| { |
| "paper_id": "P85-1028", |
| "header": { |
| "generated_with": "S2ORC 1.0.0", |
| "date_generated": "2023-01-19T09:39:29.488581Z" |
| }, |
| "title": "Explana~..: 3tructures in XSEL", |
| "authors": [ |
| { |
| "first": "Karen", |
| "middle": [], |
| "last": "Kukich", |
| "suffix": "", |
| "affiliation": { |
| "laboratory": "", |
| "institution": "Carnegie-Mellon University Pittsburgh", |
| "location": { |
| "postCode": "15213, 412-578", |
| "region": "PA" |
| } |
| }, |
| "email": "" |
| } |
| ], |
| "year": "", |
| "venue": null, |
| "identifiers": {}, |
| "abstract": "", |
| "pdf_parse": { |
| "paper_id": "P85-1028", |
| "_pdf_hash": "", |
| "abstract": [], |
| "body_text": [ |
| { |
| "text": "Expert systems provide a rich testbed from which to develop and test techniques for natural language processing. These systems capture the knowledge needed to solve real-world problems in their respective domains, and that knowledge can and should be exploited for testing computational procedures for natural language processing. Parsing. semantic ,nterpretation, dialog monitoring, discourse organization, and text gef,eration are just a few of the language processinq problems that might takeadvantage of the pre.structured semantic knowledge of an expert system. In particular, the need for explanation generation facilities for expert systems provides an opportunity to explore the relationships between the underlying knowleqge structures needed for automated reasoning and those needed for natural language processing.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "One such exploration was the development of an explanation generator for XSEL, which is an expert system that hellos a salesperson in producing a purchase order for a computer system [10] . This pager describes a technique called \"link-dependent message generation\" that forms the basis for explanation generation in XSEL. and each evaluation function is a sum of some weighted terms.", |
| "cite_spans": [ |
| { |
| "start": 183, |
| "end": 187, |
| "text": "[10]", |
| "ref_id": null |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "The terms of the evaluation function for the class disk, for example, include price, disk-pack-type, storage-capacity, average-access-time, peak-transfer-rate, and handednesa. For every candidate, XSEL computes a rank value for each term in the evaluation function. The rank value for a term is the product of the candidate's normalized SCore for the term and a weight which represents an importance factor. The essential information needed to compute a rank value for a term for a candidate is stored in a rank element, an example of which is shown in Figure 1 -2.", |
| "cite_spans": [], |
| "ref_spans": [ |
| { |
| "start": 553, |
| "end": 563, |
| "text": "Figure 1", |
| "ref_id": "FIGREF3" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "(RANK tRANK-NAME AVERAGE.ACCESS-TIME tNAME RA60-AA\" tCLASS DISK fRANK-VALUE -3 tCOEFFICIENT t tVALUE 50 tlMPORTANCE I tTOKEN G:9) A second obvious type of query asks why a certain fact has whatever value it has. e.g., why total-disk.space is 3600 megabytes. We refer to queries in this class as \"why-lact\"", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "queries. In the case of why-fact queries, the explanation generator must examine the facts that were created during the fact gathering phase, and it must determine how those facts are related through the backward-chaining process. explanations. But because the user had earlier typed \"explain more\", the explanation generator went on to explain the terms \"total-user-disk-space\" and \"sum.of.system.disk-space\", which were introduced in the first paragraph. If the user were to type \"explain more\" a second time. and then ask the same question \"why quantity total-disk-space\", the explanation generator would not stop where it did. Instead, it would go on to explain the terms user-disk.space, percent.for-expansion, and system.disk-space, which were introduced in the second and third paragraphs,", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "There is no upper bound on the number of levels of explanation the user may request. If the number of levels to explain is high.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "XSEL will keep explaining until it reaches those facts whose values were set either by user input or by default, in which case there is nothing further to explain. The user can ~lso type \"explain less\" at any time, thus decreasing the number of levels to explain. The lower bound on the number of levels to explain is one.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "The mechanism for determining which term to explain next is a queue. As new terms are introduced they are placed in the queue. The queue was originally implemented as a stack, but as explanations got longer they began to sound less coherent using the stack mechanism. So the queue was implemented, but the stack was retained. Now one can toggle between them by typing \"explain queue\" or \"explain stack\", thus producing alternatively structured explanations for the sake of comparison.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "The third ol~vious class of queries asks why a certain quantity is needed for any line-item. We refer to these as \"why-line.item\" queries, Why-line-item queries require the most complicated processing because the explanation generator must understand how the line-item that was selected relates back to the facts that determine the quantity needed, and there is usually a long sequence of forward-chaining rules as well as the whole evaluation function mechananism between the creation of the facts and the creation of the line-items. the number of levels to explain was set at two. The first two paragrapl'~ comprise the first level, so tire explanation could have 23O stopped there; the remaining two paragraphs were generated in response to terms introduced in the first two paragraphs.", |
| "cite_spans": [], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "? why q ra60 Another common early approach to explanation generation was the goal tree approach, which is very.similar to the execution trace approach. The original explanations produced by the MYCIN system were goal tree explanations [1] . This approach allowed the user to question any request for information made by the system, and the system would simply locate the goal immediately above the current one in the goal tree and report that it needed the information to resolve that higher goal. Contrast the relational explanation for the query \"why q totaldisk-space\" shown in Figure 3 -1 with the generic explanation for the same query shown in Figure 1- ", |
| "cite_spans": [ |
| { |
| "start": 235, |
| "end": 238, |
| "text": "[1]", |
| "ref_id": "BIBREF0" |
| } |
| ], |
| "ref_spans": [ |
| { |
| "start": 581, |
| "end": 589, |
| "text": "Figure 3", |
| "ref_id": null |
| }, |
| { |
| "start": 650, |
| "end": 659, |
| "text": "Figure 1-", |
| "ref_id": "FIGREF3" |
| } |
| ], |
| "eq_spans": [], |
| "section": "Introduction", |
| "sec_num": "1." |
| }, |
| { |
| "text": "The knowledge structures just described, including mas=mge~ query.terms, the query-queue, schemas and links, serve as intermediate structures between the reasoning knowledge of the expert system and the linguistic knowledge needed for language generation .4 Some of the terminology used to describe these structures, e.g., \"reason\" and \"elaboration\" relations, is derived from the work of Mann [7] and Hobbs [3] inference rules from relations. We realize that this could be a very long term goal, but it aJse has the short term benefit of providing useful explanations.", |
| "cite_spans": [ |
| { |
| "start": 394, |
| "end": 397, |
| "text": "[7]", |
| "ref_id": "BIBREF6" |
| }, |
| { |
| "start": 408, |
| "end": 411, |
| "text": "[3]", |
| "ref_id": "BIBREF2" |
| } |
| ], |
| "ref_spans": [], |
| "eq_spans": [], |
| "section": "Rationale", |
| "sec_num": "4." |
| } |
| ], |
| "back_matter": [], |
| "bib_entries": { |
| "BIBREF0": { |
| "ref_id": "b0", |
| "title": "Applications of meta level knowledge to the construction, maintenance, and use of large knowledge bases", |
| "authors": [ |
| { |
| "first": "R", |
| "middle": [], |
| "last": "Davis", |
| "suffix": "" |
| } |
| ], |
| "year": 1976, |
| "venue": "Stanford Artificial Intelligence L~oratory Memo", |
| "volume": "283", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "R. Davis. Applications of meta level knowledge to the construction, maintenance, and use of large knowledge bases. Ph.D. Th., Stanford University, 1976. Stanford Artificial Intelligence L~oratory Memo 283, Stanford, CA.", |
| "links": null |
| }, |
| "BIBREF1": { |
| "ref_id": "b1", |
| "title": "User's Manual", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "Forgy", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Ops", |
| "suffix": "" |
| } |
| ], |
| "year": 1981, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "C. L Forgy. OPS.5 User's Manual. CMU.CS-81-135, Dept of Computer Science, Carnegie-Mellon University, Pittsburgh, PA 15213, July 1981.", |
| "links": null |
| }, |
| "BIBREF2": { |
| "ref_id": "b2", |
| "title": "Towards an Understanding of Coherence in Discourse", |
| "authors": [ |
| { |
| "first": "Jerry", |
| "middle": [ |
| "R" |
| ], |
| "last": "Hobbs", |
| "suffix": "" |
| } |
| ], |
| "year": 1982, |
| "venue": "Strategies for Natural Language Processing", |
| "volume": "3", |
| "issue": "", |
| "pages": "223--247", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Jerry R. Hobbs. Towards an Understanding of Coherence in Discourse. In W. G. Lehnert and M. H. Ringle, Ed., Strategies for Natural Language Processing, Lawrence Erlbaum Associates, New Jersey, 1982, pp. 223-24,3.", |
| "links": null |
| }, |
| "BIBREF3": { |
| "ref_id": "b3", |
| "title": "A Foundation for Knowledge Acquisition", |
| "authors": [ |
| { |
| "first": "Gary", |
| "middle": [], |
| "last": "Kahn", |
| "suffix": "" |
| }, |
| { |
| "first": "Steve", |
| "middle": [], |
| "last": "Now/An", |
| "suffix": "" |
| }, |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Mcdermott", |
| "suffix": "" |
| } |
| ], |
| "year": 1984, |
| "venue": "Proceedings of the IEEE Workshop on Principles of Knowledge.Based Systems", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Gary Kahn, Steve Now/an, and John McDermott. A Foundation for Knowledge Acquisition. Proceedings of the IEEE Workshop on Principles of Knowledge.Based Systems, IEEE, Denver, CO, 1984, pp..", |
| "links": null |
| }, |
| "BIBREF4": { |
| "ref_id": "b4", |
| "title": "MEX: An OPS-based approach to explanation", |
| "authors": [ |
| { |
| "first": "Gary", |
| "middle": [], |
| "last": "Kahn", |
| "suffix": "" |
| }, |
| { |
| "first": "David", |
| "middle": [], |
| "last": "Gelier", |
| "suffix": "" |
| } |
| ], |
| "year": 1984, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Gary Kahn and David Gelier. MEX: An OPS-based approach to explanation. 1984.", |
| "links": null |
| }, |
| "BIBREF5": { |
| "ref_id": "b5", |
| "title": "XSEL as Knowledge Acquirer and Explainer", |
| "authors": [ |
| { |
| "first": "Karan", |
| "middle": [], |
| "last": "Kukich", |
| "suffix": "" |
| }, |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Mcdermott", |
| "suffix": "" |
| }, |
| { |
| "first": "Tianran", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 1985, |
| "venue": "", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Karan Kukich, John McDermott and Tianran Wang. XSEL as Knowledge Acquirer and Explainer. 1985.", |
| "links": null |
| }, |
| "BIBREF6": { |
| "ref_id": "b6", |
| "title": "Relational Propositions in Discourse", |
| "authors": [ |
| { |
| "first": "C", |
| "middle": [], |
| "last": "William", |
| "suffix": "" |
| }, |
| { |
| "first": "Sandra", |
| "middle": [ |
| "A" |
| ], |
| "last": "Mann", |
| "suffix": "" |
| }, |
| { |
| "first": "", |
| "middle": [], |
| "last": "Thompson", |
| "suffix": "" |
| } |
| ], |
| "year": null, |
| "venue": "", |
| "volume": "198", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "William C. Mann and Sandra A. Thompson. Relational Propositions in Discourse. 198,3.", |
| "links": null |
| }, |
| "BIBREF7": { |
| "ref_id": "b7", |
| "title": "A Knowledge Acquisition System for VT", |
| "authors": [ |
| { |
| "first": "Sandra", |
| "middle": [ |
| "L" |
| ], |
| "last": "Marcus", |
| "suffix": "" |
| }, |
| { |
| "first": "John", |
| "middle": [], |
| "last": "Mcdermott", |
| "suffix": "" |
| }, |
| { |
| "first": "Tianran", |
| "middle": [], |
| "last": "Wang", |
| "suffix": "" |
| } |
| ], |
| "year": 1985, |
| "venue": "Proceedings of the AAAI, AAAI", |
| "volume": "", |
| "issue": "", |
| "pages": "", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Sandra L. Marcus, John McDermott and Tianran Wang. A Knowledge Acquisition System for VT. Proceedings of the AAAI, AAAI, Los Angeles, CA, 1985, pp..", |
| "links": null |
| }, |
| "BIBREF8": { |
| "ref_id": "b8", |
| "title": "Semantic Rule Based Text Generation", |
| "authors": [ |
| { |
| "first": "Michael", |
| "middle": [], |
| "last": "Mauldin", |
| "suffix": "" |
| } |
| ], |
| "year": 1984, |
| "venue": "Proceedings of the lOth International Conference on Compu=ational Linguistic~ ACL", |
| "volume": "", |
| "issue": "", |
| "pages": "376--380", |
| "other_ids": {}, |
| "num": null, |
| "urls": [], |
| "raw_text": "Michael Mauldin. Semantic Rule Based Text Generation. Proceedings of the lOth International Conference on Compu=ational Linguistic~ ACL, Stanford University, Stanford, CA, 2-6 July 1984, pp. 376-380.", |
| "links": null |
| } |
| }, |
| "ref_entries": { |
| "FIGREF0": { |
| "num": null, |
| "uris": null, |
| "text": "Sample XSEL Rank", |
| "type_str": "figure" |
| }, |
| "FIGREF1": { |
| "num": null, |
| "uris": null, |
| "text": "Sample Why-Choice Explanations", |
| "type_str": "figure" |
| }, |
| "FIGREF2": { |
| "num": null, |
| "uris": null, |
| "text": "Sample Why-Fact Explanation This explanation would have ended immediately following the first paragraph had not the user previously asked for longer", |
| "type_str": "figure" |
| }, |
| "FIGREF3": { |
| "num": null, |
| "uris": null, |
| "text": "5 shows a sample explanation from the why-line-item class. In this example.", |
| "type_str": "figure" |
| }, |
| "FIGREF4": { |
| "num": null, |
| "uris": null, |
| "text": "Sample Why-Line.Item Explanation", |
| "type_str": "figure" |
| }, |
| "FIGREF5": { |
| "num": null, |
| "uris": null, |
| "text": "Generic vs. Relational Explanations Both of the link-dependent message generators are capable of operating in two modes, generic mode and relational mode. (The user can toggle between modes by typing \"explain generic\" or \"explain relational\".) The explanations shown above in Figures 1-4 and 1-5 are relational explanations: they explicate the relations that hold between facts. Some of those relations are arithmetic relations, such as sum and product, and some are abstract relations, such as satisfaction and allocation relations.", |
| "type_str": "figure" |
| }, |
| "FIGREF6": { |
| "num": null, |
| "uris": null, |
| "text": "generic link is a very sJmple memory element consisting of only two attributes, a source attribute and a sink attribute. The value of the source attribute is the token (i.e., unique identifier) of some fact that entered into the inference of the resultant fact; the value of the sink attribute is the token of the resultant fact. For example, the rules that fire to infer a value for the fact total-disk-233 space will deposit into working memory at lea.st five generic links, each having the token of the fact total-disk-space in its sink attribute and each having the token of a fact that entered into the calculation of the value for total-disk-space, such aS totalapplication-disk-space, programmer-disk-space, etc., in its source attribute. An example of a generic link is shown inFigure 3-2. A relational link is a sJightly richer memory element which not only names the relation that holds between two or more facts, but also categorizes it.Figure 3-3 displays one arithmetic relational link and one abstract relation link. (generic.link tsource <total-application-disk-space-token> tsink <total-disk.space-token> ) Sample Generic Link (relational-link trelation sum tcategory arithmetic tsmk <total-disk-space-token> tsourcet <total-user-disk-space-token> tsource2 <sum-of-System-disk-space-token> tSOurce3 <sum-of-page-and-swap-space-token> ) (relational-link ~retation satisfaction \u00a2category reason tsink <quantity-of-disks-token> tsource <total-disk-space-token> ) Sample Arithmetic and Abstract Relational Links The network formed by relational links is in some ;)laces more dense and in other ;)laces less dense than the network formed by genenc links; arithmetic relational links create more levels thus making the relaUonal network denser, while abstract links tend to bridge long chains of facts, thus making the network sparser. To see this distinction, consider the arithmetic formula used by XSEL to calculate the total-disk-space requirement: total-disk.space = ( (total. application -disk. space + programmer-disk-space \u00f7 total-data-file-disk-space) * 125%) + sum of system.disk.space + sum of page-and.swap-space The rules that execute this formula create at least five generic links linking total-disk.space to total-application-disk-space, programmer-disk-space, total-data-file-disk-space, one or more system-disk-sp,3ce facts, and one or more page-and-swap-space facts. At the same time they create one relational link linking total-disk-space to three new intermediate level facts, total-userdisk-space, sum.of-system-disk-space, and sum-of-page-andswap.space, and they create additional relational links linking each of the intermediate facts to their subfacts. Total.user-diskspace is a newly created intermediate fact, and a relational link, with rrelation percent, is created linking it to two more new intermediate facts, user-disk-space and percent.for-expansion. Another relational link is in turn created linking user-disk-space to the three facts total-application-disk-space, programmer-diskspace, and total-data-file-disk-space. On the other hand, the rules that determine how many RA60 disk drives are needed, for example, create a dense generic network linking all the facts that enter into the calculation of totaldisk-space to the facts that allocate some portion of that amount to fixed-disk-space. From there the network would get even denser as fixed-disk-space is linked tO the fixed.disk.unit. capabihty and quantity-of-fixed-disks facts for each candidate. In fact, these generic links are not currently created due to limitations of working memory space. In contrast to the potentially dense generic network, the relational network contains only a few abstract relation links, such as satisfaction and allocation links, that bridge many of the generic links, thus resulting in a sparser network (and in more direct explanations). There are good reasons for the existence of two complete networks. Essentially, the tradeoff is that while generic links are trivial tO create, they do not facilitate satisfying explanations. On the other hand, the creation of relatil)nal links often requires manual intervention, lout relational links facilitate direct explanations. Compare again the generic explanation inFigure 3-I to its corresponding relational explanation inFigure 1.4.", |
| "type_str": "figure" |
| }, |
| "FIGREF7": { |
| "num": null, |
| "uris": null, |
| "text": "The Generic Schema directs the message generator to create additional query.terms for all the facts that are linked to the current query-term. The second directive directs the message generator to create one message with the predicate \"IS-DEPENDENT\" and with the focus-token of term1, which is the current query.term. The surface realization of this message will be the clause \"THE VALUE OF 3600 MEGABYTES FOR TOTAL-DISK-SPACE IS DEPENDENT \". The third directive of the generic schema directs the message generator to create one additional message with the predicate \"ON\" and the focus.token of terror for each of the link terms found. These messages will emerge as prepositionalphrases in their surface form, such as \" ON 1424 KILOBYTES FOR TOTAL-APPLICATION.DISK.SPACE, 110592 KILOBYTES 3'The term so/letup wls adOl~ed fRmt ~e ~ of McKeown(11), ~ simdet smBclu=~s f~ discou~e o\u00a2~=anizatlo~. FOR PROGRAMMER.DISK.SPACE , 2816000 KILOBYTES FOR TOTAL-DATA.FILE.DISK.SPACE , 600 KILOBYTES FOR PAGE. AND-SWAP-SPACE AND 25600 KILOBYTES FOR SYSTEM-DISK-SPACE .\" When the message generator is in relational mode, it is constrained to locating relational links and using relational schemas. There are a variety of each. Currently, relational links are categorized as being either reasons, elaborations, or arithmetic links. During step 2 of the message-generation cycle, the message generator searches first for reason links, next for elaboration links, and finally for arithmetic links. In some cases, the search for arithmetic links may be suppressed. For example, some links whose relation is allocation are subcategorized as being arithmetic operations, as in \"75 percent of the total.disk. space requirement was allocated to removable-pack disks\". In these cases, expressing the arithmetic relation also would be redundant. When a relational link is located, a corresponding schema is selected. In contrast to the single generic schema, there are a variety of arithmetic and abstract relational ~chemas. Figure 3-5 illustrates the arithmetic \"plus\" schema that was used to generate the messages for the first paragraph of the \"why quantity totaJ-disk-space\" relational explanation shown in Figure 1-4. It contains five directives, one to create the new query-terms found in the arithmetic reasoning trace and four to create messages. The second message creation directive will create as many messages as are needed to account for at least 80 percent of the total value of the fact being explained. (The 80 percent factor was implemented in order to filter out insignificant facts, thus making the explanation more concise. Another process that contributes to more readable explanations is the conversion of all units in different clauses of the explanation to the same highest common denominator, eg. megabytes.) Following that, two additional messages will be created, one to mention that the remainder of the total is accounted for by other terms, and another to give an example.", |
| "type_str": "figure" |
| }, |
| "FIGREF8": { |
| "num": null, |
| "uris": null, |
| "text": "6 illustrates the \"setisfactJon\" schema that was u~=d (Schema-directives:plus-schema (make goal tgoal-name create-extra-query.terms ~status reiterate) (make goal tgoal-name create-message tfocus-token <token I > tpredicate CAPACITY.REQUIREMENT tsubname RECOMMENDED) (make goal tgoal.name create-messege tfocus-token new ?predicate CAPACITY-REQUIREMENT ~ubname GENERAL tamount 80) (make goal tgoal-name create-message tpredicate REMAINDER) (make goal tgoal.name Create-message tfocus.token new tpredicate EXAMPLE) Sample Arithmetic Schemato create the massages for the first sentence of the \"why quantity RA60\" explanation shown inFigure 1-5. It contains one directive to create an extra query-term matching the token of the new term identified in the \"satisfaction\" link, and three actions making the three messages which surface as three clauses of text in the explanation.", |
| "type_str": "figure" |
| }, |
| "FIGREF9": { |
| "num": null, |
| "uris": null, |
| "text": "Sample Satisfaction Schema natural language text. One of the premises of this work on explanation generation is that the relations, or links, that are embodied in the inference rules of a successful reasoning system are the same ones that give coherence to natural language explanations. An immediate goal of this research is to identity those relations. At the present time only twenty.six different reasoning relations, have been identified in XSEL. As more types of reasoning relations are identified and corresponding links are added to XSEL's rules, more of XSEL's reasoning will be explainable. A long term goal of this work is to continue to identify and add reasoning links and schemas until we see some generalities begin to emerge. Perhaps some domainindependent set of reasoning relations and schemas might be found. Furthermore. such relations and schemas might facilitate the design of a knowledge acquisition system that would elicit knowledge from an expert, represent it as relations, and generate", |
| "type_str": "figure" |
| }, |
| "TABREF0": { |
| "html": null, |
| "type_str": "table", |
| "text": "1.1. Overview of XSELBriefly, the function of the XSEL system is to assist a salesperson in configuring a custom-tailored purchase order for a Digital Equipment Corporation VAX computer system. XSEL works with the salesperson tO elicit the functional computing requirements of the individual customer, and then goes on to select the components that best fit those requirements. The output of an XSEL session is a purchase order consisting of a list of line-items that specify hardware and software components.There ~re two main phases to XSEL's processincj, a fact gathering phase and a component select=on phase. During the fact gathering phase XSEL carries on an interactive dialog with the salesperson to elicit values for facts that determine the customer's functional computing requirements. These might include requirements for total disk space, percent of removable disk storage, number of terminals, lines-per.minute of printing, etc. Natural language processing during the fact gathering dialog is minimal: XSEL displays menues and pre-formutated queries and accepts one-or two-word answers from the user. Once enough facts have been collected XSEL begins a silent phase of processing. During this phase a set of candidate components that satisfy the customer's basic requirements is retrieved from the DEC parts database. Within each class of component, i.e., processor, disk, terminal, etc., candidates are ranked according to their score on a~q evaluation function that measures the degree to which a candidate satisfies the customer's weighted functional requirements. The candidate with the highest score is selected and placed on the purchase order.The fact collection process is driven by backward-chaining rules. A top-level rule deposits a few \"core\" facts for which XSEL must obtain values, such as \"total.disk-space\", \"total-number.of-", |
| "num": null, |
| "content": "<table><tr><td>terminals\", etc. One at a time, XSEL solicits a value for these</td></tr><tr><td>core facts from the salesperson. If the salesperson answers</td></tr><tr><td>\"unknown\" to a solicitation, another rule fires to deposit some</td></tr><tr><td>additional facts that would enable XSEL to infer a value for the</td></tr><tr><td>unknown fact. The cycle is then repeated as XSEL solicits values</td></tr><tr><td>for each of the newly deposited facts. Any time a newly</td></tr><tr><td>instantiated fact completes the set of facts required to infer a</td></tr><tr><td>value for some other fact. the appropriate inference rule is</td></tr><tr><td>The most important knowledge structure used by XSEL during</td></tr><tr><td>the fact gathering I~ase is a fact. A fact is simply a list of</td></tr><tr><td>attribute-value pairs that represent knowledge about one of the</td></tr><tr><td>customer's functional computing requirements. Figure 1-1</td></tr><tr><td>depicts a sample facL</td></tr><tr><td>(FACT ?ATTRIBUTE TOTAL.DISK-SPACE</td></tr><tr><td>?STATUS INFERENCE TCLASS DISK</td></tr><tr><td>?UNITS MEGAB~'TE3 ?MEAN 3600</td></tr><tr><td>YTOKEN G'.29)</td></tr><tr><td>Figure 1.1: Sample XSEL Fact</td></tr></table>" |
| }, |
| "TABREF1": { |
| "html": null, |
| "type_str": "table", |
| "text": "After aJl the rank values have been computed for a candidate they are summed to obtain a total score for the candidate. The candidate with the highest total score is selected and placed on the purchase order.", |
| "num": null, |
| "content": "<table><tr><td>? why ra81</td><td/></tr><tr><td>THE RA81 IS CHEAPER THAN ANY</td><td/></tr><tr><td>ALTERNATIVE RXED PACK DISK,</td><td/></tr><tr><td>POSSIBLY BECAUSE IT HAS A SMALLER</td><td/></tr><tr><td>TOTAL STORAGE CAPACITY AND A</td><td/></tr><tr><td>SLOWER AVERAGE-ACCESS-TIME.</td><td colspan=\"2\">The component selection phase is driven by forward.chaining</td></tr><tr><td>?why rm05</td><td colspan=\"2\">rules. These rules perform the subtasks of first, retrieving</td></tr><tr><td/><td colspan=\"2\">candidates from the database, next, determining a quantity and</td></tr><tr><td>ALTHOUGH THERE ARE LESS EXPENSIVE DISK S, THE RM05 HAS A LARGER</td><td colspan=\"2\">cost for each of the candidates, next, computing a total rank</td></tr><tr><td>DISK PACK THAN ANY ALTERNATIVE</td><td colspan=\"2\">score for each candidate, and finally, selecting the candidate with</td></tr><tr><td>REMOVABLE PACK DISK.</td><td>the highest rank score.</td></tr><tr><td/><td colspan=\"2\">At present, the entire XSEL system consists of over three</td></tr><tr><td/><td colspan=\"2\">thousand OPS5 [2] rules. The explanation generator, which will</td></tr><tr><td/><td colspan=\"2\">be described shortly, comprises an additional five hundred rules.</td></tr><tr><td/><td colspan=\"2\">Anywhere from approximately five hundred to five thousand rules</td></tr><tr><td/><td colspan=\"2\">may fire during the fact gathering phase to create from fifty to five</td></tr><tr><td/><td colspan=\"2\">hundred facts, and roughly three thousand rules will fire during</td></tr><tr><td/><td colspan=\"2\">the component selection phase to create around one thousand</td></tr><tr><td/><td colspan=\"2\">rank elements. The whole process can take anywhere from ten to</td></tr><tr><td/><td colspan=\"2\">thirty minutes of real time, depending on how XSEL's queries are</td></tr><tr><td/><td>answered.</td></tr><tr><td/><td>t .2. Sample Explanations</td></tr><tr><td/><td colspan=\"2\">Three of the most obvious types of queries a user m~ght ask</td></tr><tr><td/><td>were targeted for initial explanation development.</td><td>Sample</td></tr><tr><td/><td colspan=\"2\">explanations from each of those types are given in this section.</td></tr><tr><td/><td colspan=\"2\">The following sections describe the knowledge structures and</td></tr><tr><td/><td colspan=\"2\">processes within both XSEL and the explanation generator that</td></tr><tr><td/><td colspan=\"2\">produced those explanations, as well as the goals and rationale</td></tr><tr><td/><td>behind them.</td></tr><tr><td/><td colspan=\"2\">One type of query that is likely to be asked is why a particular</td></tr><tr><td/><td colspan=\"2\">component appears on a purchase order. We refer to queries of</td></tr><tr><td/><td colspan=\"2\">this type as \"why-choice\" queries. To answer a why-choice</td></tr><tr><td/><td colspan=\"2\">query the explanation generator must compare the rank elements</td></tr><tr><td/><td colspan=\"2\">for each candidate on each term of the evaluation function in</td></tr><tr><td/><td colspan=\"2\">order to determine which attributes were responsible for the</td></tr><tr><td/><td colspan=\"2\">higher SCore of the component that was actually selected. The</td></tr><tr><td/><td colspan=\"2\">following are sample explanations from the why-choice class of</td></tr><tr><td/><td>queries.</td></tr></table>" |
| }, |
| "TABREF4": { |
| "html": null, |
| "type_str": "table", |
| "text": "desi(jn of the XSEL explanation generator was motivated by three goals: first, that explanations should be accurate. second, that explanations should be direct, and third, that some degree of generality should be attempted.Most early attempts at explanation generation adopted either a canned text or an execution trace approach. The canned text approach led to accuracy problems and the execution trace approach led to directness problems. These problems aredescribed in detail by Swartout[12]. In brief, canned explanations can suffer from a lack of accuracy in the event that any modifications or additions are made to the Performance program without the corresl0onding modifications or additions being made to the canned text. Execution trace.explanations tend to suffer from a lack of directness because every step during program execution gets reported, including what Swartout has referred to as \"computer artifacts\", as in \"Variable X was initialized to 0\".", |
| "num": null, |
| "content": "<table/>" |
| }, |
| "TABREF5": { |
| "html": null, |
| "type_str": "table", |
| "text": "Goal tree explanations tend to suffer from the same lack of directness problems that execution trace explanations suffer from.Swartout's work on an explanation generator for the DigitalisTherapy Advisor attacked the accuracy and directness problems successfully. His approach was to redesign the DTA, separating descriptive facts from domain principles and from the abstract goals of the system. This allowed the performance program to be generated by an automatic programmer, which also created a goal refinement structure in the process. The goal refinement structure captures the knowledge that goes into writing the performance program, and makes it accessible to the explanation generator, where it can be used to produce explanations that are both accurate and direct. Furthermore, as Swartout points out, such explanations can be viewed as \"justifications\" for the system's behavior.One of the major contributions of the DTA work was to demonstrate that a singte explicit representation of knowledge can and should drive both the automatic program generation process and the explanation generation process. approaches discussed above. The sheer size of XSEL would make implementing canned responses tedious. Similarly, the number of rule firings on any run would make reading execution trace explanations labonous even. or perhaps especially, if they were translated into natural lanaguage. The approach taken by Swartout of extracting the regularities and representing them separately as domain principles would work for the backwardchaining rules used during XSEL's fact gathering phase, but the forward-chaining rules used during the component selection phase are so irregular that attempting to extract regularities would result in the duplication of nearly the entire set of rules.Explanations produced by both the comparative message generator and the link-dependent message generators are certain to be accurate because they always originate from the contenfs of working memory. Special steps had to be taken to ensure the directness of the link-dependent message generators.however. Those steps will be discussed in the following sections.", |
| "num": null, |
| "content": "<table><tr><td>which describe the workings of the lipk-dependent message</td></tr><tr><td>generators in some detail. Discussion of the comparative</td></tr><tr><td>message generator and the surface generator will be reserved for</td></tr><tr><td>other occasions.</td></tr><tr><td>Further</td></tr><tr><td>research supporting the \"shared explicit knowledge\" approach</td></tr><tr><td>to automatic knowledge acquisition, rule generation, and</td></tr><tr><td>explanation generation is underway for at least three other</td></tr><tr><td>projects [8] [4] [5] [6].</td></tr></table>" |
| }, |
| "TABREF7": { |
| "html": null, |
| "type_str": "table", |
| "text": "Generic links require little effort to create because they simply incorporate the tokens of the facts that are used in an inference 234 rule. In fact, an automatic rule generator was developed for automatically creating most of XSEL's backward.chaining factgathering rules from simple arithmetic formulas such as the formula for total-disk-spsce discussed above.lit was a trivial task to have the automatic rule generator include the actions required to have the inference rules create the generic links.", |
| "num": null, |
| "content": "<table><tr><td colspan=\"2\">additional query-terms and messages suggested by the selected</td></tr><tr><td colspan=\"2\">schema, and 5) turn control over to the surface generator. Each</td></tr><tr><td colspan=\"2\">time a new query-term is created, queue-control rules decide</td></tr><tr><td colspan=\"2\">whether to place it in the query-queue, depending on such</td></tr><tr><td colspan=\"2\">factors as whether the term has already been explained and how</td></tr><tr><td colspan=\"2\">many levels of explanation the user has requested. As long as</td></tr><tr><td colspan=\"2\">The task of augmenting the fact-gathering rules to create the query-queue is not empty, the message generation cycle is</td></tr><tr><td>reiterated.</td><td>arithmetic relational links was also automatable, for the most part.</td></tr><tr><td/><td>An automatic link-creator was written to parse the arithmetic</td></tr><tr><td colspan=\"2\">formulas that were input to the rule generator and create the When the message generator is in generic mode, it b</td></tr><tr><td colspan=\"2\">appropriate links. This parser identified the main arithmetic constrained to locating generic links during step 2 of the cycle,</td></tr><tr><td colspan=\"2\">operations, created names for intermediate facts, and modified and it is constrained to selecting the generic schema during step</td></tr><tr><td colspan=\"2\">XSEL's rules to have them create the arithmetic relational links. 3 of the cycle. A simplified version of the generic schema is</td></tr><tr><td colspan=\"2\">The output of the automatic link-creator required only minor depicted in Figure 3.4. The first directive of the generic schema</td></tr><tr><td>(Schema-directives::Generic-schema</td><td>manual retouching in those cases where its heuristics for</td></tr><tr><td>(make goal tgoal-name create.extra.query.terms</td><td>creating names for intermediate facts fell short. 2 But the task of</td></tr><tr><td>tstatus reiterate) (make goal Tgoal-name create-message</td><td>augmenting the component selection rules to create the abstract</td></tr><tr><td>tgredicate IS-DEPENDENT</td><td>relational links between facts has so far resisted an automatic</td></tr><tr><td>~erml <current-focus>) (make goal rgoal-narne create-message</td><td>solution. These links are now being added manually. They</td></tr><tr><td>?predicate ON</td><td>require the effort of someone who understands the workings of</td></tr><tr><td>~terml <link.focus> tstatus reiterate)</td><td>XSEL and recognizes what explanations might be called for and.</td></tr><tr><td/><td>consequently, which rules should be modified to create relational</td></tr><tr><td/><td>links.</td></tr><tr><td/><td>3.2. Overview of Processing</td></tr><tr><td/><td>The processing of a query by a link-dependent message</td></tr><tr><td/><td>generator goes as follows. When the initial query is input, a</td></tr><tr><td/><td>query-interpretation context is entered. In this context some</td></tr><tr><td/><td>rules fire tO identify and locate the fact in question, to create a</td></tr><tr><td/><td>query-term with the same token as the fact. and to place that</td></tr><tr><td/><td>query-term in the query-queue. Following query-interpretation, a</td></tr><tr><td/><td>message generation cycle consisting roughly of the following five</td></tr><tr><td/><td>steps reiterates: 1) focus on the next query-term in the queue, 2)</td></tr><tr><td/><td>locate the links related to that query-term, 3) select an</td></tr><tr><td/><td>explanation schema 3 based on the links found, 4) create</td></tr><tr><td/><td>1XSEL's automatic ride gammer was v~ten by Samly Marcus.</td></tr><tr><td/><td>2XSEL's auSommic link-creatm ~S vmtmen by kTr.ttaet ~w~</td></tr></table>" |
| }, |
| "TABREF8": { |
| "html": null, |
| "type_str": "table", |
| "text": "on discourse organization. Mann and Hobbs independently postulate that discourse relations, such as reason and elaboration relations among others, are rasDonsible for coherence in well-organized (Schema.directives:satisfy-schema (make goal tgoal-name create-extra.query-term", |
| "num": null, |
| "content": "<table><tr><td>Hocus-token <term2>)</td></tr><tr><td>(make goal tgoal.narne create-message</td></tr><tr><td>?predicate QUANTITY.SELECTED</td></tr><tr><td>tterml <term1>)\"</td></tr><tr><td>(make goal ?goal.name create-message</td></tr><tr><td>tpredicate INORDER</td></tr><tr><td>1\"retype relational-prop)</td></tr><tr><td>(make goal tgoal-narne create-message</td></tr><tr><td>?predicate CAPACITY.REQUIREMENT</td></tr><tr><td>tsubncme SATISFY</td></tr><tr><td>tterm2 <term?.>)</td></tr></table>" |
| } |
| } |
| } |
| } |
