File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/concl/79/j79-1070_concl.xml
Size: 27,087 bytes
Last Modified: 2025-10-06 13:55:51
<?xml version="1.0" standalone="yes"?> <Paper uid="J79-1070"> <Title>American Journal of Cornput ational Linguistics TEXT UiiDERSTANDING: A SURVEY</Title> <Section position="18" start_page="1" end_page="11" type="concl"> <SectionTitle> 1 I </SectionTitle> <Paragraph position="0"> specified goal) so he can enjoy pleasure&quot; (high-level goal). The purpose of clearly defining a goal hierarchy is to enable understanding of situations in which an actor faces goal conflict and elects $0 pursue the highest goal (e.gw &quot;preserve health&quot; rather than &quot;en joy pleasure&quot;), But general knowledge about Row goals can be achieved must also be available. The &quot;D-goal&quot; is proposed as the fundamental unit of organization for suhh information. A D-goal ie a point pf access to planning information for the realization of some goal. Foe example, D-KNOW is the D-goal for the goal of knowing something, D-CQNT for the goal of physically controZling something, and D-PROX for the goal of being in proximity to something. The D-goals' associated knowledge is an ordered list of &quot;planboxee&quot;, each of which ptovides detail infoption on one method for achieving the goal. The ordering prwides thb sequence in which the methods are likely to be considered. For example, D-KNOW has an ASK planbox (as a highly likely method) which specifies the actual act of asking the appropriate question, the intended result (fee. getting the answer), and the pre-conditions necessary to successfujlly ask. Some of the pre-conditions are: communication is possi&quot;b1e; the persbn being asked knows the answer; the person being asked is disposed to answer the question, When a planbbx is ext rernely specific, it becomes a script. Using the telephone book is a method of acquiring certain kinds of knowledge that is so conventional that it is a script, Certain 'recurring sequences of B-goals are called &quot;named plans&quot;. The named plan, USE(x), stands for: D-KNOW(the location of \x)</Paragraph> <Paragraph position="2"> perform p_reparatlons and do action appropriate to x Figure 3-l(b) sham the relationship of D-goals, planboxes, scripts and named plans to each other and to the goal hierarchy.</Paragraph> <Paragraph position="3"> The representation of the plans of actors in a text would be at the KS level (with scripts). The D-goals would be explicitly represented and would be connected to actual acts at the CD level which (attempted to) implement the plan For example, the sentence: John tried to find out who ate the candy.</Paragraph> <Paragraph position="4"> would be repreaentsd arro: That is, the plan of trying to know who ate the candy was implemented BS some unspecified act (DO) which failed to achieve the plan's goal. Acts (at the CD level) which meet pre-conditions of planboxes would be linked to the D-goal they enable.</Paragraph> <Paragraph position="5"> Additional information is suggested as being appropriate and necessary for the KS level. This includes, for each character, his goals, the current status of each goal, strategies which could be used to achieve each goal and facts (true information as opposed to actual occurrences) relevant to goal understanding. This inPS onnation i.s maintained on &quot;Coal Fate Graphs&quot;, which also contain associations between characters and any &quot;Themes&quot; (large goal complexes such as BECOMING-RICH) in which he is participating.</Paragraph> <Paragraph position="6"> Page 76 Schapk and Abeleon appear to be committed to the development of a very complex system for the organization of knarledga, and the representation of such knowledge when it bccurs in stories. It should be apparent that they rare attempting to model goal related human knowledge and actions, since such knowledge and actions are common in stories. Since this knowledge pn be required to underatand particular stories it 14 not clear at what point (if any) one ceases t o study text unde rs t anding and begins modelling personality . Phillips*[75a,75b] presents probably the most comprehensive computational model of text, in that he is concerned with the representation of all types of knowledge and textual relatiorships. This breadth is informative, but necessqrily results in a lack of depth in some areas. He presents a representational scheme which he uses for the vari~us required types of knowledge, which include world knowledge, li%gulstir knowledge and the knowledge conveyed by the text. Phillips w.orld knowledge, &quot;the encyclopedia&quot;, consists of both the static data structures and the dynamic processes which operate upon these structures. The static data structure is a fairly conventional but very well defined, semantic network (closely related to the suggestions of Hays [73] ) which use nodes for entities and events, and arcs for the relationships between nodes. The set of hierarchical relationships between entities (the taxonomic- structure) is called' the paradigmatic structure. The syntagmat ic structure is the see of relationships between events and the event particip ants (case or argument relationships). Page 77 Phillip8 incroaucea a Modality node attached to every event which is always used to refer to the event and its participants as a whole. (In many systems this is represented only indirectly.) The discourse relationships of cauoali ty , time ordering and spatial relations are rep resented by arcs between the Modality nodes of events. Phillips defines another type of relationahip which he designates the Metalingual organization of knowledge. This allows him to represent a concept as a unit, and yet have a complete subnetwork elqborating the meaning of that concept. For example, the unitary concept &quot;poison&quot; would have as its composition &quot;someone ingesting something which causes that person to be ill&quot;. This rep resentational technique provides the fundamental capabilities of frames or scripts - expansion of something into its parts, and knowledge about those parts.</Paragraph> <Paragraph position="7"> The processes that operate on the semantic network are divided into two classes, The first, path-t racing, involves only following pathts through the paradigmatic structure. This type of process would be used to find that &quot;Mary gobbled caviar&quot; was a more specific instance of &quot;People eat food&quot;. The second type of process is pattern-matching, and involves constraints between the components, For example, determining that &quot;John killed himselft'is suicide while &quot;John killed ~111&quot; is not, requires a coreference constraint in the definition of suicide. Phillips observes that path-tracing is computationally equivalent to finite state automata, while pattern-matching is necessarily more complex.</Paragraph> <Paragraph position="8"> Page 78 Discourse, or text, is represented using the same representational scheme, but may be characterized by properties not applicable to discontinuous knawlddge. The first of these properties is connectedness. Two propositions are paradigmatically connected if each has an argument such that the two arguments have a common paradigmatic superordinate node, or if the first is the immediate superordinate of the the second. For example, &quot;lions&quot; and 'It igers&quot; both have &quot;animal&quot; as superordinate, so the following are connected paradigmatically: Lions are indigenous co Africa. Tigers have stripes.</Paragraph> <Paragraph position="9"> And the two propositions: Man is a hunting animal. Modern man hunts for sport.</Paragraph> <Paragraph position="10"> are also connected since the second is a more specific proposition (&quot;modern man&quot; is immediately subordinate to ''man&quot;) than the first. Two propositions are discursively connected when discursive relations (em g. causality ) exist between them.</Paragraph> <Paragraph position="11"> The second attribute of discourse is thematicity. A theme is a prescribed pattern (represented by a Metalingual construction) to which a discourse may conform. A theme may be ~'contentive&quot; like &quot;accidental drowning&quot; t l which is represented as existing when a person is caused to be in the water and is unable to act, the combination of which causes him to drown&quot;. This is contentive siqce it specifies both the parts (events) of the theme and their interrelationships. A 'lnon-contentive&quot; theme is one that provides only a structural pattern or the interrelationships between some unspecified entities &quot;clue&quot; is defined to be &quot;an unobserved act that causes the existence of something&quot;. The exact events and results are unspecified. A discourse is thematic if its propoaitiona can be matched to a theme or hierarchy of themes (i.e. a theme which matches several propositions in a discourse, such as &quot;accidental drowning&quot;, may in turn be a part of another theme, such as &quot;tragedyft). Phillipa then defines a coherent discourse as one which is both connected (each proposition is connected to at least one other proposition) and thematic.</Paragraph> <Paragraph position="12"> Phillips points out that although there is no simple one-to-one mapping between his representations and the constructs used in text analyses like those of Grimes (section 1.2.3) and Propp (section 1. .2), his structures do provide for the representation of most of the proposed relationships, Some of Grimes rhetorical predicates, such as Attribution, Specific and Collection, correspond to paradigmatic connectedness, while others, such as Covariance, correspond to discursive copnectedness. Still other rhetorical predicates, such as Respqnse and Analogy and Propp's patterns of functions in a move (as well as the functions, themselves) correspond to thematic structure. Hence, Phillips claIms to have presented computational interpretations for the principle text phenomena.</Paragraph> <Paragraph position="13"> As a test of his proposed model of text, Phillips presents a model of text understanding embodied in a computer program. His program inputs a discourse in the form of parse trees of the sentences and builds the knowledge representation of the discourse. An interesting use of the Metalingual const mcti on is its function in replacing non-cognitive surface words by the Page 80 appropriate cognitive structure. For example, the preposition &quot;through&quot; is replaced by the cognitive st tucture for &quot;in-contact-wLth&quot; in a sentence like: The Abominable Snwman walked through the snow.</Paragraph> <Paragraph position="14"> &quot;fhus, the same representational mechanism is used for such divergent types of knowledge as syntact ically-related information and thematic structures. Once the parse tree has been converted into the knowledge representation of the proposition, the understanding process involves two major steps. The first fs begun by matching the input proposition (IP) against the encyclopedia to find a corresponding generalized proposition (GP). Thus, the IP The boat contains Horatio Smith.</Paragraph> <Paragraph position="15"> is matched to the GP Things contain peaple.</Paragraph> <Paragraph position="16"> Notice that Phillips uses GPs to capture the same knowledge that features and selectional restrictions capture in many other systems. So the correct interpretation of a GP is that it is a plausible proposition, not a necessary truth. Once the IP has been matched to a GP, three additional checks are made starting from the matched GP. One is to determine if any of the terms of the proposition have Metalingual definitions. If so, new knowledge correspondin& to the definition is added to the discourse representation. For example, if the discourse included that &quot;John was poisoned&quot; the elaborated information that &quot;John ingested something that caused him to be ill&quot; would be added. A second check is to see if the matched GP is discursively connected to any Page 81 other GPs. If so, new k~lowledge is added to the discourse which corresponds to the related GP and the d$.ecursive connection. If the rnatche'd GP were &quot;People are injured&quot; and it had a causal link to &quot;People are unable to act&quot;, then the IP &quot;John was injured&quot; would result in the addition oE &quot;John was unable to act&quot; with a causal link from the original IP, It is by using these interrelationships of GPs that Phillips accomplishes inference. The third test is to determine if the matched GP is a part of any 'contentive theme. (The encyclopedia's GPs have pointers to all contentive themes which contain them.) If so the theme is matched against the discourse as a whole to see if all components are present and all constraints satisfied, If they are, the theme is addtid to the discourse rep resent ation. &quot;Accidental drowning1' would be added to a discourse representation which had matched the GPs and connections I' (&quot;People contact- water&quot; and &quot;~eople cannot act&quot;) cause (&quot;People drown&quot;) &quot; with a coreference constraint on &quot;people&quot;, After processing Ips and adding the related structure to the discourse representation, the second major step is testing the discourse for coherence. This involve two tests. The first: determines if the discourse is connected. Then, thematicity, is tested by checking to see if a single undominated theme has been found. It should be noted that non-contenrive themes, since they have no component actions, and thus cannot be pointed to by GPs, must be tested for in a serial fashion. If the discourse passes both tests, then it is judged coherent.</Paragraph> <Paragraph position="17"> Page 82 To illustrate the process, a very brief description of the understanding of a atory will be given. Given the story: Phillips principle goal was a computational model of text* His model of text understanding was intended to demonst rate that the proposed representations could actually be built from an input text. His text model does seem to present a reasonable representation for a number of text phenomena, several of which have not been considered by other computational models, Paradigmatic connectedness and some types of thematic structure seem particularly important. Iiowever, several object ions must be raised to his model of text underslanding. The first objection is that his testa for coherence are applied only to the complete discourse, and are not formulated in such a way as to suggest strategies to avoid incorrect interpretation of propositions pnd relationships, If this were done, the sequential processing of input p topositions would continually test for coherence, and incoherence would immediately suggest that some misinterpretation might have been made, or some inference omitted. A second objection i& that the definitions of connectedness and thematici ty are inadequate. A common superordinate node is simply not sufficient to explain paradigmatic connectedness. The example of sentences about lions and tigers would be connected only in some context that explained why these statements were being made (e, g. &quot;All* I know about lions and tigers is . . .&quot;). Thematicity is defined without respect to how much of the discourse the theme accounts for. A theme cannot account for all of the propositions in a discourse unless the discourse is very trivial, and yet if a theme matched only the first three propositionS of a one hundred proposition discourse, it could hardly be called the theme of the discourse. Finally, a number of aspects of the understanding processes are not convincingly shown to be computationally practical. The problem of avoiding incorrect additions to Page 84 the discourse representation from a rich encyclopedia is ignored. Each exawle has exactly the right infamation available. Also, it ie mot clear that: the use of generalized propositions would work when there am ~averrrl .levels of generalization possible. For example, the g perelized proposition &quot;People contact water&quot; has associated knowledge, but the higher level proposition &quot;People contact thingstt would also need to be present with its associated knowledge. And finally, the methods of accessing themes - by pointers to all occurrences of generalized propositions or by serial search both seem comput at ionally unaccep table. Propositions like &quot;~eople go&quot; would result in, a combinatoric explosion of possibilities. For all of these reasons, Phillips' understanding model is useful more in suggesting the kinds of problems involved, than in providing an actual model of text understanding.</Paragraph> <Section position="1" start_page="1" end_page="11" type="sub_section"> <SectionTitle> 3.3 Other Work </SectionTitle> <Paragraph position="0"> Space limitations and the narrowly defined scope of this sunrey have combined to eliminate certain interesting work from detailed consideration.</Paragraph> <Paragraph position="1"> Wilks [75, 76, 771 has described a text understanding (and translation) system which uses a meaning representation called Preference Semantics. The system normally opercltes in a basic mode, dealing with sentences individually, but is capable of entering an extended mode when a reference problem occurs. An example is &quot;it&quot; in John drank the whiskey from the glass and - it felt warm in hi3 stoomch.</Paragraph> <Paragraph position="2"> Several processes are involved in the attempt to determine the correct referent. &quot;Ext ractionl' is the addition of logically true propositions, only Page 85 wtlicSt in the text, but available from the meanings of the unite. Thus, both &quot;ltm was in John's stomach, The vhiskey is in a part of Johnroipld ath be extracted. These two, pqopositions can be identified, thus resolving &quot;it&quot; as &quot;the whiskey&quot;.</Paragraph> <Paragraph position="3"> *I Bwever, had the above example been . and it was good.&quot; the original pruposifions, as vell as relevant extractions, would then be subjected to m ccuncm sense inference rules&quot; such as 1. <1 drink 2) -> (1 judges 2) 2, (1 is good) <-> (2 wants 1) rt3ich would be used to find the shortest inference chain identifying the referent. For the example, when &quot;the whiskey'' is tested, the established (John drhk whiskey) -I-> (John judges whiskey) -subset-> (John wants whiskey) <-2- (Whiskey is good) where %ant&quot; events are a subset of &quot;judge&quot; events. A recent addition to Wilks' system is large knowledge structures (like frames) called &quot;pseudo-texts&quot;. They are suggested as containing other knavledge which might be required in solving reference problems (although they have other important use in understanding individual words).</Paragraph> <Paragraph position="4"> Page 86 At the levei of text, Wilks differs from Rieger and Charniak primarily in his insistence upun entering an extended made only when a problem requires it. Reference p toblems are the only problems he discusses as triggering thia mode His use of pseudo-text B in establishing textual connections is still too briefly described to critically evaluate.</Paragraph> <Paragraph position="5"> Rieger ~([75b, 75c, 76a, 76b1, Rieger and Grinberg [771) has described a complex system for the organized representation of cause and effect knwledge, .a& plans utilizing this knowledge. A set of decision trees (&quot;selection networks&quot;) are postulated which. given a goal, select certain common sense algorithms&quot; capable of realizing that goal. Rieger asserts that the same knowledge st-ructures used for planning, should be used for understanding the intentional acts of others, To do this in text, he requires knwledge which allows prediction of goals and actions likely to be made in response to the occurrence of some event or state, A subsequent sentence is tested to see if it confirms any prediction by being a step in an expected action or in an algorithm to achieve an expected goal, This testing requires that all algorithms be indexed by step. Thus, it must b~ possible to find all occurrences of (X GOT0 Y) in some set of algorithms, A confirmation is used to more- coslfidently predict the course of action being followed, Clearly, Rieger is dealing with the same problems as Schank and Abelson (section 3,2.1.2), but ha.s not yet clearly demonstrated the utility or practicality of his approach to this aspect of text understanding.</Paragraph> <Paragraph position="6"> Page 87 Hobbs 176, 771 has discussed analyses of various texts based on his own system of semantics. Possible fntersentential relations are described by pattern - action pairs which, when matched by input sentences modify the text representation tree apprap riately. These relations include causality, time ordering, paraphrase, examp 1.2, contrast, parallel const.ruction and vlolated expectation. These relationship8 indicate that Hobbs does not make a distinction between content and form (author imposed) relations. Thc complexity of Hobbs' system (in the large number of rules and their interactions) makes evaluation difficult until he has completed his computer implementation.</Paragraph> <Paragraph position="7"> Other work that should be mentioned includes Schmidt ([761, Schmidt and Sridharan [77]), who discusses the problem of recognitioh of plans from actions, as well as Novak [76] and Bobrow, et a1 [77] who both use frame like knowledge structures in specialized language understandi rag systems.</Paragraph> </Section> <Section position="2" start_page="11" end_page="11" type="sub_section"> <SectionTitle> 3.4 Discussion And Conclusions </SectionTitle> <Paragraph position="0"> What have the discrissed computational models added to the previously described model of text understanding? Charniak, Riegex and Wilks have demonstrated how-it is possible to infer information only implicit in a text.</Paragraph> <Paragraph position="1"> Schank has especially examined the richness and complexity of causal connections, which are often implicit. This capability raises many additional questions, however, including when, and how many, inferences should be made.</Paragraph> <Paragraph position="2"> Whi le Cllarniak and Rieger suggest making many infe rences whencvv r possible, Wilks argues for making them only as required. Neither approach has yet been Page 88 applied to sufficient ly 1 argc trxt s and knwledge bases to convincingly demonstrate its validity.</Paragraph> <Paragraph position="3"> Pre-existcnt kn~wledgr structures have been suggested by Schaak and Philips to avoid the extrc:mc8ly difficult problem of making matiy-step inference chains to establish inlpli ci t con~~cct ions. Although use of these structures has demonstrated their ability to meet this goal, as well as their usefulness in generation of summaries and paraphrases, problems have appeared. Complex knowledge structures and real texts present many difficult.kea la matching an input proyosltlon with an element in the structure. When the match is imperfect, or- when many choices are (computationally) posgible, it is very difficult to perform the required matching correctly, In an attempt to deal with this difficulty (as well as others), and to recognize the fact that novel situations are also unde rstandable, Schank, Abelson, Rieger and Schmidt have been concerned with recognizing the motivations and intentions of actors. This aspect of understanding is clearly important, for the stated rt3asons, but the approaches described have almost certainly been too simplif icd for the understanding ,of actual texts.</Paragraph> <Paragraph position="4"> It Thus *the requi rcmcnt that an understar~ding of a text contains as much inferred inforniat~o~~ as h~!c~ss~iry to mqet some minimal level of ...</Paragraph> <Paragraph position="5"> coherence&quot; (section 1.4) is wen to bc a computationally difficult problem. The orgat-lizing cor~tt~i~t SI 111, tt~rl~~; arc i~sc~ful in dealing with this problem, but create new problems. Knocll c-dgc) nrltl rr~p resr~nt at ion of plans and intentions is introduced tn hnndalr* sor.\c3 of thcsr problems. Tt should be noted that the previously dj sc~~r;sotl nrotlc~l tl~~ not spcrl f i c,illy distinguished this type of Page 89 knowledge. Coherence has been defined in tllcse computational models as connectivity established by inferrencing, matching a knowledge structure or being a step in a plan. The exact distinctions between these three types of knowledge are still blurred, and it remains to be demonstrated that these distinctions are appropriate and adequate.</Paragraph> <Paragraph position="6"> Only Phillips and Hobbs have addressed representing the form of texts, bqt these suggestions have not beeh computationally adequate, nor has there been any clear distinction between form and content structures. A better defined notion of text form should certainly play a role in the ongoing determinatiion of textual coherence that occurs during comp rehension in the text understanding model.</Paragraph> <Paragraph position="7"> With regard to forgetting, little has been added, although higherlevel knowledge structures could be used to summarile their more detailed components (which could then be &quot;forgotted'). Schank has also suggested that the least connected propositions in a representation are those most likely to be forgotten. However, these ideas have not been seriously investigated. In general, computational models understand texts perfectly (if at all), and do not contain any imperfect ret rieval processes. Permanent learning and integration with prior knowledge has not been investigated in these systems, nor has any explanation been offered for occasional recall of surface text, --These systems generally do not keep the surface text at all, but could easily do sv. However, it would result in perfect recall of this information.</Paragraph> <Paragraph position="8"> Page 90 The principal value of computational models has been the demonstration of the great difficulty in actually specifying comprehension algorithms. It is all too easy, when explaining hw a particular result is achieved, to ignore the problem of avoiding incorrect paths. The fact that higher-level content st ructures are computationally uaeful, and also psychologically indicated, is an important confirmation. The investigations of more, actual texts by systems having more knowledge available (not just the relevant knowledge) is an important step in validating the models being proposed.</Paragraph> <Paragraph position="9"> Page 91</Paragraph> </Section> </Section> class="xml-element"></Paper>