peculiar to this representation. In general, it would be a mistake to identify any of the correla­ tors too strongly with conventional cases: the 'mean ing' of a correlator is given in the specialistp~ gram attached to the predicate governing thecorre: lators and which makes use of the correlators when doing inferences. Their names can give only a very rough intuitive feel for what they mean.

Not all correlators are obligatorily present, even for the same predicate (cf. planes 7 and II, where the predicate is in both cases DEPLACER (to move) modified essentially in the same way, but where the correlators are different). Thus, when a parti­ cular correlator is used, its name must appear. Ri­ gid rules specify the argument frames for eachpre­ dicate, determining not only what correlators must appear in a particular argument frame for that predicate or its modifications, but also what sort of entities can fill any particular argument. The argument frame for 'rep+PRODUIRE' is a fairly simple example :

[rep+PRODUIRE {<SUJ=Vedette> <OBJ=lexical item>B specifies that 'rep+PRODUIRE' (which can be roughly translated as 'to re-do' or 'to correct') musthave at least two correlators SUJ and OBJ and that the SUJ slot must be filled with a star, the OBJ slot by a lexical item.

The predicates themselves should now be con~­ dered. There are at the moment only five, although nothing is fixed or magical about the number. Alth­ ough we intend to keep the number of predicates as small as possible, more may well prove necessary. DEPLACER(to move), AVOIR-ATTITUDE (to have an at­ titude), PRODUIRE (to produce), ETRE-AFFECTE (tobe affected by) and ETRE-PRESENT (to be present) are chosen to be mnemonic. Nonetheless they correspond only minimally to the normal use in natural language of the same verb. Once again, the real meaning of the predicates is given by the specialist attached to each predicate. A better idea of what the predi­ cates represent will be gained when we do, finally, take a closer look at the example planes.

Intuitively, five predicates alone are not ade quate to represent the sort of complex data with which we want to deal, and a few minutes experimen­ tation with the example text will soundly reinforce intuition. A set of 'modulators' are therefore ured to modify the sense of the predicate. (cf. Mel'chuls Zholkovskij 1970). These modulators, in the graphic representation used in this paper, appear in small letters to the left of the predicate, attached to it by a + sign. The 'rep' in the argument frame a­ bove is, in fact, a modulator, and adds a sense of ~oing over again' to the basic 'produce' sense of PRODUIRE. Another example can be found in plane 5, where 'ment' adds a sense of 'intellectual activity' to the same predicate. Modulators may be combined, as in plane 6, where 'soc+mult+ment' attached to PRODUIRE add, respectively, the notion that theac­ tion of the predicate is part of the career of the SUJ (soc) (i.e. the sort of thing he might put in his curriculum vitae), that the action is an action done before, may be several times (mult), and the same notion of intellectual activity just discussed (ment).

Discussion of the planes

After this rather lengthy general discussion, we now have sufficient background to look at the representation of the example text in some detail.

Plane I needs very little comment. Vedettes-1 has already been explained. The predicate 'soc+const +ETRE-AFFECTE' with an object attached does, how~~ deserve some explanation. The 'soc' simply says that the being-affected was, in some way, part of the of­ ficial career of the ambassadors. 'canst' is a littl e more complicated. The predicates are essentially static in their nature, so there has to be some way of representing changes of state (cf. Abels:m 1975). This is done by three modulators : incep, fin and canst. 'incep' marks the creation of a new state, 'fin' the end of the state, 'canst' that the state is, at the time specified, in full force. All three relate to the first time slot in the plane. When any one of the three is present in the descriptive part of the plane, the second time slot remains emp­ ty. Thus plane I does not tell us that a group of persons were given a position on april 14th, but that by that date they already had the position.

The use of OBJ in plane I illustrates one si­ tuation where a correlator does not map in a obvious way onto a conventional .case with the same name, and of how the meaning of any particular correlator is in part determined by the predicate with which it is associated. ETRE-AFFECTE is used when the OBJ can be conceived of as being, at least temporarily or provisionally, a characteristic of the SUJ.

'universite' (unlike the 'moral' person Univ~ site-de-paris) is a lexical item, which means that its meaning is defined not by a program (as is the case with predicates, correlators and modulators) but by relationship with other lexical items (Orna­ ta, Zarri 1976).

Plane 3 is fairly clear. DEPLACER (to move) is always transitive, and always involves a sense of actual distance covered. The modulator 'ment' is used when it is something expressed (by means of a letter or whatever) which is 'moved' from the SUJ­ ject to the DESTination (cf. plane 7). If a SOURCE correlator is used with DEPLACER unmodified, it is interpreted as 'at the instigation of'. Similarly, in plane 5, the SOURCE with 'ment+PRODUIRE' can be interpreted as 'on the orders of' or 'on behalf of'.

Plane · 7 offers an example of the use of a 'can­ pletive' plane. The OBJect is itself described by an independant plane, whose number fills theOBJect slot. The completive plane itself, plane 8, is mar­ ked with an asterisk to shew that the assertion it represents is false. As was said in the introductio~ this is a situation common in dealing with our type of data. The inference routines too must take ac­ count of false assertions.

The <CAUSE2 10 (COORD2 7 II)> appearing af­ ter plane 9 is (like the CONFER line after plane2) an indication of relationships between planes. COORD2 associates two planes of the same overall type which are both governed by the same correla­ tor-pointer. In fact, COORD2, and its parallelcor­ relator COORD! which functions inside planes, are equivalent to first order predicate calculus 'and', with the restriction that only similar things may be conjoined. Another pair of correlators, ALTERN! and ALTERN2, are equivalent to exclusive 'or'.

CAUSE2 specifies that its first argument is the result of its second argument. A parallel cor­ relator-pointer FINAL is used when the result of an episode seems probable but is not yet confirmed by the evidence so far available. For example, 'Charles VI sent the Comte de St. Pol to arrest Pierre d'Ailly at Cambrai' would give rise to three planes, one reporting that Charles VI sent the Count to Cambrai, another that the Count arre~ ted Pierre d'Ailly, and the third that Pierre d'Ailly was, at some time later, in prison. But the latter two planes would be marked as only probable by the use of FINAL instead of CAUSE2 to linkthese three planes together. RESEDA is a very cautious inference maker. It will not make inferences based on probable information. On the other hand, if a plane marked as probable is later confirmed, it will be up-graded to certain by changing the poin­ ter and the order of the arguments.

The remainder of the coding contains verylit­ tle that has not been already commented on. Plane II introduces a new modulator 'int' which gives a sense of intention or of 'wanting that' to the pre­ dicate to which it is attached.

AVOIR-ATTITUDE in plane 12 means, roughly, 'feels about' or 'has an attitude towards', and the modulator 'contre', used only with this predi­ cate, makes it into 'opposes'. A VOIR-ATTITUDE gives us a good example of how one predicate can be use~ with appropriate modification, to represent quite subtle notions. As one might expect, it bears most of the burden of representing relationships between people. With appropriate modulators and argument frames it has been used to express all of the fol­ lowing :

Someone supports (opposes) someone (optionally, about something).

Someone promises (threatens) someone with somethin~

Two (or more) persons are involved in litigation (in opposition or on the same side) concerning some matter.

Two (or more) persons oppose each other by violent means.

Someone (not immediately involved), supports oro~ poses someone in a legal matter.

Someone wants something to come about.

Section Ill :using the data base to answer questions

As was said earlier, the sort of understanding we are aiming at is not the very deep detailed un­ derstanding of some other systems. Nonetheless we hope to be able to answer most of the type of ques­ tions a historian would be likely to ask about the people mentioned in our data base. This involves an attempt to write deduction programs based on human attitudes and motives, and is, we think, sufficien­ tly difficult and of sufficient intrinsic interest to warrant the degree of effort involving in cons­ tructing RESEDA.

The main burden of the inferencing is carried by the specialists attached to each predicate. Even if it were worked out in full detail for all five predicates, it would be clearly impossible to give a full description here. But it seems worthwhile to take one single question, imagining that the da­ ta base is as given earlier (i.e. just the repre­ sentation of our example text, with some more detail attached to the lexical items involved) and shewing how the inference routines would find the answer to that question.

Let us imagine that a user asks

'Why did Clamanges, after august 1395,completely give up his activity as official secretary to the University of Paris ?'

This would have to be coded, since natural language input is not allowed, in the same formalization as that used for the data base

~in+soc+mult+ment+PRODUIRE {<SUJ=Clamanges Paris> <OBJ=lettre-officielle <SOURCE=universite-de­ paris Paris> <CAUSE I=? >}

{(apres-aout-1395) ()B

Figure 3

CAUSEI is the equivalent within planes of CAUSE2 between planes; here its value is simply a question mark, shewing that the information requi­ red by the user is the value of that argument. For the rest, the representation matches that of the data base, except that no bibliographic authority is specified. So the predicate and its instantiated arguments give the known information, the ? shews what is required.

If a plane existed in the data base which mat­ ched, item by item, the question, the answer would be found directly, and the plane or lexical item in the ? position would give the user the informa:ion he wanted. Direct match here is a little more com­ plicated than is often the case : if one of the it­ ems being searched for is a lexical item defined by a tree structure (see Ornata, Zarri 1976) the search is generalized upwards in the definition tree. No such directly matching plane exists in the present case.

Therefore the specialist attached to PRODUIRE is called upon. (It is always the predicate contai­ ned in the question which determines the specialist to be tried). The modulator 'fin' and the empty CAUSEI correlator attached to PRODUIRE together tell us that we are seeking the cause of an interruption in the predicate activity. The modulator 'ment' adds the additional information that what is involved is an intellectual activity. This points us to a sub­ part of PRODUIRE which gives information about PRODUIRE as an intellectual activity, including as­ pects like possible reasons for terminating the activity. There we find that if 'soc' is also at­ tached to the predicate and if SOURCE is present in the question, then we are justified in assuming that the SOURCE delegates power to the SUJ. All this leads to the formation of an hypothesis 'some­ one stopping working for someone else -in a situa­ tion involving the delegation of intellectual po­ wer- may be the result of the employee, who repre­ sents the employer, acting, in the course of ·his official duties, in a way not consonant with the official views of his employer'. (Employee and em­ ployer are here used only as shorthand for 'some­ one representing an official body' and 'the offi­ cial body represented' : no money need change hands).

This hypothesis is tested by searching in the data base for episodes which fit the suggestion that Clamanges may have acted in such a way with regard to the University of Paris.

The search is done by trying to find in the data base planes which match · the schemata given in figure 4, which represent a formalization within RESEDA of the hypothes~ given above. After the for­ mal representation, an english transcription of the schema is given.

VI and V2 are, respectively, the SUJ and the SOURCE of the original question, V3 and V4 are oth­ er, unspecified characters who may turn out to be the same person, q and r refer to instantiated versions of the schemata with the same name and a is a variable. The exclamation marks before q andr shew that they are contradictory (i.e. if 'pour' appears in one 'centre' must appear in the other). Obviously the dating of any planes found must fit with the question and with the other planes.

In our present case, where the data base is as shewn earlier, p and q are satisfied by planes 7 and 8. Plane 8, it will be remembered, was marked as false by the use of an asterisk, and indeed ph­ ne 12, which flatly contradicts 8, also satisfies r above. The particular form of the contradiction -plain falsity vs. established truth- allows us to conclude that it is unnecessary to search for a p~· ne to satisfy s. s is there to cater for the most general case, where the two planes satisfying p and q ·are simply contradictory but neither is known to be downright false. In that case, one must find some evidence (plane s) that V2, the original SOURCE, has informed some other person of whatever is reported in r. But if, as in this case, theatti tude of the employer is well known, there is no - need to search for other evidence.

Even after episodes have been found which sup­ port the hypothesis, it is not given to the user as a definite answer before an attempt has been made to verify it by finding some concrete action done by the employer against the employee to shew that what Clamanges did, was in fact sufficient to ac­ count for a break in the relationships between him and the University. In other words, we search for immediate reactions of the University against Cla­ manges'action. The inference rules give a number of conditions which are grounds for assuming a con­ crete manifestation of disagreement with someone. To remove a living from someone, to forbid the publication of someone's book, to contest someone's authority, for example, are all possible grounds. Amongst these conditions 'rep+PRODUIRE' (re-doing something) appears as one possible equivalent to 'real+contre+AVOIR-ATTITUDE' (the concrete manifes­ tation of disagreement), . providing that the SUJ of 'rep+PRODUIRE' was the SOURCE in a plane where the OBJ was the same as that of the 'rep+PRODUIRE' pla­ ne. The SUJ of the plane containing the SOURCE is the OBJ of the AVOIR-ATTITUDE plane.

In order to make use of this rule, we first look for any CAUSE2 pointer which contains p, i.e. 7, as its second argument. This gives us two plane~ both 9 and 10 are said to be caused by 7 (and by II, but we are only interested now in 7). 10 has universite-de-paris as SUJ and tells us that the University corrected the critical letter. This gi­ ves us t .he verification we need. If even more con­ firmation had been required, 9 would have provided it.

In the present version of RESEDA, the specia­ list programs for the predicates use a collection of fairly ad-hoc rules. Although this preliminary model despite its limited number of rules allows us to make a many useful inferences within prede­ fined contexts, we hope to develop a more flexible and more unified version. But it is clear that any further development will depend heavily on the em­ pirical work done in the first stage described hera

Conclusion

RESEDA, the system described informally here, is an attempt to apply well-established techniques from the domain of Artificial Intelligence within a useful application, whilst at the same time de­ veloping new inferences rules to account for the peculiarity of the data. It is hoped that the re­ presentation and inference methods developed in the design and implementation of this system will prove sufficiently supple to deal with any kind of biographical data. Ghosts are not necessarily old­ fashioned.