* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Download On Mending a Torn Dress: The Frame Problem
Malay grammar wikipedia , lookup
American Sign Language grammar wikipedia , lookup
Lithuanian grammar wikipedia , lookup
Esperanto grammar wikipedia , lookup
Germanic weak verb wikipedia , lookup
French grammar wikipedia , lookup
Scottish Gaelic grammar wikipedia , lookup
Cognitive semantics wikipedia , lookup
Germanic strong verb wikipedia , lookup
Old Irish grammar wikipedia , lookup
Modern Greek grammar wikipedia , lookup
Macedonian grammar wikipedia , lookup
Swedish grammar wikipedia , lookup
Japanese grammar wikipedia , lookup
Udmurt grammar wikipedia , lookup
English clause syntax wikipedia , lookup
Ukrainian grammar wikipedia , lookup
Polish grammar wikipedia , lookup
Modern Hebrew grammar wikipedia , lookup
Chinese grammar wikipedia , lookup
Navajo grammar wikipedia , lookup
Portuguese grammar wikipedia , lookup
Sotho verbs wikipedia , lookup
Old English grammar wikipedia , lookup
Kannada grammar wikipedia , lookup
Turkish grammar wikipedia , lookup
Hungarian verbs wikipedia , lookup
Spanish grammar wikipedia , lookup
Italian grammar wikipedia , lookup
Kagoshima verb conjugations wikipedia , lookup
Russian grammar wikipedia , lookup
Ancient Greek grammar wikipedia , lookup
Georgian grammar wikipedia , lookup
Latin syntax wikipedia , lookup
Serbo-Croatian grammar wikipedia , lookup
Yiddish grammar wikipedia , lookup
On Mending a Torn Dress: The Frame Problem and WordNet∗ Sandiway Fong NEC Research Institute 4 Independence Way Princeton NJ [email protected] Abstract Verbs fall into a variety of semantic classes. In particular, change of state verbs express events where an “affected” grammatical object undergoes a change of state. Properties of the object as expressed by adjectival modification hold true at the start of the event. The problem of determining if such properties are preserved at event culmination is a sub-case of the well-studied Frame Problem. This paper describes the use of WordNet in conjunction with a parser for solving instances of adjective-verb opposition. 1 Introduction Recently, there has been much interest in the form and content of lexical knowledge concerning the event semantics of verbs, (Pustejovsky, 1995), (Rappaport Hovav and Levin, 1998) and (Fillmore et al., 2001). Lexical semantic knowledge can also be applied to problems of logical inference. For example, (Pustejovsky, 2000b) describes an instance of the Frame Problem for change of state verb objects and its solution in the formal Generative Lexicon framework.1 Armed with knowledge about the semantics of verbs and their relation to other grammatical categories such as adjectives, it is possible to draw inferences about events and the effect on the participants. This paper describes how WordNet can be employed in conjunction with an event semantics lexicon to solve instances of the Frame Problem. We consider minimal pairs of the form shown in (1), adapted from (Pustejovsky, 2000a). (1) a. Cathie mended the torn dress b. Cathie mended the red dress infer that as a result of the action of mending, the object dress becomes or has property mended. Using Rappaport Hovav and Levin (1998)’s notation, the event template for mend can be expressed as (2): (2) [x cause [become [ y <mended>]]] where x and y denote the subject Cathie and the affected object dress, respectively. The sub-template become [ y <mended>] encodes the fact that y undergoes a change of state. For change of state verbs, an adjective modifying the object y expresses a property that is true at the onset of the event. The question is whether that property still holds true at the completion or culmination of the event. Returning to (1a), to use Pustejovsky’s terminology, torn and mended are in semantic opposition with respect to each other. That is, they cannot be simultaneously true for a given entity. From this, we can conclude the dress is no longer in state torn as a result of the action of mending. By contrast, if no semantic opposition obtains, there is no reason to assume any change in property status. In other words, there is no reason to conclude that the dress changes color from red in (1b). In this paper, we make precise the nature of the semantic opposition with respect to the network of synonym/antonym relations in WordNet, (Fellbaum, 1998), and describe how WordNet can be employed by an event semanticsaware parser to evaluate cases of adjective-verb opposition. 1.1 Further Examples Further examples of the paradigm from (Pustejovsky, 2000a) are given in (3) through (10).2 (3) Change of state verbs like mend refer to some action resulting in a change of state for their direct objects. For example, in both (1a) and (1b), we can a. The plumber fixed every leaky faucet b. The plumber fixed every blue faucet (4) Mary fixed the flat tire ∗ 2 Pustejovsky The author wishes to thank Christiane Fellbaum for discussion on this topic. 1 In logic, the Frame Problem is defined as the problem of whether statements that hold true of a given state continue to hold after some action has been performed. See, for example (Kowalski, 1979). also discusses general cases of semantic opposition not involving adjectival modification, as in The woman on the boat jumped into the water and The prisoner escaped from the prison. For the purpose of this experiment, we restrict our attention to examples of the type shown in (3) through (10). Figure 1: Shortest path between mend and tear (5) John mixed the powdered milk into the water (6) The father comforted the crying child (7) John painted the white house blue (8) Mary rescued the drowning man (9) Mary cleaned the dirty table As the event templates in (12b) and (13b) indicate, by definition the “effected” object does not exist prior to and subsequent to the event for creation and destruction verbs, respectively. Hence, the solution to the Frame Problem for these cases is technically trivial in the sense it does not involve semantic opposition with respect to an individual verb. The waiter filled every empty glass with water 2 (10) An accurate algorithm must be able to discriminate between leaky and blue in (3a) and (3b), respectively, without ruling out flat in (4). Semantic opposition is not limited to change of state verbs. Activity verbs such as sweep, wipe, broom, paint in (7) and brush have a change of state interpretation when they are modified by a resultative. For example, in (11), the cancellation of the state dirty is only implied in the latter case, i.e. the semantic opposition is between the resultative encoding the end state clean and dirty. (11) a. John brushed the dirty carpet b. John brushed the dirty carpet clean However, the same analysis does not extend to all verb classes. Verbs of creation and destruction are exceptions, as the examples in (12a) and (13a) show. (12) a. Nero built the (gleaming) temple b. [x cause [become<built> [ y exist(+)]]] (13) a. Nero ruined/(destroyed) the (splendid) temple b. [x cause [become<ruined> [ y exist(-)]]] WordNet WordNet organizes verb, nouns and adjectives into fairly distinct networks consisting of synonym set nodes called synsets.3 Synsets are linked via semantic relations such as hypernymy (“instance of”), meronymy (“part of”) and antonymy. In the case of adjectives, the model of clustering around direct antonyms allows an indirect antonym relation to be formed as well. By exploiting these and other relations via transitivity, we can establish whether an appropriate semantic chain exists between a given adjective and verb. Figure 1 illustrates the shortest path between mend and tear. Mend and repair belong to the same synset. Repair and break are antonyms. Break and one sense of bust are in the same synset. A second sense of bust and tear belong to another synset. The link between the two very similar senses of bust is made through a common lexicographers’ source file, i.e. they are both verbs of contact.4 3 For example, the only direct link between the adjective and verb networks is the adjectival participle relation ppl which consists of a total of just 90 links, linking examples like cooing and coo. For this experiment, we use past and present participles derived from comlex to supplement these links. See section 3. 4 According to WordNet 1.6, bust as a verb has 5 distinct senses, 3 of them being verbs of contact. The two senses being automatically linked here are bust as in “ruin completely” and “separate or cause to separate abruptly”. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. mend and fix in same synset fix (200177962,3) and fix (200339066,2) in synsets related by verb.change fix instance of attach attach (200885494,1) and attach (200881541,1) in synsets related by verb.contact attach instance of touch touch (200820504,1) and touch (200820743,1) in synsets related by verb.contact touch see also touch down touch down instance of land land (201348563,1) and land (201349748,3) in synsets related by verb.motion land and shoot down in same synset shoot down (201349748,2) and shoot down (201405541,3) in synsets related by verb.motion shoot down and tear in same synset Figure 2: Antonym-less path from mend to tear There are another five ways (all involving a longer chain) to get from mend to tear using WordNet. Four of these also involve a single antonym relation consistent with the chain from figure 1. However in general, it is reasonable to assume that the longer the chain, the less certain or reliable the information imparted by that chain. For example, the somewhat implausible-looking chain for mend to tear shown in figure 2 has a dozen links.5 Also, crucially, it does not involve an antonym relation. Hence, we restrict our attention to shortest chains only. 3 Implementation A prolog-based sentence parser was used to analyze the Pustejovsky examples given in (1) through (10). The parser computes both a parse tree based on event structure and the shortest WordNet path for relevant configurations. An example of the output for Mary rescued the drowning man, example (8), is given in figure 3. As explained in section 1, an appropriate configuration will contain an NP object modified by an adjective in the context of a change of state verb (or an activity verb modified by a resultative). If a single antonym link is present in the shortest path, the modifying adjective is marked with the feature cancelled to signify semantic opposition, i.e. the property no longer holds true on event completion. In the case where no antonym link is found or the verb does not allow a change of state reading, the property represented by the adjective remains uncancelled. We briefly describe the relevant components of the parser below: • The prolog version of the WordNet verb/adjective system was employed. This is a network of approximately 174K nodes and 600K links. Breadth-first search, implemented using a hybrid prolog/C program, was used to compute shortest chains.6 5 Synset and offset numbers are supplied for some of the relations shown to distinguish between different senses of the same word. 6 The WordNet databases are stored in prolog for flexi- • The broad-coverage lexicon used for the experiments was formed by combining parts of comlex, (Grishman et al., 1994), with an event semantics verb lexicon derived from (Levin, 1993). The adjective subsystem is composed of 6.2K basic adjectives from comlex (WordNet has about 20K entries) plus approximately 9.5K past and present participle verb forms. 4 Results and Conclusions The crucial notion being tested in this paper is the general transitivity of semantic relations. A pertinent question to ask then is: What are the limits to transitivity, especically when applied to heterogeneous relations? Empirically, how many relations can be chained together before reliability is compromised and we end up with degraded or unwanted inferences? Can some sort of thresholding be applied to this problem? The table in figure 4 reports the results for examples (1), (3) through (10). Clearly, the shortest path algorithm appears to work well, producing the correct value of the semantic opposition feature in 9 out of the 11 cases. A few comments are in order at this point: The Value of Thresholding Currently, there is no upper limit on the length of the shortest chain. However, a length threshold is difficult to establish. For example, fix and blue are connected by an 11-link chain, which is arguably too long to be trustworthy. However, the semantic opposition between rescue bility and accessed via (efficient) first-argument indexing during searches. The breadth-first tree is implemented in C for efficiency and to avoid triggering prolog garbage collection for reclaiming search tree storage. On a 200MHz Sun Sparc, the search rate is approximately 26K nodes/sec. As figure 4 attests, finding the shortest chain takes no more than a second even for the largest search. In future work, a bi-directional breadth-first algorithm will be substituted for further efficiency gains. Figure 3: Mary rescued the drowning man and drowning relies on a shortest chain containing 13 links. The Shortest Path Criterion The implemented algorithm fails to discover a semantic opposition relation for fix -flat, example (4). There are 18 senses of flat as an adjective, one of them directly referring to a flat tire. Flat and deflated are clustered adjectives. Fix and deflate can be connected by the 7-link shortest chain shown in figure 5. This chain lacks an antonym link. Actually, WordNet is missing a link here between deflated as an adjective and deflate as a verb due to the incompleteness of the adjective participle of verb relation (ppl).7 Even assuming the existence of this link, our algorithm will still fail to find any semantic oppo7 See note 3. sition in this case. However, there does exist a chain (not shown here) containing the required single antonym relation between deflate and fix ; the only problem being that it is one link longer than the shortest chain reported above. The Utility of Domain-specific Opposition The algorithm also fails to discover an antonym chain between blue and white, example (7). In WordNet, there is no direct or indirect antonym relation between colors.8 8 A reviewer pointed out that blue and white, like rich and poor, are contraries, cf. contradictions such as present and absent. Both propositions forming a contrary may be false. For example, a house may be painted in a third color, and someone may be neither rich nor poor. We make two observations here: (1) WordNet does not make a principled distinction between contradictions and contraries. For example, it contains direct antonym links between pairs such as Candidate Pair mend-torn mend-red fix-leaky fix-blue fix-flat mix-powdered comfort-crying blue-white rescue-drowning clean-dirty fill-empty Shortest Chain (No. of links) 5 – 5 11 – 6 9 – 13 1 1 Semantic Opposition Yes No Yes No No* Yes Yes No* Yes Yes Yes Size of Search (No. of Nodes) 1261 11974 12167 14553 12286 11931 11359 24431 9142 61 48 Figure 4: Shortest path results 1. 2. 3. 4. 5. 6. 7. deflate instance of collapse collapse instance of fold fold (200872449,1) and fold (201042123,1) in synsets related by verb.contact fold instance of lace lace (201042567,6) and lace (201045661,1) in synsets related by verb.contact lace instance of tie tie instance of fix Figure 5: Shortest path from deflate to fix This deficiency can be temporarily rectified by assuming a default rule that semantic opposition obtains whenever no chain can be found. However, although it correctly accounts for the data when limited to this test set, it is unlikely to work in general because it implies an antonym-less shortest chain must exist for all cases not exhibiting semantic opposition. More plausibly, we can restrict the domain of the default rule to color verbs and adjectives only. All three of the above difficulties can arguably be laid at WordNet’s door. Thresholding and the shortest path criterion can only work if the length of the chain is inversely correlated with reliability. In the case of the colors example, WordNet could be augmented with (possibly domain-specific) rules about semantic opposition. To summarize, we have described how WordNet can be used to help solve the semantic opposition task. An experimental system was constructed and tested on the Pustejovsky examples. Finally, note that WordNet was constructed without the benefit of a corpus. To further explore and more precisely determine the limits on general transitivity of semantic relations, future plans include testing the rich and poor, left and right, and early and late, as well in the case of present and absent. Hence, WordNet is simply incomplete with respect to the color domain. (2) In the case of a change of state verb, the final state is asserted. In other words, there is no empirical distinction between a contrary and contradiction in this context. system with its broad-coverage lexicon on a large semantic opposition dataset to be extracted from corpora. References C. Fellbaum, editor. 1998. WordNet. MIT Press. C. J. Fillmore, C. Wooters, and C. F. Baker. 2001. Building a large lexical database which provides deep semantics. In Proceedings of the 15th Pacific Asia Conference on Language, Information and Computation (PACLIC-15), pages 3–25. R. Grishman, C. Macleod, and A. Meyers. 1994. Comlex syntax: Building a computational lexicon. In COLING ’94, pages 268–272, Kyoto, Japan. R. Kowalski. 1979. Logic for Problem Solving. North-Holland. B. Levin. 1993. English verb classes and alternations: A preliminary investigation. University of Chicago Press. J. Pustejovsky. 1995. The Generative Lexicon. MIT Press. J. Pustejovsky. 2000a. Event-based models of change and persistence in language. IRCS Colloquium Series handout, University of Pennsylvania. J. Pustejovsky. 2000b. Events and the semantics of opposition. In C. Tenny and J. Pustejovsky, editors, Events as Grammatical Objects, pages 445– 482. CSLI Publications. M. Rappaport Hovav and B. Levin. 1998. Building verb meanings. In Butt and Geuder, editors, The Projection of Arguments: Lexical and Compositional Factors. CSLI Lecture Notes No. 83.