Learning Micro-Planning Rules for Preventative Expressions Keith Vander Lindeny Information Technology Research Institute University of Brighton email:
[email protected]
Barbara Di Eugenio Computational Linguistics Carnegie Mellon University email:
[email protected]
8 May 1996
Abstract
corpus analysis on a representative collection of the text produced by human authors in that domain and to induce a set of micro-planning rules guiding the generation process in accordance with the results. Some fairly simple rules usually jump out of the analysis quickly, mostly based on the analyst’s intuitions. For example, in written instructions, user actions are typically expressed as imperatives. Such observations, however, tend to be gross characterisations. More accurate microplanning requires painstaking analysis. In this paper, for example, the micro-planner must distinguish between phrasing such as “Don’t do action X” and “Take care not to do action X”. At this level of detail, it is far from clear how the decision can best be made. Some form of automation would clearly be desirable. Unfortunately, corpus analysis techniques are not yet capable of automating the initial phases of the corpus study (nor will they be for the foreseeable future). There are, however, techniques for rule induction which are useful for the later stages of corpus analysis and for implementation. In this paper, we focus on the use of such rule induction techniques in the context of the microplanning of preventative expressions in instructional text. We define what we mean by a preventative expression, and go on to describe a corpus analysis in which we derive three features that predict the grammatical form of such expressions. We then use the C4.5 learning algorithm to construct a micro-planning sub-network appropriate for these expressions. We conclude with an implemented example in which the author is allowed to set the three relevant features, and the system generates the appropriate expressions in English
Building text planning resources by hand is timeconsuming and difficult. Certainly, a number of planning architectures and their accompanying plan libraries have been implemented, but while the architectures themselves may be reused in a new domain, the library of plans typically cannot. One way to address this problem is to use machine learning techniques to automate the derivation of planning resources for new domains. In this paper, we apply this technique to build microplanning rules for preventative expressions in instructional text.
1 Introduction Building text planning resources by hand is timeconsuming and difficult. Certainly, much work has been done in this regard; there are a number of freely available text planning architectures, but it is frequently the case that while the architecture itself can be reused in a new domain, the library of text plans developed for it cannot. In particular, micro-planning rules, those rules that specify the low-level grammatical details of expression, are highly sensitive to variations between sublanguages, and are therefore difficult to reuse. When faced with a new domain in which to generate text, the typical scenario is to perform a This work is partially supported by the Engineering and
Physical Sciences Research Council (EPSRC) Grant J19221, by BC/DAAD ARC Project 293, and by the Commission of the European Union Grant LRE-62009. y After September 1, Dr. Vander Linden’s address will be Department of Mathematics and Computer Science, Calvin College, Grand Rapids, MI 49546, USA.
1
and in French.
communication, and the features of the surrounding text being generated. In this section we will start with a discussion of our corpus, and then de2 Preventative Expressions tail the function and form features that we have coded. We will conclude with a discussion of the Preventative expressions are used to warn the intercoder reliability. reader not to perform certain inappropriate or potentially dangerous actions. The reader may be told, for example, “Do not enter” or “Take care 3.1 Corpus not to push too hard”. Both of these examples The raw instructional corpus from which we take involve negation (“do not” and “take care not”). all the examples we have coded has been collected Although this is not strictly necessary for preven- opportunistically off the internet and from other tative expressions (e.g., one might say “stay out” sources. It is approximately 4 MB in size and rather than “do not enter”), we will focus on the is made entirely of written English instructional use of negative forms in this paper, using the fol- texts. The corpus includes a collection of recipes, lowing categorisation: two complete do-it-yourself manuals RD, 1991;
McGowan and DuBern, 19911 , the Sun Opennegative imperatives proper (termed DONT windows on-line instructions, and a collection of imperatives) — These are characterised by administrative application forms. As a collection, the negative auxiliary do not or don’t, as in: these texts are the result of a variety of authors working in a variety of instructional contexts. The (1) Your sheet vinyl floor may be vinyl size of these portions of the corpus are shown in asbestos, which is no longer on the Table 2. market. Don’t sand it or tear it up We broke the corpus texts into expressions usbecause this will put dangerous ing a simple sentence breaking algorithm and then asbestos fibers into the air. collected the negative imperatives by probing for other negative imperatives (termed neg-TC expressions that contain the grammatical forms imperatives) — These include take care and we were interested in (e.g., expressions containbe careful followed by a negative infinitival ing phrases such as don’t and take care). The first complement, as in the following examples: row in Table 1 shows the frequency of occurrence for each of the grammatical forms we probed for. (2) To book the strip, fold the bottom third These grammatical forms, 1215 occurrences in or more of the strip over the middle of all, constitute 2.6% of the expressions in the full the panel, pasted sides together, taking corpus. We then filtered the results of this probe care not to crease the wallpaper in two ways: sharply at the fold. (3)
If your plans call for replacing the wood base molding with vinyl cove molding, be careful not to damage the walls as you remove the wood base.
1. When the probe returned more than 100 examples for a grammatical form, we randomly selected around 100 of those returned. We took all the examples for those forms that returned fewer than 100 examples. The number of examples that resulted is shown in row 2 of Table 1 (labelled “Raw Sample”).
3 Corpus Analysis In terms of text generation, our interest is in finding mappings from features related to the function of these expressions, to those related to their grammatical form. Functional features include the semantic features of the message being expressed, the pragmatic features of the context of
2. We removed those examples that, although they contained the desired lexical string, did not constitute negative imperatives. This pruning was done when the example was not 1
These do-it-yourself manuals were scanned by Joseph Rosenzweig.
2
Raw Grep Raw Sample Final Coding
DONT don’t/do not never 802 108 199 108 167 40 207
Neg-TC make sure be careful 229 52 104 52 3 46 72
take care 21 21 17
be sure 71 71 6
Table 1: Distribution of negative imperatives
3.3 Function Features
an imperative (e.g., “If you don’t see the Mail Tool window : : : ”) and when the example was not negative (e.g., “Make sure to lock the bit tightly in the collar.”). The number of examples which resulted is shown in row 3 of Table 1 (labelled “Final Coding”).
We will now discuss three of the function features we have coded: INTENTIONALITY, AWARENESS, and SAFETY. We illustrate them in turn using to refer to the prevented action and using “agent” to refer to the reader and executer of the instructions.
As shown in Table 1, the final corpus sample is made up of 279 examples, all of which have been coded for the features to be discussed in the next two sections. Table 2 also shows the number of examples from this sample that came from each type of instructional text.2
3.3.1
INTENTIONALITY
This feature encodes whether the agent consciously adopts the intention of performing . We settled on two values:
CON is used to code situations where the writer expects the agent to intend to perform . This often happens when the agent is aware Because of its syntactic nature, the form feature that is one of his or her possible alternacoding was very robust. The possible feature valtives, as in: ues were: DONT — for the do not and don’t forms discussed above; NEVER, for imperatives (4) Dust-mop or vacuum your parquet containing never; and neg-TC — for take care, floor as you would carpeting. Do not make sure, ensure, be careful, be sure, be certain scrub or wet-mop the parquet. expressions with negative arguments.3 The two coders agreed on their coding of this feature in all If this negative imperative were not included, cases. the agent might have intentionally chosen Note that this is only a preliminary categorito scrub or wet-mop because these actions sation. We believe that we have included the would result in deeper cleaning, and because most common preventative forms, but have not he or she was unaware of the bad conseattempted to quantify this. Furthermore, the work quences; reported here does not distinguish between the sub-forms for the main categories. For example, UNC is perhaps a less felicitous name because we certainly don’t mean that the agent may we do not address the subtle distinction between perform actions while being unconscious! do not and don’t (cf. Horn, 1989), nor have we Rather, we mean that the agent doesn’t redistinguished the various sub-forms of neg-TC, alize that there is a choice involved (cf. Di which are unlikely to be functionally equivalent. Eugenio, 1993, Section 4.2). It is used in 2 Note that there are number of examples from DiEugetwo situations: when is totally accidental, nio’s thesis which were included as excerpts, so we do not as in: know the full size of original corpus for that portion.
3.2 Form
3
We also code for a number of other syntactic features which are not relevant for this paper.
(5)
3
Be careful not to burn the garlic.
Instruction type Recipes Do-it-yourself Dieugenio’s thesis Software instructions Administrative forms Other Totals
Corpus size 1.7M 1.26M 52K 264K 317K 565K 4.2M
# of preventatives 83 99 69 0 9 19 279
Table 2: Distribution of examples from sample In the domain of cooking, no agent would are: consciously burn the garlic. Alternatively, an example is coded as UNC when has BADP is used when the agent’s safety is put at risk by performing , as in: to be intentionally planned for, but the agent may not take into account a crucial feature (7) Never mix cleaners containing acid or of , as in: ammonia with chlorine bleach. The (6) Don’t charge – or store – a tool where chemical reaction releases the chlorine the temperature is below 40 degrees F as a poisonous gas. or above 105 degrees. Here, performing the mixing action could While the agent will clearly intend to perlead to the agent’s death; form charging or storing a tool, he or she is likely to overlook (at least in the writer’s NOT-BADP is used when it is not unsafe to perform . Example 5 (“Be careful not to burn opinion) that temperature could have a negthe garlic”) is coded as NOT-BADP because ative impact on the results of such actions. the agent’s safety it not put at risk by burning the garlic. 3.3.2 AWARENESS This feature captures whether the agent is AWare 3.4 Intercoder reliability or UNAWare that the consequences of are bad. These features are detailed now: Each author independently coded each of the features for all the examples in the sample. The UNAW is used when the agent is perceived to be percentage agreement for each of the features is unaware that is bad. Example 6 (“Don’t shown in the following table: charge – or store – a tool where the temperature is below 40 degrees F or above 105 feature percent agreement degrees”) is coded as UNAW because it is intentionality 74.9% unlikely that the agent will know about this awareness 93.5% restriction; safety 90.7%
form 100% AW is used when the agent is aware that is bad. Example 5 (“Be careful not to burn the garlic”) is coded as AW because the agent is Until very recently, these values would most likely well aware that burning things when cooking have been accepted as a basis for further analysis. This has changed somewhat since the 1995 AAAI them is bad. workshop on empirical methods in discourse. Following Isard and Carletta’s suggestion 1995, we 3.3.3 SAFETY have looked into the content analysis literature, This feature captures whether the agent’s safety is and have agreed with their conclusions that the put at risk by performing . The possible values K coefficient suggested by Siegel and Castellan 4
1988 is a good measure of coder agreement. It not only measures agreement, but also factors out chance agreement. The K statistic is used for nominal (or categorical) scales, namely, scales in which numbers or other symbols are used to classify an object. In nominal scales, there is no relation between the different categories, and classification induces equivalence classes on the set of classified objects. In our coding scheme, the features are independent, so we report separate K values for each of the features coded. If P (A) is the proportion of times the coders (raters) agree, and P (E ) is the proportion of times that coders are expected to agree by chance, K is computed as follows:
which are more subjective in nature, engender more disagreement among coders, as shown by the K values in the following table: feature INTENTIONALITY AWARENESS SAFETY
K 0.46 0.76 0.71
According to this table, therefore, the AWARENESS and SAFETY features show “substantial” agreement and the INTENTIONALITY feature shows “moderate” agreement. We have coded other functional features as well, but they have either not proven as reliable as these, or are not as useful in text planning. P (A) ? P (E ) In addition, Siegel and Castellan 1988 point out K = 1 ? P (E ) that it is possible to check the significance of K when the number of objects is large; this involves Thus, if there is total agreement among the coders, computing the distribution of K itself. Under this K will be 1; if there is no agreement other than approach, the three values above are significant at what is expected to occur by chance, K will be the .000005 level. 0. There are various ways of computing P (E ); according to Siegel and Castellan, 1988, most researchers agree on the following formula, which 4 Automated Learning we also adopted: The corpus analysis results in a set of examples m coded with the values of the function and form P (E ) = p2j features. This data can be used to find correlaj =1 tions between the two types of features, correlawhere m is the number of categories, and pj is the tions, which, in text generation, are typically implemented as decision trees or rule-based systems proportion of objects assigned to category j . The mere fact that K may have a value k greater mapping from function features to forms. In this than zero is not sufficient to draw any conclusion, study, we used the 174 coded examples on which however, as it must be established whether k is the two coders agreed on all features as input to significantly different from zero. There are sug- the learning algorithm. gestions in the literature that allow us to draw general conclusions without these further compu- 4.1 Data Mining tations. For example, Rietveld and van Hout 1993 suggest the correlation between K values and in- As noted above, the correlations are difficult to tercoder reliability shown in the following table: divine and to implement. Work in machine learning, however, has addressed the question of how to find them, and in some cases of how to build Kappa Value Reliability Level decision networks or rules implementing them. .00 – .20 slight Quinlan’s C4.5 Quinlan, 1993 is such a system. .21 – .40 fair Given a set of discrete coded features designated .41 – .60 moderate as either input or output, it produces a decision .61 – .80 substantial tree or a set of rules. .81 – 1.00 almost perfect To aid in the construction of the decision For the form feature, the Kappa value is 1.0, indi- trees themselves, we have employed Clementine cating perfect agreement. The function features,
X
5
Figure 1: The Clementine learning environment
Figure 2: The micro-planner system network derived from the decision tree
Clementine, 1995, a data mining tool which allows rapid reconfiguration of various data manipulation facilities, including C4.5. Figure 1 shows the basic control stream we used for learning and testing decision trees. Data is input from the splitoutput file node on the left of the figure and is passed through filtering modules until it reaches the output modules on the right. The two select modules (children of the main input node) select the examples reserved for the training set and the testing set respectively. The upper stream processes the training set and contains a type module which marks the main syntactic form (i.e., DONT, NEVER, or Neg-TC) as the variable to be predicted and the AWARENESS, SAFETY, and INTENTIONALITY features as the inputs. Its output is passed to the C4.5 node, labelled mform, which produces the decision tree. We then use two copies of the resulting decision tree, represented by the diamond shaped nodes marked with mform, to test the accuracy of the testing and the training sets. One run of the system gave the following decision tree:
intention CON/SUB -> DONT intention UNC awareness AW -> NEG-TC awareness UNAW safety NOT -> DONT safety BADP -> NEVER This tree takes the three function features and predicts the DONT, NEVER, and Neg-TC forms. It matches our basic intuition that never imperatives are used when personal safety may be endangered (coded as safety=“BADP”), and that Neg-TC forms are used when the reader is expected to be aware of the danger that may arise. More features are required to distinguish the various Neg-TC forms. This decision tree accurately predicted the grammatical form of 72.5% of the 142 training examples, and 84.4% of the 32 testing examples. It is somewhat unusual, although not impossible, for the network to perform better on the testing set than on the training set as it did in this case.
6
4.2 Integration
builds decision trees. This gives rise to three distinctly non-systemic features of these learned Because it is common for us to rebuild this denetworks: cision tree frequently during analysis, we implemented a routine which automatically converts 1. The realisation statements are included only the decision tree into the appropriate KPML-style at the leaf nodes of the network. We have system networks with their associated choosers, built no intelligent facility for decomposing inquiries, and inquiry implementations Bateman, the realisation statements and filtering com1995. This makes the network compatible with mon realisations up the tree. the DRAFTER micro-planner, a decendent of IM2. The learning algorithm will freely reuse sysAGENE (Vander Linden and Martin, 1995). The tems (i.e., features) as various points in the conversion routine takes the following inputs: tree. This did not happen in Figure 2, but occasionally one of the features is indepen the applicable language(s) — C4.5 prodently used in different sub-trees of the netduces its decision trees based on exwork. We are forced, therefore, to index the amples from a particular language, and system and feature names with integers to KPML is capable of being conditiondisambiguate. alised for particular languages. Thus, we may perform separate corpus anal3. There is no meta-functional distinction in the yses of a particular phenomenon for network, but rather, all the features, regardvarious languages, and learn separate less of their semantic type, are included in micro-planning trees; the same tree. the input feature(s) — The sub-network being built must fit into the overall catThe sub-network derived in this section was egorisations of the full micro-planner, spliced into the existing micro-planning network and thus we must specify the text funcfor the full generation system. As mentioned tions that would trigger entry to the new above, this integration was done by manually sub-network (we’ll discuss integration specifying the desired input feature for the sublater in this section); network when the micro-planning rules are built. the decision tree itself; For the preventative expression sub-network, this a feature-value function — To traverse turned out to be a relatively simple matter. The the new sub-network, the KPML inmodel of procedural relations used by the system quiries require a function that can deterincludes a warning relation which may be attached mine the value of the features for each by the user when appropriate. The micro-planner, pass through the network; therefore, is able to identify those portions of the grammatical form specifications — The procedure which are to be treated as warnings, sub-network must eventually build senand to enter the derived sub-network appropritence plan language (SPL) commands ately. This same process could be done with any for input to KPML, and thus must be of the other procedural relations. This assumes, told the appropriate SPL terms to use to however, the existence of a core set of microspecify the required grammatical forms; plans which perform the procedural categorisa an output file name. tion properly. We have only just begun to experiment with the possibility of building the entire For our example, the system sub-network shown network automatically from a more exhaustive in Figure 2 is produced based on the decision corpus analysis. network shown above.4 It is important to note here that the machine learning algorithm is no respecter of systemic linguistic theory. It simply 5 A DRAFTER Example 4
The realisation statements, choosers, inquiries, and inquiry implementations are not shown in the figure.
Given the corpus analysis and the learned system networks discussed above, we will present an 7
Figure 3: The procedure graph for the example
complex of action and object entities that represent the reader’s repairing of some device. This node may be expressed in any number of different grammatical forms depending upon context. The particular CNL sentence for the warning is more complicated. It can be loosely translated as warning the reader against damaging the cover. It also includes information regarding the three semantic and pragmatic features identified in the corpus analysis as relevant for the expression of warning messages:
example of how preventative expressions can be delivered in DRAFTER, an implemented text generation application. DRAFTER is a instructional text authoring tool that allows technical authors to specify a procedural structure, and then uses that structure as input to a multilingual text generation facility Paris et al., 1995. The instructions are generated in English and in French. To date, our domain of application has been manuals for software user interfaces, but because this domain does not commonly contain preventative expressions, we have extended DRAFTER’s domain model to include coverage for do-it-yourself applications. Although this switch has entailed some additions to the domain model, DRAFTER’s input and generation facilities remain as they were.
5.1 Input Specification
In DRAFTER, technical authors specify the content of instructions in a language independent manner using the DRAFTER specification tool. This tool allows the authors to build up procedural hierarchies which include user goals, plans, actions, and a number of procedural relations (cf. Delin et al., 1994). Among the procedural relations is the Plan-Warning relation, which associates preventative or ensurative warnings with plans. Figure 3 shows a procedure which we will use as an example. The procedure describes the repair of some unspecified device. There are three sub-actions, consulting the manual, unplugging the device, and removing the cover. There is also a warning node which warns the reader not to damage the service cover. The actions in the graph are all expressed in terms of the DRAFTER controlled natural language (CNL). For the simple actions, the meaning is relatively clear. “Repair Device” refers to the
AWARENESS is indicated by the first word in the CNL string. The writer can either tell the reader that the action is inappropriate, indicating that the reader is likely to be unaware of this fact , or, as in the example, remind the reader, indicating that the reader is already aware of the danger. INTENTIONALITY — The action may be done either purposely or inadvertently, indicating that the reader is likely to consciously perform the inappropriate action, or, as in the example, to do it unintentionally. SAFETY — The action may either be unsafe or, as in the example, merely unhelpful.
We rely on the author to set these feature values because there is no straight-forward way in which they could be determined automatically (cf. Ansari, 1995). Figure 4 shows a snapshot of the CNL specification in process. Here, the author has already specified the three features by choosing the appropriate words from menus (i.e., remind, inadvertent, and unhelpful), and is now proceeding to specify the action to be prevented by selecting its CNL term from the list of possibilities in the “Select a pattern” window. Because there are many possible actions, the author is given a 8
Figure 4: The controlled natural language input for the example warning message
2. D´ebrancher le dispositif. 3. Enlever le couvercle de service. ´ Eviter d’endommager le couvercle de service.
sub-string search capability to locate the desired action. The items in square brackets are nodes that can be expanded.
5.2 Text Generation Once the input procedure is specified, the author may initiate text generation from any node in the procedural hierarchy. When the technical author generates from the root goal node in Figure 3, for example, the following English and French texts are produced:
Figure 5 shows a portion of the micro-plan that underlies the body of the English text. The warning, shown at the bottom of the graph, is marked as a “Neg-TC-imperative” which is translated into the appropriate SPL commands for the KPML generator. Note that the French version employs e´ viter (avoid) rather than the less common prendre soin de ne pas (take care not). This is possible because the French text is produced by a separate microplanning sub-network. This separate network was not based on a corpus study of French preventatives, but rather was implemented by taking the learned English decision tree, modifying it in accordance with the intuitions of a French speaker, and automatically constructing French systems
To repair the device 1. Consult the repair manual. 2. Unplug the device. 3. Remove the service cover. Take care not to damage the service cover.
R´eparation du dispositif 1. Se reporter au manuel de r´eparation. 9
Figure 5: A portion of the micro-plan for the example text
from that modified decision tree. Clearly, a corpus study French of preventatives is still needed, but this does show DRAFTER’s ability to make use of KPML’s language conditionalised resources. Were we to replace the warning with other sorts of warnings, the expression would also change according to the learned micro-planning network. If authors, for example, wish to prevent the reader from performing the action of dismantling the frame of the device, and they decide that the reader is unaware of this danger, that the action is consciously performed and not unsafe, DRAFTER produces the following text: Do not dismantle the frame. Ne pas d´emonter l’armature.
6 Conclusion In this paper we have discussed the use of machine learning techniques for the automatic construction of micro-planning sub-networks. We noted that because the automatic derivation of useful, welldefined features for corpus analysis is beyond the current state of the art, the painstaking process of corpus analysis must still be performed manually. On the positive side, however, we noted that the learning and construction of planning networks is possible and greatly simplifies the tasks of testing and building text planning resources for new domains. We intend to continue this work by applying the technique to larger portions of the planning resources. We also presented an analysis of English preventative expressions which served as the basis for the learning. We intend to continue this part of the work by addressing more preventative forms and by extending it to other languages.
If authors wish to prevent the reader from performing the action of disconnecting the ground, and they decide that the reader is unaware of this danger, that the action would be unconsciously performed, and that the consequences are indeed life-threatening, DRAFTER produces the following Acknowledgements text: Never disconnect the ground. The authors wish to acknowledge valuable discusNe jamais d´econnecter la borne de terre. sions with Tony Hartley, Adam Kilgarriff, C´ecile 10
the Fourteenth International Joint Conference on Artificial Intelligence, August 20– 25, Montr´eal, Canada, pages 1398–1404. Also available as ITRI report ITRI-95-11.
Paris, Richard Power, and Xiaorong Huang, as well as detailed comments from the anonymous reviewers.
References
Quinlan, J. R. (1993). C4.5: Programs for Machine Learning. Morgan Kaufmann.
Ansari, D. (1995). Deriving procedural and warning instructions from device and environ- RD (1991). Reader’s Digest New Complete DoIt-Yourself Manual. ment models. Master’s thesis, Department of Computer Science, University of Toronto. Rietveld, T. and van Hout, R. (1993). Statistical Techniques for the Study of Language and Bateman, J. A. (1995). KPML: The KOMETLanguage Behaviour. Mouton de Gruyter. Penman (Multilingual) Development Environment. Technical report, Institut fu¨ r InSiegel, S. and Castellan, Jr., N. J. (1988). Nontegrierte Publikations- und Informationssysparametric statistics for the behavioral sciteme (IPSI), GMD, Darmstadt. Release 0.8. ences. McGraw Hill. Clementine (1995). Clementine User Guide, VerVander Linden, K. and Martin, J. H. (1995). Exsion 2.0. Integral Solutions Limited. pressing local rhetorical relations in instructional text: A case-study of the purpose relaDelin, J., Hartley, A., Paris, C., Scott, D., and tion. Computational Linguistics, 21(1):29– Vander Linden, K. (1994). Expressing pro57. cedural relationships in multilingual instructions. In Proceedings of the Seventh International Workshop on Natural Language Generation, Kennebunkport, ME, 21–24 June 1994, pages 61–70. Di Eugenio, B. (1993). Understanding Natural Language Instructions: A Computational Approach to Purpose Clauses. PhD thesis, University of Pennsylvania. also available as IRCS Report 93-52. Horn, L. R. (1989). A Natural History of Negation. University of Chicago Press, Chicago. Isard, A. and Carletta, J. (1995). Replicability of Transaction and Action Coding in the Map Task Corpus. In AAAI Spring Symposium on Empirical Methods in Discourse Generation and Interpretation, pages 60–66, Stanford, CA. McGowan, J. and DuBern, R., editors (1991). Home Repair. London: Dorlin Kingersley Ltd. Paris, C., Vander Linden, K., Fischer, M., Hartley, A., Pemberton, L., Power, R., and Scott, D. (1995). A support tool for writing multilingual instructions. In Proceedings of 11
