<?xml version="1.0"?><!DOCTYPE article SYSTEM "/project/take/software/searchbench_offline_processing/paperxml_generator/aclextractor/src/python/../resource/dtd/paperxml.dtd"><article><header><firstpageheader><page local="1" global="115"/><title>Towards a Grounded Model for Ontological Metaphors</title><pubinfo>Proceedings of the Student Research Workshop associated with RANLP 2011, pages 115-120, Hissar, Bulgaria, 13 September 2011.</pubinfo><author surname="Nayak" givenname="Sushobhan"><org  name="Indian Institute of Technology" country="India" city="Kharagpur"/></author></firstpageheader><frontmatter><p><b>Towards a Grounded Model for Ontological Metaphors</b></p><p><b>Sushobhan Nayak</b></p><p>Indian Institute of Technology, Kanpur, India</p><p>snayak@iitk.ac.in</p></frontmatter><abstract>A novel approach to learning metaphors with­out any prior knowledge is proposed, in which ideas are acquired as concrete concepts and later on develop their abstraction. A grounded model of linguistic concepts and a hierarchical probability map is used to interpret/generate ontological metaphors. </abstract></header><body><section number="1" title="Introduction"><p>Consider the following sentences:</p><doubt alpha="65.7" length="70" tooSmall="False" monospace="0.0">• You arewastingmy time.     (TIME IS MONEY)(Lakoff and Johnson(1980))</doubt><p>• We need to <i>combat </i>inflation. (INFLATION IS AN ENTITY)(Lakoff and Johnson(1980))</p><p>These usages are so ingrained in everyday conver­sation that we fail to recognize that they are ac­tually metaphors that try to describe abstract con­cepts in terms of concrete ones(Lakoff and John-son(1980)). A proper understanding of language and thought therefore calls for an increased and focused research into metaphors and the way they are acquired, interpreted and generated.</p><p>As such, there have been many attempts at in­terpreting metaphors over the years. The ear­lier works like Wilks(1975); Fass(1991) are based on a violation of selectional restrictions in a given context. Kintsch(2000) effectively uses LSA to interpret metaphors like "My lawyer is a shark". However, these models are incapable of handling Lakoff and Johnson(1980))'s view of metaphors. The works that are closest to the modern view, in their attempt at interpret­ing common text from financial or other domains, encompass Narayanan(1997); Narayanan(1999). Shutova et al.(2010) also show metaphor para­phrasing using noun-verb clustering. However, they both need a hand-coded seed set or metaphor repository to do further learning. The model pro­posed here assumes no prior knowledge. From multimodal input, we first ground some basic con­cepts and using them, exploit the language syntax to learn/interpret/generate metaphorical mappings in a natural way that emulates language learning in an early learner.</p></section><section number="2" title="Proposed Model"><p>The models till date have looked at metaphors as something that is acquired/interpreted differ­ently than common language. However, the treat­ment by Lakoff and Johnson(1980)) suggests that we should look at metaphor acquisition as we look at language acquisition, and not something that is interpreted/acquired <i>after </i>we have acquired the basic nuances of the language. There is am­ple evidence in literature to suggest that basic linguistic forms can be grounded(Roy and Re-iter(2005)). Consider an early learner who has acquired the grounded concepts of the verb <i>im­pale </i>and understands that, based on its experience till date, only a living entity is capable of exe­cuting it. Then it comes across this:"<i>'Arrogance impaled him." </i>Now based on this linguistic input only, in the absence of any other physical stim­uli, what is to stop the learner from interpreting 'arrogance' as a living entity? It is only when it comes across other usages of the concept 'arro­gance' that its initial idea that 'arrogance is an en­tity' might be modified to 'arrogance is an abstract concept'. However, as one might notice, 'arro­gance is an entity' is actually a well established metaphor. This leads us to look at metaphor ac­quisition in a different light. Why look at 'arro­gance' etc. as abstract concepts that <i>need </i>to be un­derstood through grounded concepts <i>later? </i>Why not look at them as <i>grounded </i>concepts, which <i>later </i>acquire the abstractness, there by being im­parted with metaphorical mappings, which is sug­gested by the example alluded to before? The proposed model takes this approach, where we start with grounding some very basic verbs/nouns and then go on to acquire/interpret/generate on-tological metaphors (a metaphor in which ab­stract notions are projected through concrete con­cepts of objects/entities/substance etc.<page local="2" global="116"/>, i.e. some physical entity.), just as we would do any other words/concepts.In the discussion that follows, it's assumed that language representation incorporates Langacker(1987))'s <i>image schema, </i>and that lan­guage understanding incorporates Embodied Con­struction Grammar(ECG) (Bergen et al.(2004)).</p><subsection number="2.1" title="Grounded Forms"><p>It is more or less established in literature that lin­guistic concepts are cognitively characterized in terms of <i>image schemas, </i>which are schematized recurring patterns from the embodied domains of force, motion, and space(Langacker(1987); Lakoff and Johnson(1980)). Before we go into learning ontological metaphors, we create a grounded system that helps modify/nurture image schemas of new concepts as they come along.</p><p>Almost all types of ontological metaphors come under three broad categories of an Object, a Sub­stance or a Container. In fact, these concepts emerge directly for an early learner through phys­ical experience(Lakoff and Johnson(1980)). A probable scenario for an artificial learner can be to try grounding concepts from multimodal inputs of image, sound and written transcripts. Consider the work by Mukerjee et al.(2011), where they try to discover coreference using image grounded verb argument structure. From multimodal input of a video, and associated narration, they have been able to learn the verb structure CHASE(A,B). This can be further extended to derive the im­age schemas of CHASE and the actors A and B. Mukerjee et al.(2011) use velocity features of the objects in the video to unsupervisingly clus­ter them to find the cluster of CHASE(A,B). This cluster that contains much of the velocity re­lated information for concept CHASE can be pre­sented as an image schema for CHASE(A,B); and since CHASE is a two-party interaction, learning CHASE means the concept of agents A and B are also learned. In the present model, the cen-troid, maxima, and minima of the feature cluster have been stored to represent CHASE<footnote anchor="1"/>, whereas</p><p>'This is just the chosen representation. One might well like to store the whole cluster for representing the action. The claim here is not that the present representation is the one that is actually there; the idea is just to demonstrate that even such a crude model works.</p><p>A and B are being treated as <i>point objects, </i>i.e. en­tities whose behavior would remain unchanged if they are replaced by geometrical points in the vi­sual space they act in. Storing maxima/minima also helps us create image schemas of adjectives</p><p>SLOW() and FAST(). In this model, these func­tions take an action, say CHASE, as argument, and output the minimum or maximum of the ac­tion's velocity feature. With this sort of grounding at hand, we can handle the mental simulation part in ECG(Bergen et al.(2004)) used for understand­ing linguistic occurrences. A <i>fast(chase) </i>would mean the simulation runs with the image schema of CHASE, with the velocity features being main­tained at their maximum.</p><p>We now have grounded forms of CHASE(A,B),</p><p>SLOW(),FAST() and entities(point-objects). We next take cues from Mukerjee and Sarkar(2007) and learn IN(A,B), OUT OF(A,B), INTO(A,B) and the 'container'. Mukerjee and Sarkar(2007) use Voronoi-based features to distinguish space into the interior or exterior of an object. In this model, the image schemas of IN(A), OUT OF(A), INTO(A) and the 'container' are interpreted as be­ing interconnected. While the boundaries of the object-container in the visual input are taken as the boundaries of concept 'container', IN(A,B) is represented by substance/entity A inside container</p><p>B. INTO(A,B) and OUT OF(A,B) are schemas in two states, where the object/substance A is in/out of the container B in one state, and changes its po­sition in the other. Substance is crudely grounded as something that can't be represented by a point object, i.e. something that is not executing rigid body motion. Essentially, combined with, say INTO(A,B), if in the motion schema simulation of action INTO(A,B), A can't be represented by a point, it is taken as a Substance. This fur­ther allows us to ground adjectives MORE(A) and LESS(A) for Substances, based on the change they bring about in the volume of the Voronoi interior.</p><p>To reiterate, the objective of this section is not to claim that a proper image schema has been de­veloped for the above concepts. The goal here has been to show that even from a simple multimodal input like a video and the associated commen­tary, an intelligent agent can get a crude grounded model of linguistic concepts. This can only mean that an early human learner will be much better at this job. To summarize, the model has at its disposal, the grounded concepts of Entity/Object,<page local="3" global="117"/></p><doubt alpha="61.2" length="85" tooSmall="False" monospace="0.0">Substance, Container, CHASE(A,B), INTO(A,B), IN(A,B), OUT OF (A,B), MORE(A), LESS(A),</doubt><p>SLOW() and FAST(). However, these alone are insufficient to show how metaphors are acquired ('chase' verb is sparsely used in general litera­ture), and we therefore <i>assume </i>the availability of GIVE(AgentA, AgentB, Entity/Substance C), SOME(Substance A) and SPEND(Substance A) hereafter.</p></subsection><subsection number="2.2" title="Concept Acquisition"><p>The vast majority of our vocabularies are learned later purely from the linguistic in-put(Bloom(2000)). The goal from here-on will thus be to acquire language concepts with the help of the aforementioned grounded concepts and a text corpus. To determine how far language usage alone can help shape the concept of metaphors, we compiled a list of sentences from Lakoff and John-son(1980)) and Lakoff et al.(1991) that correspond to the metaphor-mappings for Containers, Objects and Substances. The salient findings are:</p><p>• Of the 85 sentences denoting Container metaphors,in 65, the abstract idea was im­parted the image schema of a container based only on the prepositions <i>in/out. </i>In the rest 20, adjectives <i>(full, empty) </i>and verbs <i>(explode, erupt, fill) </i>took the mantle.</p><p>• In all of the 63 sentences for Object metaphor, the Object property was imparted to the concept because VERB(A,B) took ob­ject arguments, i.e. verbs were the primary basis of metaphor mapping.</p><p>• Of the 42 sentences for Substance metaphors, 17 mappings were done based on adjectives <i>(more, less) </i>while the rest were of the type Container contains Substance, i.e. first the Container property was imparted, and then whatever was supposed to be inside the con­tainer was called a substance.</p><p>Based on this observation, we construct a model of concept acquisition which incorporates the fol­lowing bold (unproven) claims:</p><p><b>Claim 1 </b>Verbs, adjectives, nouns and preposi­tions all play roles in concept acquisition and have varying importance in different forms of concept acquisition</p><p><b>Claim 2 </b>There are only a limited representative verbs/adjectives that are grounded (i.e. have stable and distinct image schema), and the rest import their schemas and modify them to suit themselves. So, this structure is hier­archically organized.</p><p><b>Claim 3 </b>They can be represented in terms of a probability map, whence we can get an idea of the interplay between different concepts.</p><p><b>Claim 4 </b>Abstract concepts are acquired as con­crete concepts first, just like any other grounded concept, and later they acquire their abstraction due to emergence of future evi­dence.</p><p>Claim 1 is supported by the observation that precedes it. Claim 2 is somewhat self-explanatory, and better understood through examples. Con­sider verbs impart, provide, shower, bombard, donate etc. A close look will re­veal that they can all be derived from GIVE(). For instance, SHOWER() (as in 'shower some­body with praises') can be a combination of verb</p><p>GIVE() and adjective MORE(). In fact this kind of representation seems more memory-efficient. If we are required to store image schemas of the mil­lions of words that we come across, our memory will be a mess. Storing only a select few and com­bining them to derive the others is a more struc­tured, systematic and efficient mechanism. Fur­thermore, we should also notice that we had to take help of an adjective to describe a verb, which reinforces Claim 1. The first level of hierarchy consists of these grounded forms which are dis­tinguishable. The second level is the derived one that draws from all the nodes of the first level. Claim 3 asserts that this interplay can be repre­sented through a probability map. For instance, for a single verb GIVE(), the adjectives can be as­signed probabilities based on how frequently they modify GIVE() to produce an understandable im­age schema, so that we have an idea of which ones are more probable of appending to the verb when a new concept involving the same emerges. We will later see how this map can incorporate many aspects of the metaphor acquisition task.</p><subsubsection number="2.2.1" title="The Model"><p>We now describe the metaphor acquisition model based on Claim 4. We first have a repository of grounded concepts. Then as the learner is exposed to more sentences, the sentences are searched for contexts similar to the ones already learned.<page local="4" global="118"/> The noun arguments, which might be new concepts, are assigned a dynamic probability of belonging to one of the classes of Object/Substance/Container. With more evidence, these probabilities are mod­ified within a reward/penalty scenario. All con­cepts are treated concrete unless evidence to the contrary crops up.</p><p>The model is better understood through ex­amples. Let the learner come across the sentence - 'I can't give you much time' = GIVE(MUCH(time)). Now MUCH() takes a Substance as an argument. So time is assigned the schema of Substance with probability 1. Then GIVE(time), which takes either a Substance or an Object as its argument, dynamically changes the probabilities to 2/3 for Substance and 1/3 for Object. When it further comes across "In time, you will understand", i.e. IN(time), the probabil­ities are modified to 1/4, 2/4 and 1/4 for container, substance and object respectively. This assign­ment helps us in two ways - firstly, it prevents us from exclusively assigning the concept to any sin­gle class, thereby allowing us to model metaphors that contextually take up the properties of differ­ent classes. And secondly, it also gives us an idea of the affinity of the abstract concepts for different classes. To avoid confusion, as of now, from the above example, we can only assume that time is a concrete concept that has properties of Object, Container and Substance <footnote anchor="2"/>. <b>ACQUISITION: </b>We next tackle the problem of distinguishing between abstract and concrete con­cepts. Consider an early learner who comes across the following:</p><p>• Llama is a four-legged animal.</p><p>• Anger is a red-eyed demon.</p><p>How does the subject distinguish between con­crete Llama and abstract Anger? One way to look at it would be that as soon as the learner comes across these, it incorporates the features in the im­age schema of the concept. The schema field is then searched for possible conflicts. If two prop­erties are in conflict, they are brought to the CON­TEST field in the image schema, where 'voting' takes place between the conflicting scenarios. Vot­ing is done to take care of two possibilities. First, the conflict that arises might be due to a false evi­dence. If one of the properties is discarded based on some false evidence, the schema might become erroneous. So both are kept, but they are assigned probabilities of expected occurrence. Secondly, it gives room for metaphorical descriptions. For instance, suppose Anger has been assigned Ob­ject and Substance properties before this occur­rence. Now when it acquires the schema of 'red-eyed demon', the properties it assumes are, say, physical appearance and adjectives describing a demon. The conflict arises between physical ap­pearance and Substance (because a substance can't have eyes). So they are brought to the CONTEST field. Based on how often these concepts occur in the corpus, they are assigned probabilities. For ex­ample, in this case, the probability previously as­signed with Substance property is converted to the equivalent vote and red-eye is given vote 1. On next occurrence of 'anger is a substance', we fol­low a reward/penalty scenario <footnote anchor="3"/>, where the vote of Substance is increased by one, and that of 'red­eye' is reduced by one. When the vote of one concept is reduced to zero, the CONTEST field is cleared of the two. Assuming that false alarms are very less in number compared to correct usages, this process will reach stability. One might also note that this process doesn't in anyway harm the objective of the second sentence. The idea that it wanted to convey about Anger remains there inside Anger's image schema in the form of the adjectives - only the physical appearance schema is eliminated, which is the ideal behavior. Given enough time to settle down, the concrete concepts would thus have some sort of physical character­istic in their image schema that is NOT derived from Object/Substance/Container. The abstract concepts, on the other hand, would only be linked to the basic schemas, without any distinguishing and particular physical characteristics. For exam­ple, 'Box is a container' and 'Love is a Container'. Both will imbibe concepts of enclosure, bound­ary etc. from the container. But BOX would have additional schemas of a lid, wooden material etc. (which would actually vary subjectively). Since the concept of 'Love is a Container' is ingrained in the learner, we will say that the subject has <i>ac­quired </i>the metaphor.<page local="5" global="119"/> This method of metaphor ac­quisition eliminates any argument regarding a vi­olation of selectional restrictions and the need for basic seed metaphors to understand others. In this scenario, metaphors are learned like any other lin­guistic concept. The 'concrete first-then abstrac­tion' should thus score over 'abstract first-concrete grounding' approach.</p><footnote label="2">Which creates no conflict since Substances can act as agents/point-objects. Similarly containers and objects are in­terchangeable based on the context.</footnote><footnote label="3">It is to be noted that this scenario is only followed for schemas that are in the CONTEST class, not for all the schemas. This prevents unnecessary removal of non-conflicting schemas.</footnote><p><b>INTERPRETATION:   </b>This representation also helps in understanding of metaphor­ical occurrences. Previous works like KARMA(Narayanan(1997)), when they come across a metaphorical occurrence, search in a repository of metaphorical mappings to under­stand the statement. Understanding in the present system can naturally flow through the ECG approach. The ECG asserts that when we are interpreting 'Harry fell in water', we actually simulate Harry falling in water to understand the utterance. In the present approach, the occur­rence 'Harry fell in love', when simulated, will behave like this - the sentence will first be con­verted to concept FALL(Harry, love), then using FALL(Object,Container), HARRY and LOVE would import those schemas. HARRY would also import its own physical characteristic schemas while LOVE would have no such schema. Then motion schema of FALL() would be brought along and these three will be composited to produce the final simulation. Metaphorical mapping would thus be understood as any other linguistic concept, the only distinguishing factor would be that while concrete concepts would bring in their associated physical properties, abstract concepts will be described by a bare-bones schema. The idea of assigning probabilities of a concept's association to different base classes helps us in another elegant way. Previously, to understand 'Harry is in love' and 'Love led to his demise', the models had to invoke two different metaphorical mappings of 'Love is a container' and 'Love is an entity'. Wheres as, in the present model, the concept Love has already been assigned to both the Container and Object class, and based on the context one of the assignments gets highlighted. This reduces the memory inefficient and crude method of having a repository of metaphor-maps.</p><p><b>GENERATION: </b>Claim 3 helps us generate new metaphors or use established metaphors just as natural language is used, without conscious effort.</p><p>Once the system needs to convey an abstract idea, it has at its disposal a probability map through which the idea is connected to other concrete con­cepts. Those concepts are further connected to verbs/adjectives etc. with certain other probabili­ties. A path through this map, which can represent a coherent structure, would lead to a metaphor­ical mapping. Which metaphorical mapping is more culturally accepted would of course depend on the rating of the path (in this case, simply the multiplication of probabilities). This model how­ever would be extremely good in interpretation of newly created metaphors. The understanding process would just involve simulating the ideas by ECG. The older models, since they rely on a metaphor list, would have a hard time understand­ing new metaphors because they might not fit into the established scenarios.</p></subsubsection></subsection></section><section number="3" title="Experiments"><p>The Brown Corpus was used to test the ideas and derive some possible metaphorical mappings. All the occurrences of the grounded concepts, viz.</p><doubt alpha="57.9" length="38" tooSmall="False" monospace="0.0">CHASE(A,B), INTO(A,B), IN(A,B), OUT OF</doubt><p>(A,B), MORE(A), LESS(A), SLOW() and FAST() were found out and the sentence structure was con­verted to these functional structures using a very crude method - the first occurrence of a singu­lar or mass noun(NN) in the tagged corpus was assigned to the concept. For example, the sen­tence fragment 'into a hot cauldron' is converted to INTO(cauldron). Using this very basic method, some of the possible metaphor mappings that were found were:</p><p><b>Container </b>The following concepts were common a strong Container metaphorical map<i>-future battle fight mission darkness violence chaos silence water mind religion language</i> <b>Substance </b>The following concepts were common between GIVE() and MORE(), leading to a strong affinity for Substance mapping: <i>affec­tion information emphasis interest protection time</i></p><doubt alpha="64.3" length="42" tooSmall="False" monospace="0.0">between IN(), INTO(), OUT OF(), leading to</doubt><p>To have a flavor of how an abstract concept is connected to the base classes, we examined all oc­currences of noun LOVE in the corpus. While 90% of the time it acted like Object/Substance, 10% of the time, it acted like container, hinting that the affinity of Love for Container is mini­mal. <footnote anchor="4"/> The Object and Substance cases are al­most indistinguishable except for when substance-specific adjectives like 'more/less' are used.<page local="6" global="120"/> Oth­erwise, Love is considered a physical material, and it is not usually distinguishable whether its an Ob­ject or a Substance.</p><p>To look into how deeper mappings like 'Time is Money' might be deciphered, we also looked into SPEND(), and WASTE(). 60% of us­ages of SPEND() were SPEND(Time)(in various forms like day/year etc.), while the rest were SPEND(Money), with very minimal(two or three occurrences of the 200 odd) SPEND<i>(other sub­stance). </i>Similarly, of the 10 occurrences of verb WASTE(), 9 are concerning Time and the rest con­cerns Substance. The trend also points to one im­portant assumption we have made - that is, ab­stract concepts are first learned as concrete. As we see, these verb usages correspond to abstract concepts much more readily than they do to their concrete counterparts. So it's but natural on the part of an early learner to assume them as concrete ideas.</p></section><section number="4" title="Conclusion and Future Work"><p>While the above description pointed towards a new approach of handling metaphors that is closer in spirit to the view that metaphors are an inte­gral part of thought and language usage and not just poetic devices, the work might still look in­complete. This is so because even if the basic ideas and claims have been supported, the system is still not fully functional. To be precise, as of now we have the grounded concepts described in Section 2.1 and based on that extremely small test set, we have tried to learn some metaphorical map­pings. As more concepts are grounded in some way or other, the system will be better equipped to handle other mappings.</p><p>The ultimate aim would be to finally simulate all this in an ECG framework to show that the model is capable of emulating human behaviour. However, this must again be reiterated that the aim of this paper was not to show a working model fully capable of handling ontological metaphors, which is under construction, but that the new ap­proach might be better and more natural than pre­vious works that depended on hand-coded knowledge in some form or other. <b>Acknowledgement</b></p><footnote label="4">The container metaphor arose almost exclusively in the usage 'in love'.</footnote><p>I would like to thank Dr. Amitabha Mukerjee for his constant guidance throughout this work.</p></section><references><p>Bergen, B, N Chang, and S Narayan. 2004. Simu­lated action in an embodied construction grammar. In <i>Proc. of the 26th Annual Meeting of the Cognitive Science Society.</i></p><p>Bloom, Paul. 2000. <i>How Children Learn the Meanings of Words. </i>MIT Press, Cambridge, MA.</p><p>Fass, D. 1991. Met*: A method for discriminating metonymy and metaphor by computer. <i>Computa­tional Linguistics </i>17(1):49-90.</p><p>Kintsch, W. 2000. Metaphor comprehension: A com­putational theory. <i>Psychonomic Bulletin and Review </i>pages 257-266.</p><p>Lakoff, George, Jane Espenson, and Alan Schwartz. 1991. <i>Master metaphor list: 2nd draft copy.</i></p><p>Lakoff, George and Mark Johnson, editors. 1980. <i>Metaphors We Live By. </i>University of Chicago Press.</p><p>Langacker, RW, editor. 1987. <i>Foundations ofCogni-tive Grammar I: Theoretical Prerequisites. </i>Stanford University Press.</p><p>Mukerjee, Amit, Kruti Neema, and Sushobhan Nayak. 2011. Discovering coreference using image-grounded verb models. In <i>Proceedings ofRANLP.</i></p><p>Mukerjee, Amitabha and Mausoom Sarkar. 2007. Grounded perceptual schemas: Developmental ac­quisition of spatial concepts. In <i>Spatial Cognition V Reasoning, Action, Interaction, </i>Springer Berlin / Heidelberg, volume 4387, pages 210-228.</p><p>Narayanan, S. 1997. <i>Knowledge-based Action Rep­resentations for Metaphor and Aspect (KARMA).</i></p><p>Ph.D. thesis, CS, UC Berkeley.</p><p>Narayanan, S. 1999. Moving right along: A computa­tional model of metaphoric reasoning about events. In <i>Proc. ofNCAI. </i>pages 121-129.</p><p>Roy, Deb and Ehud Reiter. 2005. Connecting language to the world. <i>Artificial Intelligence: Special Issue on Connecting Language to the World </i>167:112.</p><p>Shutova, E, L Sun, and A Korhonen. 2010. Metaphor identification using verb and noun clustering. In <i>Proc. ofCOLING. </i>pages 1002-1010.</p><p>Wilks, Y. 1975. A preferential pattern seeking seman­tics for natural language inference. <i>Artificial Intelli­gence </i>6(1):53-74.</p></references></body></article>