<?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="980"/><title>Unsupervised Learning of Selectional Restrictions and Detection of Argument Coercions</title><pubinfo>Proceedings of the 2011 Conference on Empirical Methods in Natural Language Processing,pages 980-990, Edinburgh, Scotland, UK, July 27-31, 2011. ©2011 Association for Computational Linguistics</pubinfo><author surname="Roberts" givenname="Kirk"><org  name="Duke University" country="USA" city="Durham"/></author><author surname="Harabagiu" givenname="Sanda"><org  name="University of Texas at Dallas" country="USA" city="Richardson"/></author></firstpageheader><frontmatter><p><b>Unsupervised Learning of Selectional Restrictions and Detection of</b></p><p><b>Argument Coercions</b></p><p><b>Kirk Roberts and Sanda M. Harabagiu</b></p><p>Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083, USA (kirk,sanda}@hlt.utdallas.edu</p></frontmatter><abstract>Metonymic language is a pervasive phe­nomenon. Metonymic type shifting, or ar­gument <i>type coercion, </i>results in a selectional restriction violation where the argument's se­mantic class differs from the class the predi­cate expects. In this paper we present an un-supervised method that learns the selectional restriction of arguments and enables the de­tection of argument coercion. This method also generates an enhanced probabilistic reso­lution of logical metonymies. The experimen­tal results indicate substantial improvements the detection of coercions and the ranking of metonymic interpretations. </abstract></header><body><section number="1" title="Introduction"><p>Metonymic language is pervasive in today's social interactions. For example, it is typical to find ques­tions that require metonymic resolution:</p><p>(Q1) Did you enjoy War and Peace? (Q2) Does anyone have any advice on how to start a bowling team?<footnote anchor="1"/></p><p>In order to process such questions and capture the intention of the person that posed them, coercions are needed. Question (Q1) is interpreted as whether you enjoyed <i>reading </i>"War and Peace", while (Q2) is interpreted as asking for advice on <i>organizing, forming, </i>or <i>registering </i>a bowling team. The qual­ity of the answers therefore depends on the ability to (1) recognize when metonymic language is used, and (2) to produce coercions that capture the user's intention. One important step in this direction was taken by SemEval-2010 Task 7, which focused on the ability to recognize (a) an argument's selectional restriction for predicates such as <i>arrive at, cancel, </i>or <i>hear, </i>and (b) the type of coercion that licensed a correct interpretation of the metonymy. Details of the task are reported in (Pustejovsky et al., 2010). Approaches to metonymy based on this task are lim­ited, however, because (a) the task is focused only on semantically non-ambiguous predicates and (b) the selectional restrictions of the arguments were cho­sen from a pre-defined set of six semantic classes (artifact, document, event, location, proposition, and sound). However, metonymy coercion systems ca­pable of providing the interpretations of questions (Q1) and (Q2) clearly cannot operate with the sim­plifications designed for this task.</p><footnote label="1">Both questions taken from Yahoo Answers.</footnote><p>Inspired by recent advances in modeling selec-tional preferences with latent-variable models (Rit­ter et al., 2010; O Seaghdha, 2010), we propose an unsupervised model for learning selectional re­strictions. The model assumes that (1) arguments have a single selected class exemplified by the se-lectional restriction, and (2) the selected class can be inferred from the data, in part by modeling how coercive each predicate is. The model is capable of operating with both ambiguous and disambiguated predicates, producing superior results for predicates that have been disambiguated. The selectional re­strictions and coercions detected by the model re­ported in this paper can be used to enhance the logi­cal metonymy approach reported in Lapata and Las-carides (2003). The experimental results show a sig­nificant improvement in the ranking of interpreta­tions.</p><page local="2" global="981"/><p>The rest of this paper is organized as follows. Section 2 discusses related work. Section 3 de­tails unsupervised models that inform detection of metonymies. Section 4 outlines a method for disam­biguating ambiguous predicates. Section 5 describes the enhanced interpretation of logical metonymies when conventional constraints are known. Section 6 outlines our implementation and experimental de­sign. Section 7 presents our experimental results in three broad tasks: (i) semantic class induction, (ii) coercion detection, and (iii) logical metonymy inter­pretation. Section 8 summarizes the conclusions.</p></section><section number="2" title="Previous Work"><p>Lapata and Lascarides (2003) propose a probabilis­tic ranking model for logical metonymies. They es­timate these probabilities using co-occurrence fre­quencies of predicate-argument pairs in a corpus. Shutova (2009) extends this approach to provide sense-disambiguated interpretations from WordNet (Fellbaum, 1998) by using the alternative interpre­tations to disambiguate polysemous words. Shutova and Teufel (2009) extend this approach further by clustering these sense-disambiguated interpretations into distinct groups of meaning (e.g., <i>(read, browse, look through} </i>and <i>(write, produce, work on} </i>for <i>"enjoy book"). </i>Not only do these approaches as­sume logical metonymies have already been iden­tified, but they are susceptible to providing interpre­tations that are themselves logical metonymies (e.g., <i>finish book). </i>In this paper, we propose an enhance­ment to resolving logical metonymies by ruling out event-invoking predicates in order to provide more semantically valid interpretations.</p><p>Recently, the resolution of several linguistic prob­lems has benefited from Latent Dirichlet Alloca­tion (LDA) (Blei et al., 2003) models. O Seaghdha (2010) examines several selectional preference mod­els based on LDA in predicting human judgements on predicate-argument plausibility. Both LDA and an extension, ROOTH-LDA (based on Rooth et al. (1999)), perform well at predicting plausibility on unseen predicate-argument pairs. Inspired by these results, we propose to extend selectional preference models in order to learn selectional restrictions.</p><p>Alternatively, unsupervised algorithms exist that both induce semantic classes (Rooth et al., 1999; Lin and Pantel, 2001) and cluster predicates by their selectional restrictions (Rumshisky et al., 2007) but none of these provide a sufficient framework for de­termining if a specific argument violates its predi­cate's selectional restriction.</p></section><section number="3" title="Unsupervised Learning of Selectional Restrictions"><p>In predicate-argument structures, predicates impose selectional restrictions in the form of semantic ex­pectations on their arguments. Whenever the seman­tic class of the argument meets these constraints a <i>selection </i>occurs. For example, the predicate <i>"hear" </i>imposes the semantics related to sound on the ar­gument <i>"voice". </i>Because the semantic class for <i>"voice" </i>conforms to these constraints, we call its se­mantic class the <i>selected class. </i>However, when the semantic class of the argument violates these con­straints, we follow Pustejovsky et al. (2010) and re­fer to this as a <i>coercion. </i>In this case, we call the argument's semantic class the <i>coerced class. </i>For ex­ample, <i>"hear speaker" </i>is a coercion where the ar­gument class, person, is implicitly coerced into the voice of the speaker, a sound.</p><subsection number="3.1" title="A Baseline Model"><p>We consider the LDA-based selectional prefer­ence model reported in O Seaghdha (2010) as a baseline for modeling selectional restrictions. For­mally, we define our LDA baseline model as follows. Let V be the predicate vocabulary size, let A be the argument vocabulary size, and let K be the number of argument classes. Let a" be the ith (non-unique) argument realized by predicate v. Let cv be the class for a". Let 6" be the class distribution for predicate v and ffk be the argument distribution for class k. The graphical model for this LDA is shown in Fig­ure 1(a). The generative process for LDA is:</p><doubt alpha="64.7" length="34" tooSmall="False" monospace="0.0">For each argument class k = 1..K :</doubt></subsection></section><section number="1." title="Choose (f k ~ Dirichlet(ß)"><p>For each unique predicate <i>v </i>= 1<i>..</i><i>V:</i></p><doubt alpha="64.0" length="25" tooSmall="False" monospace="0.0">2. Choose6"~ Dirichlet(a)</doubt><doubt alpha="64.3" length="28" tooSmall="False" monospace="0.0">For every argument i =1..n":</doubt><doubt alpha="62.1" length="29" tooSmall="False" monospace="0.0">3. Choosec"~ Multinomial(6" )</doubt></section><section number="4." title='Choose a" ~ Multinomial(f c i )'><p>Following Griffiths and Steyvers (2004), we col­lapse 6 and ff and estimate the model using Gibbs<page local="3" global="982"/></p><p>For each unique predicate v <i>= 1</i><i>..V</i>:</p><doubt alpha="0.0" length="1" tooSmall="False" monospace="0.0">0</doubt><doubt alpha="28.6" length="7" tooSmall="False" monospace="0.0">(a) (b)</doubt><figure caption="Figure 1: Graphical models for (a) LDA, and (b) coercion LDA (cLDA)."></figure><p>Sampling. This yields the update equation:</p><doubt alpha="66.7" length="12" tooSmall="False" monospace="0.0">fvk+ a fak+ß</doubt><doubt alpha="52.9" length="17" tooSmall="False" monospace="0.0">p(cv= k\av;a, ß)a</doubt><p>fv + <i>Kafk</i><i> </i><i>+</i><i> Aß</i></p><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(1)</doubt><p>Where <i>fak </i>is the frequency of argument <i>a </i>being as­signed class k; <i>fk</i><i> </i>is the frequency of class <i>k</i><i> </i>being assigned to any argument; <i>f</i><i>vk </i>is the frequency of predicate v having an argument of class k; and fv is the total number of non-unique arguments for pred­icate <i>v.</i></p><subsection number="3.2" title="A Coercion Model"><p>We now incorporate our assumptions for selec-tional restriction modeling. Namely: (1) there is one selected class per predicate, and (2) the predi­cate's selected class can be chosen from the classes of its arguments. To accomplish this, we must also account for the coerciveness of each predicate. We assign a latent variable t<i>v </i>for each predicate v that controls how coercive <i>v </i>should be. The additional hyper-parameters 70 and 71 act as priors on tv. The generative process for this coercion LDA model, which we denote cLDA, is:</p><p>For each argument class k <b>= 1 </b>..K:</p><p>1. Choose <i>4&gt;k</i><i> </i><i>-</i><i> </i>Dirichlet(ß) 3. Choose 9v <i>- </i>Dirichlet(a)<footnote anchor="2"/></p><doubt alpha="66.7" length="24" tooSmall="False" monospace="0.0">2. Choosesv-Uniform(1,K)</doubt><doubt alpha="56.5" length="23" tooSmall="False" monospace="0.0">4. Choosetv-Beta(7o,71)</doubt><p>For every argument <i>i</i><i> </i><i>= 1</i><i>..nv</i>:</p><p>5. Choose cv - Multinomial(9v) 6. Choose xv - Bernoulli(Tv)</p><p>7. If xv <b>= 1, </b>Choose <i>av</i><i> </i>- Multinomial^<footnote anchor="0"/>") Else Choose av - Multinomial(&lt;/&gt;s" )</p><p>The model variable sv represents the selected class for predicate v.  The coerced class is represented for each argument i by cv, where xv chooses be­tween the selected and coerced class. The variable xv is similar to switching variables in other graph­ical models such as Chemudugunta et al. (2007) and Reisinger and Mooney (2010), where switch­ing variables are used to choose between a back­ground distribution and a document-specific distri­bution. In this case, the switching variable chooses between a specific class and a predicate-specific dis­tribution. The graphical model for cLDA is shown in Figure 1(b). Note that cLDA is virtually equiva­lent to LDA when tv is 1 and 71 is small because the selected class will be ignored. In this way, highly co­ercive predicates have less of an impact on the argu­ment clustering because they are more reliant on the multinomial 9. We use Gibbs sampling to perform model inference and collapse 9, 0, and t, integrat­ing them out using multinomial-Dirichlet conjugacy (the Beta distribution used by t is just a special case of the Dirichlet with only two parameters). The update formula for the selected class sv is:</p><footnote label="2">With the exception that the probability of drawing the se­lected class s v is zero. This can be seen as drawing the multi­nomial 6 v  from a Dirichlet distribution with K-1 components.</footnote><doubt alpha="60.0" length="20" tooSmall="False" monospace="0.0">p(sv=k\av,cv,xv;a,ß)</doubt><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">nv</doubt><doubt alpha="50.0" length="16" tooSmall="False" monospace="0.0">l\P(av\sv= k; ß)</doubt><doubt alpha="62.5" length="8" tooSmall="False" monospace="0.0">favk + ß</doubt><doubt alpha="100.0" length="2" tooSmall="False" monospace="0.0">OC</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(2)</doubt><p>Where nv is the number of argument observations for predicate v; <i>Sv </i>is the set of arguments of v that are selections; and fav k is the frequency of word avbeing assigned to class k for any predicate. We then sample cv and xv jointly:</p><doubt alpha="52.0" length="75" tooSmall="False" monospace="0.0">p(cv=k,xv=q\sv,c-v,x-v,av;a, ß, 7)a p(cv=k;a)p(xv=q; 7)p(av\sv,cv,x-v,av;ß)</doubt><doubt alpha="63.2" length="19" tooSmall="False" monospace="0.0">fvk +afvq +7qfaz +ß</doubt><doubt alpha="53.8" length="26" tooSmall="False" monospace="0.0">fv +Ka fvo+70+fvi+71fz +Aß</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(3)</doubt><page local="4" global="983"/><p>Where <i>fvq</i>, fv0, and fv1 is the frequency of x values that equal q, 0, and 1, respectively, for predicate v; <i>faz </i>is the frequency of word <i>a </i>being in class z and <i>fz</i><i> </i>is the frequency all words being in class z, where z is defined as being equal to k when xv = 1, or svwhen xv = 0.</p><p>Note that Equation (2) results in a sampling of the selected class for v proportional to the number of arguments in each class for v, fulfilling our sec­ond assumption. Also note from Equation (3), the second term corresponds to the coerciveness of the predicate. When the predicate is very coercive, the marginal probability associated with xv = 0 will be very low. If all predicates become entirely coercive, most x values will become 1 and the cLDA will be­come almost equivalent to an LDA model.</p></subsection><subsection number="3.3" title="Coercion Detection"><p>After the latent parameters have been estimated, we still require a method to determine if a given predicate-argument pair is a coercion or not. We assign a score in [0,1] instead of a binary value. Higher scores (near 1) indicate high likelihood of selection, while lower scores (near 0) indicate coer­cion. The LDA model must rely on a scoring method using the predicate-class and argument-class mix­tures:</p><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">K</doubt><doubt alpha="50.0" length="24" tooSmall="False" monospace="0.0">Ci(v,a) =P(k\v)P(a\k)k K</doubt><doubt alpha="100.0" length="1" tooSmall="False" monospace="0.0">k</doubt><p>Where 9kv represents the probability ofany argument of v being in the class k and <i>&lt;\</i>)ka represents the prob­ability of the argument a being in class k for any predicate. <i>C1 </i>is also available as a scoring method for cLDA by including the proportion ofthe selected class sv in 9. Note that since 9 and &lt; are integrated out for both LDA and cLDA, we instead use their frequencies smoothed with a and ß, respectively, which is their maximum likelihood estimate.</p><p>The cLDA model contains two useful parameters that can identify selections and coercions: the se­lected class s and the coercion indicator x. This yields two more coercion scoring metrics:</p><doubt alpha="38.9" length="18" tooSmall="False" monospace="0.0">C2(v,a) = P (a\sv)</doubt><doubt alpha="25.0" length="8" tooSmall="False" monospace="0.0">= &lt;av(5)</doubt><doubt alpha="37.5" length="24" tooSmall="False" monospace="0.0">Cs(v,a) = P (xa = 0\v,a)</doubt><doubt alpha="60.0" length="5" tooSmall="False" monospace="0.0">I vv'</doubt><doubt alpha="16.7" length="6" tooSmall="False" monospace="0.0">\1- a\</doubt><p>Where sv is the selected class for predicate v; /a is the set of predicate-argument instances for pred­icate v and argument a; and xv is 0 for a selection and 1 for a coercion. Of the three metrics, <i>C3 </i>is the most direct measure of a coercion as it represents the average decision the model learned on the same predicate-argument pair. However, <i>C3 </i>requires a large sample of instances for a particular predicate and argument, and so may be quite sparse. In prac­tice, these different metrics have their own strengths and weaknesses and the best performing method of­ten depends on the final task.</p></subsection></section><section number="4" title="Predicate Sense Induction"><p>Our assumption of a single selected class per predi­cate ignores predicate polysemy. However, the same lexical item may have multiple meanings, each with a separate selected class. We therefore propose a method of partitioning a predicate's arguments by the induced senses of the predicate. This allows sep­arate induced predicates to each select a separate ar­gument class. Consider the verb <i>fire, </i>which has at least two distinct common senses: (1) to shoot or propel an object (e.g., to fire a gun), and (2) to lay someone off (e.g., to fire an employee). The first sense selects a weapon (e.g., gun, bullet, rocket), while the second sense selects a person (e.g., em­ployee, coach, apprentice).</p><p>Specifically, we employ tiered clustering (Reisinger and Mooney, 2010) using the words in the predicate's context. Tiered clustering is a discrete clustering method, as opposed to methods such as (Brody and Lapata, 2009) that assign a distribution of word senses to each word instance. Tiered clustering has several advantages over other discrete clustering approaches. First, tiered clustering learns a background word distribution in addition to the clusters. This reduces the impact that words common to most senses have on the cluster­ing process and allow clusters to form around only the most salient words.  Second, tiered clustering<page local="5" global="984"/></p><p>Table 1: Context word clusters resulting from tiered clus­tering for the verb <i>fire </i>(includes the number of unique words belonging to each cluster).</p><p>uses a Chinese Restaurant Process (CRP) prior to control both the formation of new clusters (senses) and the bias toward larger clusters (more common senses). This conforms with our intuition of how word senses are distributed: a few common senses with a gradual transition to a long tail of rare senses. When deciding which cluster to use for a given predicate-argument pair, we use the cluster most associated with the argument.</p><p>We use a 10-token window around the predicate as features. The result of predicate induction on the verb <i>fire </i>is shown in Table 4. The first three clusters can be interpreted to be about (1) firing weapons, (2) sport-related shots (e.g., <i>"fired the puck"), </i>and (3) lay-offs. One must be careful in choosing the param­eters for induction, however, as it is possible to par­tition a unique word sense such that coercions and selections are placed in a separate clusters. Section 6 discusses our parameter selection experiments.</p></section><section number="5" title="Logical Metonymy Interpretation"><p>Logical metonymies are a unique class of coercions due to the fact that their eventive interpretation can be derived from verbal predicates. For instance, for the logical metonymy <i>"enjoy book", </i>we know that <i>read </i>is a good candidate interpretation because (1) books are objects whose purpose is to be read and (2) reading is an event that may be enjoyed. We therefore expect to see many instances of both <i>"read book" </i>and <i>"enjoy reading"" </i>(Lapata and Lascarides, 2003). Conversely, for coercions with non-eventive interpretations, such as <i>"arrive at meeting"", </i>the in­terpretation <i>(location of) </i>is more dependent on the predicate <i>(arrive) </i>than the function of its argument <i>(meeting).</i></p><p>In this section, we limit our discussion of logical metonymy to the verb-object case, its correspond­ing baseline for ranking interpretations, and our pro­posed enhancements. However, similar baselines exist for other types of logical metonymy, such as adjective-noun and noun-noun. Since our enhance­ment does not depend on any syntactic information beyond the predicate-argument instances needed for Section 3.2, it could easily be applied to those as well.</p><p>Lapata and Lascarides (2003) propose a proba­bilistic ranking model where the probability of an interpretation e for a verb-object pair (v, o) is pro­portional to the probability of all three in a verb-interpretation-object pattern.<footnote anchor="3"/> For example, the probability that <i>read </i>is the correct interpretation of <i>"enjoybook" </i>is proportional to the likelihood ofsee-ing <i>"enjoy reading book" </i>expressed as a syntactic dependency in a sufficiently large corpus. Due to data sparsity, they approximate this likelihood of seeing the object given the verb and interpretation to simply the likelihood of seeing the object given the interpretation. We denote this logical metonymy ranking method as LMLL, formally defined as:</p><doubt alpha="58.3" length="12" tooSmall="False" monospace="0.0">LMLL(e; v,o)</doubt><doubt alpha="45.5" length="11" tooSmall="False" monospace="0.0">Pc(v, e, o)</doubt><doubt alpha="56.5" length="62" tooSmall="False" monospace="0.0">Pc(e)Pc(v\e)Pc(o\e,v) Pc(e)Pc(v\e)Pc (o\e)fc(v,e)fc(o,e)Nfc(e)</doubt><doubt alpha="0.0" length="3" tooSmall="False" monospace="0.0">(7)</doubt><p>Where Pc and fc indicate probability and frequency, respectively, derived from corpus counts. See Lap-ata and Lascarides (2003) for a detailed explanation of how these frequencies are obtained.</p><p>This model, which we consider our baseline, is only partially correct as the corpus will contain co­ercions that form invalid interpretations. Consider the phrases <i>"enjoy finishing a book"" </i>and <i>"enjoy discussing a book"". </i>Both <i>"finish book"" </i>and <i>"dis­cuss book" </i>are coercions (and logical metonymies) themselves, and do not form a valid interpretation.<footnote anchor="4"/><page local="6" global="985"/></p><footnote label="3">They use two patterns: "v e-ing o" and "v to e o", where e is tagged as a verb.</footnote><footnote label="4">For evidence of the frequency of these phrases, at the time of this writing, "enjoy finishing a book" and "enjoy finishing the book" have a combined 728 Google hits, while "enjoy dis­cussing a book" and "enjoy discussing the book" have a com-</footnote><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Cluster 1</p></td><td class="cell"><p>Cluster 2</p></td><td class="cell"><p>Cluster 3</p></td><td class="cell"><p>Cluster 4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>(18,391)</p></td><td class="cell"><p>(16,651)</p></td><td class="cell"><p>(18,749)</p></td><td class="cell"><p>(11,833)</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>shots</p></td><td class="cell"><p>ball</p></td><td class="cell"><p>hire</p></td><td class="cell"><p>gun</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>gun</p></td><td class="cell"><p>puck</p></td><td class="cell"><p>letter</p></td><td class="cell"><p>imagination</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Israeli</p></td><td class="cell"><p>hired</p></td><td class="cell"><p>Yeltsin</p></td><td class="cell"><p>grill</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>missiles</p></td><td class="cell"><p>owner</p></td><td class="cell"><p>minister</p></td><td class="cell"><p>laser</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>rockets</p></td><td class="cell"><p>shots</p></td><td class="cell"><p>workforce</p></td><td class="cell"><p>cells</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>officers</p></td><td class="cell"><p>coaches</p></td><td class="cell"><p>executives</p></td><td class="cell"><p>engine</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>soldiers</p></td><td class="cell"><p>net</p></td><td class="cell"><p>employee</p></td><td class="cell"><p>brain</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>rounds</p></td><td class="cell"><p>circle</p></td><td class="cell"><p>managers</p></td><td class="cell"><p>I</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>bullets</p></td><td class="cell"><p>Johnson</p></td><td class="cell"><p>hired</p></td><td class="cell"><p>engines</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>weapons</p></td><td class="cell"><p>Williams</p></td><td class="cell"><p>union</p></td><td class="cell"><p>fire</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>Thus, when discovering interpretations for logical metonymies, we must be aware ofthe selectional re­strictions of candidate interpretations.</p><p>We propose to incorporate the coercion probabil­ity learned by our cLDA model in order to rank only those interpretations that are considered selections:</p><doubt alpha="33.3" length="33" tooSmall="False" monospace="0.0">LM ' (e;v,o)=P (v,e,o,xO = 0) (8)</doubt><p>However, due to the approximations made to esti­mate Pc(v, e, o), this probability cannot be directly calculated as not all the frequencies reflect verb-object counts. Instead, we can combine the corpus probability Pc(v, e, o) with the probability that the verb-object pair (e, o) is a coercion in our model. We denote this probability <i>Px</i><i> </i>(e, o), and it may be derived from the scoring metrics in Equations (4), (5), or (6) above. We further propose three methods for enhancing the LMLL baseline using Px (e, o) to approximate Equation 8.</p><p>A naive method for including information from our cLDA model is to consider the corpus prob­ability, Pc(v, e, o) and the coercion probability, Px(e, o), to be independent:</p><doubt alpha="47.2" length="36" tooSmall="False" monospace="0.0">LMind(e; v,o) =Pc(v,e,o)Px (e,o) (9)</doubt><p>In other words, the rank of an interpretation is dic­tated by the unweighted combination of its corpus probability Pc and its coercion probability Px. How­ever, these two quantities are not likely to be inde­pendent. Most instances where e is used with either v or o are in fact selective.<footnote anchor="5"/> We therefore experiment with two shallow learning methods for combining these two quantities.</p><p>The first method is a filtering approach where a threshold is learned for Px:</p><doubt alpha="35.7" length="28" tooSmall="False" monospace="0.0">lm™(e;v,o)={pc(v0e,o)^-5(10)</doubt><p>Where the threshold 5 is learned from a development set. We expect this model could suffer from noisy Px values or to simply choose a threshold ofzero due to the prominence of Pc.</p><p>Finally, we include a weighted linear model to</p><p>bined 7,040 Google hits.</p><footnote label="5">For comparison, "enjoy reading a book" and "enjoy read­ing the book" have a combined 6.5 million Google hits</footnote><p>discover the relative value of Pc and Px:</p><doubt alpha="34.1" length="44" tooSmall="False" monospace="0.0">LMwt(e; v,o) =       (v,e, o)+W2Px(e,o) (11)</doubt><p>Where <i>w1 </i>and <i>w2 </i>are learned weights. We dis­cuss how the parameters for LMTH and LMWT are learned in the experimental setup below.</p></section><section number="6" title="Experimental Setup"><p>We use the NYT subsection of the English Gigaword Fourth Edition (Parker et al., 2009) for a total of 1.8M newswire articles. The Stanford Dependency Parser (de Marneffe et al., 2006) is used to extract verb-object relations <i>(dobj) </i>that form the input to our model. To reduce noise, we keep only verbs listed in VerbNet (Kipper et al., 1998) with at least 100 ar­gument instances, discarding <i>have </i>and <i>say, </i>which are too semantically flexible to select from clear se­mantic classes and so common they distort the class distributions. This results in 4,145 unique verbs with 51M argument instances (388K unique arguments). Additionally we use the dependency parser to ex­tract open clausal complements of verbs (e.g., <i>"like to swim") </i>for use in logical metonymy interpreta­tion. We believe this to be a more reliable alter­native to the phrase chunk extraction patterns used in Lapata and Lascarides (2003). We keep clausal complements <i>(xcomp) </i>where the dependent is either a gerund or infinitive in order to estimate Pc(v\e) in Equation (7).</p><p>For tiered clustering we use the same implemen­tation as Reisinger and Mooney (2010)<footnote anchor="6"/> to partition the surface form of the verb into one or more in­duced forms. Instead of using a fixed number of iterations, the clustering was run for 100 iterations past the best recorded log-likelihood in order to find the best possible fit to the data. We tuned the hyper­parameters by maximizing the log-likelihood on a small held-out set of 20 predicate-argument pairs (10 selections, 10 coercions). The resulting parti­tions were fairly conservative, yielding 12,332 in­duced verbs or about 3 induced verb forms for every surface form, with 305 verbs not being partitioned at all.</p><p>We implemented both LDA and cLDA as de­scribed in Sections 3.1 and 3.2. For the a and ß hyper-parameters, we used the MALLET (McCal-lum, 2002) defaults of 1.<page local="7" global="986"/>0 and 0.1, respectively, for both LDA and cLDA. We used the 20 predicate-argument pairs mentioned above to tune the 7 hyper­parameters as well as the number of iterations. Both 70 and 71 were set to 100. We observed that for both LDA and cLDA, longer runs (in iterations) re­sulted in improved model log-likelihood but infe­rior results in terms of detecting coercions. It is not uncommon in topic modeling for model likeli­hood to not be completely correlated with the score on the task for which the topic model was intended (see Chang et al. (2009)). Both LDA and cLDA were found to perform best at 50 iterations on this data, after which their class distributions were less "smooth" and became rigidly associated with just a few classes, thus having a negative impact on coer­cion detection. While further iterations hurt coer­cion detection, only minor gains in model likelihood are seen. We believe the small number of iterations necessary for the model to converge is therefore a function of the data. In traditional topic modeling, documents are generally of similar size (i.e., within an order of magnitude). But in our data, many pred­icates have 10,000 times more instances than others. We have not yet empirically explored the impact of using a more uniform number of arguments for each predicate. This issue also makes it difficult to take multiple samples, which we experimented with un­successfully.</p><footnote label="6">Available at http://github.com/joeraii/UTML-Latent- Variable-Modeling-Toolkit</footnote><p>Our a priori intuition was that as the number of classes was increased, LDA would improve and cLDA would degrade due to its assumption of a sin­gle selected class. However, this did not always bear out in the results for every task described below. As such, instead of choosing a specific number of classes for each model, we describe results for each model with <i>K = </i>10, 25, and 50.</p><p>For logical metonymy, both LMth and LMwtrequire learned parameters. LMth needs a learned threshold while LMwt needs two learned weights. For both, we split the data set into two partitions, learn the optimal threshold/weights on one partition, and use it as the parameters for the other partition. Both methods are trained on the final scoring metric, described in Section 7.3. For threshold learning, this involves finding the optimal cut-off to maximize the score. For weight learning, we use an exhaustive search over the range {1.0, 0.9, ..., 0.2, 0.1, 10-2, 10-3, ..., 10-14 } for both <i>wi </i>and <i>w2</i>.</p></section><section number="7" title="Results and Discussion"><subsection number="7.1" title="Semantic Class Induction"><p>For the evaluation of the argument classes in­duced by our method, we use a subset of the Word­Net lexicographer files, which correspond to coarse­grained semantic classes. We chose this form of evaluation because, unlike a named entity corpus, no sentential context is required and is therefore more consistent with the information available to our model. We use six of the larger, more seman-tically coherent WordNet classes: artifact, person, plant, animal, location, and food. We consider each of these a cluster and compare them to clusters com­posed of the top ten non-polysemous words (accord­ing to WordNet) in each of the classes generated by both the baseline (LDA) and our model (cLDA). Words not in both sets of clusters are removed. The result of this evaluation, compared with six cluster­ing metrics, is shown in Table 2. For descriptions of NMI, Rand, and cluster F-measure, see Manning et al. (2008); for the B<footnote anchor="3"/> metrics (Cluster and Element), see Bagga and Baldwin (1998); for the MUC met­ric, see Vilain et al. (1995). Each metric has differ­ent strengths and biases in regards to the number and distribution of clusters, so all are provided to give a general picture of class induction performance.</p><p>The best performing model on all metrics is cLDA with induced predicates using 10 classes. However, as the number of classes is increased and the gran­ularity of the induced classes becomes more fine­grained, LDA (predictably) outperforms cLDA on most metrics. This is consistent with our intuition is a direct result of the sampling for the predicate-argument pair in question and can thus be expected to perform poorly on rare predicate-argument pairs.<page local="8" global="987"/> Given that many of the arguments in this data are rare or unseen in the Gigaword data (e.g., <i>"cancel Renaulf"), </i><i>C</i><i>3"s </i>poor performance is understandable.</p><table caption="Table 2: Clustering scores for induced classes." class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>induced predicates?</p></td><td class="cell"><p>N</p></td><td class="cell"><p>Y</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p># classes</p></td><td class="cell"><p>10</p></td><td class="cell"><p>25</p></td><td class="cell"><p>50</p></td><td class="cell"><p>10</p></td><td class="cell"><p>25</p></td><td class="cell"><p>50</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LDA</p></td><td class="cell"><p>NMI</p></td><td class="cell"><p>.382</p></td><td class="cell"><p>.448</p></td><td class="cell"><p>.389</p></td><td class="cell"><p>.435</p></td><td class="cell"><p>.391</p></td><td class="cell"><p>.383</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Rand</p></td><td class="cell"><p>.717</p></td><td class="cell"><p>.731</p></td><td class="cell"><p>.721</p></td><td class="cell"><p>.760</p></td><td class="cell"><p>.723</p></td><td class="cell"><p>.730</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Fl</p></td><td class="cell"><p>.425</p></td><td class="cell"><p>.319</p></td><td class="cell"><p>.192</p></td><td class="cell"><p>.543</p></td><td class="cell"><p>.311</p></td><td class="cell"><p>.205</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Ba (C)</p></td><td class="cell"><p>.553</p></td><td class="cell"><p>.513</p></td><td class="cell"><p>.444</p></td><td class="cell"><p>.525</p></td><td class="cell"><p>.476</p></td><td class="cell"><p>.341</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Bd (E)</p></td><td class="cell"><p>.453</p></td><td class="cell"><p>.351</p></td><td class="cell"><p>.223</p></td><td class="cell"><p>.521</p></td><td class="cell"><p>.324</p></td><td class="cell"><p>.234</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MUC</p></td><td class="cell"><p>.545</p></td><td class="cell"><p>.545</p></td><td class="cell"><p>.531</p></td><td class="cell"><p>.500</p></td><td class="cell"><p>.532</p></td><td class="cell"><p>.544</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p>NMI</p></td><td class="cell"><p>.446</p></td><td class="cell"><p>.403</p></td><td class="cell"><p>.360</p></td><td class="cell"><p><b>.510</b></p></td><td class="cell"><p>.430</p></td><td class="cell"><p>.366</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Rand</p></td><td class="cell"><p>.736</p></td><td class="cell"><p>.719</p></td><td class="cell"><p>.716</p></td><td class="cell"><p><b>.788</b></p></td><td class="cell"><p>.734</p></td><td class="cell"><p>.711</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Fl</p></td><td class="cell"><p>.448</p></td><td class="cell"><p>.291</p></td><td class="cell"><p>.183</p></td><td class="cell"><p><b>.567</b></p></td><td class="cell"><p>.329</p></td><td class="cell"><p>.184</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Bd (C)</p></td><td class="cell"><p>.575</p></td><td class="cell"><p>.484</p></td><td class="cell"><p>.312</p></td><td class="cell"><p><b>.593</b></p></td><td class="cell"><p>.495</p></td><td class="cell"><p>.313</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Ba (E)</p></td><td class="cell"><p>.473</p></td><td class="cell"><p>.321</p></td><td class="cell"><p>.205</p></td><td class="cell"><p><b>.556</b></p></td><td class="cell"><p>.346</p></td><td class="cell"><p>.205</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>MUC</p></td><td class="cell"><p>.500</p></td><td class="cell"><p>.521</p></td><td class="cell"><p>.507</p></td><td class="cell"><p><b>.595</b></p></td><td class="cell"><p>.541</p></td><td class="cell"><p>.571</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>The use of predicate sense induction based on tiered clustering to overcome the single-class as­sumption caused significant degradation in perfor­mance on this task. Using automatically induced predicates instead of the surface form caused an av­erage degradation of 2.6 points across the twelve tests. A potential explanation for this is that the evaluated predicates have a single dominant sense, meaning the single class assumption may be valid for these predicates (the task-defined selected classes are: location for <i>arrive, </i>event for <i>cancel </i>and <i>finish, </i>proposition for <i>deny, </i>and sound for <i>hear). </i>Therefore it would be interesting to evaluate it on a set of highly polysemous predicates with multi­ple dominant senses. Furthermore, the introduction of predicate sense induction was designed to help cLDA, and the performance degradation for these nine tests was not as large as it was for LDA. For cLDA, Ci had an average degradation of 3.5 points compared to LDA's C1 average degradation of 6.5 points. cLDA's C2 had an average degradation of only 2.5 points and <i>C3 </i>was actually improved by 2.1 points. This suggests that there is value in assign­ing different selected classes via sense induction, but that the two-step approach is not beneficial for these common predicates. This could be overcome by a joint approach of inducing predicate classes while simultaneously detecting coercions, as the presence of many coercions would be an indicator that more induced predicates are necessary.</p></subsection><subsection number="7.3" title="Logical Metonymy Interpretation"><p>For the evaluation of logical metonymy, we use both an existing data set and a newly created data set. Shutova and Teufel (2009) annotated 10 verb-object logical metonymies from Lapata and Las-carides (2003) with sense-disambiguated interpreta­tions and organized the interpretations into clusters representing different possible meanings. For evalu­ation purposes we ignore the sense annotations and clusters and consider all lexical matchings of one of the annotated interpretations to be correct. The that a single-class assumption degrades as the num­ber of classes increases.</p><table caption="Table 3: Accuracy on SemEval-2010 Task 7 data."></table><p>For this evaluation, predicate induction also im­proved LDA for smaller numbers of classes, but not to the degree that it improved cLDA. Without pred­icate induction, LDA outperforms cLDA on all six metrics for 25 and 50 classes. With predicate in­duction, LDA outperforms cLDA on only one metric for 25 classes and five metrics for 50 classes. Thus the induced predicates do reduce the negative im­pact caused by the single selected class assumption for semantic class induction.</p></subsection><subsection number="7.2" title="Coercion Detection"><p>For the evaluation of coercion detection, we use the SemEval-2010 Task 7 data (Pustejovsky et al., 2010). This data uses the most common sense for each of five predicates <i>(arrive, cancel, deny, fin­ish, </i>and <i>hear) </i>with a total of 2,070 sentences an­notated with the argument s source type (the argu­ment s semantic class) and target type (the predi­cate s selected class for that argument). We ignore the actual argument classes and evaluate on the coer­cion type, which is a selection when the source and target type match, and a coercion otherwise.</p><p>In order to evaluate unsupervised systems on this data, we use the corresponding training set (1,031 examples) to learn a threshold for coercion detec­tion. At test time, if the model output is below the threshold, a coercion is inferred. Otherwise it is con­sidered a selection. Therefore, the better a model can rank selections over coercions, the more accu­rate threshold it will learn. The results for this eval­uation are shown in Table 3. The baseline for this task (threshold = 0, or all selections) is 67.4.</p><p>The best overall model on this data is cLDA us­ing the C1 coercion scoring method (Equation (4)). This method consistently outperforms the baseline LDA, especially for smaller numbers of classes, per­forming best with <i>K = </i>25. The second metric, C2, was not as reliable. The third metric, <i>C3, </i>performed poorly on the task. As discussed in Section 3.3, <i>C3</i><page local="9" global="988"/></p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>induced predicates?</p></td><td class="cell"><p>N</p></td><td class="cell"><p>Y</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p># classes</p></td><td class="cell"><p>10</p></td><td class="cell"><p>25</p></td><td class="cell"><p>50</p></td><td class="cell"><p>10</p></td><td class="cell"><p>25</p></td><td class="cell"><p>50</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>LDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>74.4</p></td><td class="cell"><p>78.7</p></td><td class="cell"><p>80.5</p></td><td class="cell"><p>69.7</p></td><td class="cell"><p>70.1</p></td><td class="cell"><p>73.4</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>80.6</p></td><td class="cell"><p><b>81.2</b></p></td><td class="cell"><p>80.9</p></td><td class="cell"><p>76.2</p></td><td class="cell"><p>78.4</p></td><td class="cell"><p>77.5</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c2</p></td><td class="cell"><p>75.4</p></td><td class="cell"><p>75.9</p></td><td class="cell"><p>78.9</p></td><td class="cell"><p>73.5</p></td><td class="cell"><p>68.3</p></td><td class="cell"><p><b>80.8</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c3</p></td><td class="cell"><p>67.8</p></td><td class="cell"><p>70.8</p></td><td class="cell"><p>67.4</p></td><td class="cell"><p>70.9</p></td><td class="cell"><p>67.4</p></td><td class="cell"><p>74.1</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>Table 4: Mean average precision (MAP) scores on the Shutova and Teufel (2009) data set. The bold items indicate the best scores with/without induced predicates as well as using/not using a threshold-based interpretation method.</p><p>Table 5: Mean average precision (MAP) scores on 100 logical metonymies manually annotated with interpretations. The bold items indicate the best scores with/without induced predicates as well as using/not using a threshold-based interpretation method.</p><p>data contains an average of 11 interpretations per metonymy and has a reported 70% recall.</p><p>In order to create a larger data set, we identified 100 verb-object logical metonymies, including those used in Lapata and Lascarides (2003). Three anno-tators were asked to provide up to five interpreta­tions for each metonymy (they were not provided with any verbs from which to choose, only the verb-object pair). The annotators provided an average of 4.6 interpretations per metonymy. Because our goal was recall, inter-annotator agreement was necessar­ily low, and each logical metonymy had an average of 11.7 unique interpretations. All annotators agreed on at least one interpretation for 40 metonymies, while for 14 they had no interpretations in common.<footnote anchor="7"/></p><p>Since logical metonymy interpretation is usually evaluated as a ranking task, we score our methods using mean average precision (MAP):</p><footnote label="7">Data available at http://www.hlt.utdallas.edu/~kirk/data/lmet.zip</footnote><doubt alpha="59.5" length="42" tooSmall="False" monospace="0.0">MAP=lfE^prec(n)^rel(n)(12) Q q=1interps(q)</doubt><p>Where Q is the number of metonymies evaluated; N is the number of interpretations ranked; prec(n) is the precision at rank n; rel(n) = 1 if interpreta­tion n is valid, 0 otherwise; and interps(q) is the number of valid interpretations for the metonymy q. We rank all 4,145 verbs as interpretations except for those removed by the threshold technique, as they have a score of zero. This can give LMth artifi­cially high MAP scores since it may remove some valid interpretations that are low-ranking. However, since a smaller, higher precision list may be useful for many applications we still consider MAP a valid metric and indicate both the highest scoring method and the highest scoring non-threshold method. The results on the Shutova and Teufel (2009) data are shown in Table 4.<page local="10" global="989"/> The results on our own data are shown in Table 5.</p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>induced predicates?</p></td><td class="cell"><p>N</p></td><td class="cell"><p>Y</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p># classes</p></td><td class="cell"><p>10    1    25    1 50</p></td><td class="cell"><p>10    1    25    1 50</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMLL</i></p></td><td class="cell"><p></p></td><td class="cell"><p>0.381</p></td><td class="cell"><p>0.365</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMIND</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p><i>Ci</i></p></td><td class="cell"><p>0.415</p></td><td class="cell"><p>0.406</p></td><td class="cell"><p>0.383</p></td><td class="cell"><p>0.386</p></td><td class="cell"><p>0.412</p></td><td class="cell"><p>0.395</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p><i>c\</i></p></td><td class="cell"><p>0.408</p></td><td class="cell"><p>0.412</p></td><td class="cell"><p>0.412</p></td><td class="cell"><p>0.407</p></td><td class="cell"><p><b>0.468</b></p></td><td class="cell"><p>0.439</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>c2</i></p></td><td class="cell"><p>0.415</p></td><td class="cell"><p>0.447</p></td><td class="cell"><p>0.419</p></td><td class="cell"><p>0.414</p></td><td class="cell"><p>0.415</p></td><td class="cell"><p>0.434</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>c3</i></p></td><td class="cell"><p>0.416</p></td><td class="cell"><p>0.453</p></td><td class="cell"><p><b>0.455</b></p></td><td class="cell"><p>0.395</p></td><td class="cell"><p>0.416</p></td><td class="cell"><p>0.402</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMth</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.599</p></td><td class="cell"><p>0.568</p></td><td class="cell"><p>0.588</p></td><td class="cell"><p>0.479</p></td><td class="cell"><p>0.520</p></td><td class="cell"><p>0.551</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.571</p></td><td class="cell"><p>0.644</p></td><td class="cell"><p>0.751</p></td><td class="cell"><p>0.497</p></td><td class="cell"><p>0.620</p></td><td class="cell"><p><b>0.708</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c2</p></td><td class="cell"><p>0.544</p></td><td class="cell"><p>0.496</p></td><td class="cell"><p>0.633</p></td><td class="cell"><p>0.457</p></td><td class="cell"><p>0.635</p></td><td class="cell"><p>0.660</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c3</p></td><td class="cell"><p>0.601</p></td><td class="cell"><p>0.677</p></td><td class="cell"><p><b>0.767</b></p></td><td class="cell"><p>0.472</p></td><td class="cell"><p>0.622</p></td><td class="cell"><p>0.571</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LM\yt</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.383</p></td><td class="cell"><p>0.381</p></td><td class="cell"><p>0.379</p></td><td class="cell"><p>0.365</p></td><td class="cell"><p>0.356</p></td><td class="cell"><p>0.361</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.380</p></td><td class="cell"><p>0.387</p></td><td class="cell"><p>0.381</p></td><td class="cell"><p>0.386</p></td><td class="cell"><p>0.377</p></td><td class="cell"><p>0.321</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c2</p></td><td class="cell"><p>0.317</p></td><td class="cell"><p>0.342</p></td><td class="cell"><p>0.350</p></td><td class="cell"><p>0.338</p></td><td class="cell"><p>0.340</p></td><td class="cell"><p>0.345</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c3</p></td><td class="cell"><p>0.378</p></td><td class="cell"><p>0.370</p></td><td class="cell"><p>0.350</p></td><td class="cell"><p>0.387</p></td><td class="cell"><p>0.382</p></td><td class="cell"><p>0.384</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p></p></td><td class="cell"><p>induced predicates?</p></td><td class="cell"><p>N</p></td><td class="cell"><p>Y</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p># classes</p></td><td class="cell"><p>10    1    25    1 50</p></td><td class="cell"><p>10    1    25    1 50</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMLL</i></p></td><td class="cell"><p></p></td><td class="cell"><p>0.274</p></td><td class="cell"><p>0.248</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMIND</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.291</p></td><td class="cell"><p>0.286</p></td><td class="cell"><p>0.294</p></td><td class="cell"><p>0.263</p></td><td class="cell"><p>0.267</p></td><td class="cell"><p>0.255</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.296</p></td><td class="cell"><p>0.298</p></td><td class="cell"><p>0.285</p></td><td class="cell"><p>0.280</p></td><td class="cell"><p>0.274</p></td><td class="cell"><p>0.288</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c2</p></td><td class="cell"><p>0.291</p></td><td class="cell"><p>0.287</p></td><td class="cell"><p>0.288</p></td><td class="cell"><p>0.283</p></td><td class="cell"><p>0.271</p></td><td class="cell"><p>0.285</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c3</p></td><td class="cell"><p>0.318</p></td><td class="cell"><p>0.317</p></td><td class="cell"><p><b>0.333</b></p></td><td class="cell"><p>0.298</p></td><td class="cell"><p>0.285</p></td><td class="cell"><p><b>0.307</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LMth</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p>Ci</p></td><td class="cell"><p>0.478</p></td><td class="cell"><p>0.534</p></td><td class="cell"><p>0.534</p></td><td class="cell"><p>0.414</p></td><td class="cell"><p>0.495</p></td><td class="cell"><p>0.479</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p><i>c\</i></p></td><td class="cell"><p>0.449</p></td><td class="cell"><p>0.504</p></td><td class="cell"><p>0.541</p></td><td class="cell"><p>0.391</p></td><td class="cell"><p>0.495</p></td><td class="cell"><p><b>0.513</b></p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>c2</i></p></td><td class="cell"><p>0.505</p></td><td class="cell"><p>0.478</p></td><td class="cell"><p>0.456</p></td><td class="cell"><p>0.398</p></td><td class="cell"><p>0.429</p></td><td class="cell"><p>0.440</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>c3</i></p></td><td class="cell"><p>0.449</p></td><td class="cell"><p>0.496</p></td><td class="cell"><p><b>0.577</b></p></td><td class="cell"><p>0.382</p></td><td class="cell"><p>0.439</p></td><td class="cell"><p>0.446</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>LM\yt</i></p></td><td class="cell"><p>LDA</p></td><td class="cell"><p><i>Ci</i></p></td><td class="cell"><p>0.276</p></td><td class="cell"><p>0.270</p></td><td class="cell"><p>0.271</p></td><td class="cell"><p>0.248</p></td><td class="cell"><p>0.251</p></td><td class="cell"><p>0.249</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>cLDA</p></td><td class="cell"><p><i>Ci</i></p></td><td class="cell"><p>0.271</p></td><td class="cell"><p>0.272</p></td><td class="cell"><p>0.270</p></td><td class="cell"><p>0.257</p></td><td class="cell"><p>0.259</p></td><td class="cell"><p>0.265</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p><i>c2</i></p></td><td class="cell"><p>0.274</p></td><td class="cell"><p>0.274</p></td><td class="cell"><p>0.266</p></td><td class="cell"><p>0.250</p></td><td class="cell"><p>0.259</p></td><td class="cell"><p>0.261</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>c3</p></td><td class="cell"><p>0.271</p></td><td class="cell"><p>0.273</p></td><td class="cell"><p>0.274</p></td><td class="cell"><p>0.253</p></td><td class="cell"><p>0.262</p></td><td class="cell"><p>0.259</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><p>The scores reported in the Shutova and Teufel (2009) data are noticeably higher than the data we annotated. Since the metonymies in our data are a super-set of those in their data, and since for those metonymies our annotators provided approximately the same number of interpretations (110 versus 120), this likely indicates the remaining metonymies in our data are more difficult.</p><p>In all cases the best reported scores use cLDA. Unlike coercion detection on the SemEval data, <i>C3</i>performs very well, achieving the highest scores when no predicate sense induction is used. Also un­like coercion detection, LDA scores do not increase as the number of classes increase. We suspect both these differences have to do with the fact that the ar­guments in this data are far more common. Since LDA is a selectional preference model and its co­ercion scores correspond roughly to the plausibility of seeing a predicate-argument pair, it is less able to distinguish coercions in common arguments.</p><p>Of the logical metonymy ranking methods, LMth consistently produces the highest MAP scores. However, as stated before, by using a cut-off and removing low-ranking valid interpretations, the MAP score is increased, which might not be applica­ble to some applications. The best non-thresholded ranking method is <i>LMind </i>, which naively combines the <i>LMll </i>score with the coercion probability. In almost every case this beats out LMwt . Upon in­spection, we observed that the range of scale for the <i>LMll </i>scores are very inconsistent. This can make it difficult to learn a linear model using these scores as features, and as a result the learned weights were forced to ignore the coercion score and rely entirely on <i>LMll. </i>We attempted other scaling methods, such as a rank-based method, but these had poor re­sults as well, so we leave the problem of the super­vised learning these weights to future work.</p><p>Using induced senses did not result in the dras­tic and consistent degradation in performance seen on the SemEval data, and the highest non-threshold result for the Shutova and Teufel (2009) data used predicate induction. Both metonymy data sets were limited to the verbs found in Lapata and Lascarides (2003), which are still quite common <i>(attempt, be­gin, enjoy, expect, finish, prefer, start, survive, try,</i> <i>want).</i><i> </i>However, the verbs used in our data set had a greater number of WordNet senses attested in a cor­pus than the SemEval data (an average of 4.4 senses for our data versus 3.0 senses for the SemEval data). This suggests the potential value of sense induction for highly polysemous predicates and further moti­vates the integration of sense induction within a se-lectional restriction model.</p></subsection></section><section number="8" title="Conclusion"><p>We have presented a novel topic model that ex­tends an unsupervised selectional preference model (LDA) to an unsupervised selectional restriction model (cLDA) using two assumptions. For the first assumption, that each predicate has a single selected class, we proposed a predicate induction method to overcome predicate polysemy. This improved re­sults for semantic class induction but proved harmful for detecting coercions on common predicates with a single, dominant sense. For the second assump­tion, that the selected class can be inferred from the data, we proposed a sampling method based on the classes of the predicate s arguments. Superior per­formance on coercion detection shows the merit of this assumption.</p><p>Additionally, we proposed methods for improving an existing task, logical metonymy interpretation, using the learned parameters of our model, showing positive results.</p><p>It is clear that our model may be improved by more accurate predicate sense induction. To this end, we plan to develop a model that simultane­ously induces predicates and learns coercions, using knowledge of a predicate s coerciveness to inform the induction mechanism.</p></section><section title="Acknowledgements"><p>We would like to thank Diarmuid O Seaghdha, Bryan Rink, and Anna Rumshisky for several help­ful conversations during the course of this work. We thank Mirella Lapata and Ekaterina Shutova for making the data from their experiments available as well as the organizers of SemEval-2010 Task 7 for the associated data set. Additionally, we thank Srikanth Gullapalli, Aileen McDermott, and Bryan Rink for annotating the data set used in our exper­iment. Finally, we thank the anonymous reviewers for their suggestions on improving this work.</p><page local="11" global="990"/></section><references><p>Amit Bagga and Breck Baldwin. 1998. Algorithms for scoring coreference chains. In <i>In Proceedings of the First International Conference on Language Re­sources and Evaluation Workshop on Linguistic Coref-erence.</i></p><p>David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet allocation. <i>Journal of Machine Learning Research, </i>3:993-1022.</p><p>Samuel Brody and Mirella Lapata. 2009. Bayesianword sense induction. In <i>Proceedings of the 12th Confer­ence ofthe European Chapter ofthe Association for Computational Linguistics, </i>pages 103-111.</p><p>Jonathan Chang, Jordan Boyd-Graber, Sean Gerrish, Chong Wang, and David M. Blei. 2009. Reading tea leaves: How humans interpret topic models. In <i>Neural Information Processing Systems, </i>pages 1-9.</p><p>Chaitanya Chemudugunta, Padhraic Smyth, and Mark Steyvers. 2007. Modeling General and Specific As­pects of Documents with a Probabilistic Topic Model. In <i>Advances in Neural Information Processing Sys­tems 19.</i></p><p>Marie-Catherine de Marneffe, Bill MacCartney, and Christopher D. Manning. 2006. Generating Typed Dependency Parses from Phrase Structure Parses. In <i>Proceedings ofthe Fifth International Language Re­sources and Evaluation.</i></p><p>Christiane Fellbaum, editor. 1998. <i>WordNet: An Elec­tronic Lexical Database. </i>MIT Press.</p><p>Thomas L. Griffiths and Mark Steyvers. 2004. Find­ing scientific topics. <i>Proceedings of the National Academy ofSciences ofthe United States ofAmerica, </i>101(Suppl 1):5228.</p><p>Karin Kipper, Hoa Trang Dang, and Martha Palmer. 1998. Class-based construction of a verb lexicon. In <i>Proceedings ofAAAI/IAAI.</i></p><p>Maria Lapata and Alex Lascarides. 2003. A Probabilistic Account of Logical Metonymy. <i>Computational Lin­guistics, </i>21(2):261-315.</p><p>Dekang Lin and Patrick Pantel. 2001. Induction of Se­mantic Classes from Natural Language Text. In <i>Pro­ceedings ofACMSIGKDD Conference on Knowledge Discovery and Data Mining, </i>pages 317-322.</p><p>Christopher D. Manning, Prabhakar Raghavan, and Hin-rich Schütze. 2008. <i>Introduction to Information Re­trieval. </i>Cambridge University Press.</p><p>Andrew Kachites McCallum. 2002. Mallet: A machine learning for language toolkit.</p><p>Diarmuid O Seaghdha. 2010. Latent variable models of selectional preference. In <i>Proceedings of the 48th Annual Meeting ofthe Association for Computational Linguistics, </i>pages 435-444.</p><p>Robert Parker, David Graff, Junbo Kong, Ke Chen, and Kazuaki Maeda. 2009. English Gigaword Fourth Edi­tion. <i>The LDC Corpus Catalog., </i>LDC2009T13.</p><p>James Pustejovsky, Anna Rumshisky, Alex Plotnick, Elisabetta Jezek, Olga Batiukova, and Valeria Quochi. 2010. SemEval-2010 Task 7: Argument Selection and Coercion. In <i>Proceedings of the 5th International Workshop on Semantic Evaluation, </i>pages 27-32. Joseph Reisinger and Raymond J. Mooney. 2010. A Mixture Model with Sharing for Lexical Semantics. In <i>Proceedings ofthe Conference on Empirical Methods in Natural Language Processing, </i>pages 1173-1182. Alan Ritter, Mausam, and Oren Etzioni. 2010. A La­tent Dirichlet Allocation method for Selectional Pref­erences. In <i>Proceedings ofthe 48th Annual Meeting of the Association for Computational Linguistics, </i>pages 424-434.</p><p>Mats Rooth, Stefan Riezler, Detlef Prescher, Glenn Car­roll, and Franz Beil. 1999. Inducing a semantically annotated lexicon via EM-based clustering. In <i>Pro­ceedings ofthe 3 7th Annual Meeting of the Association for Computational Linguistics. </i>Anna Rumshisky, Victor A. Grinberg, and James Puste-jovsky. 2007. Detecting selectional behavior of com­plex types in text. In <i>Fourth International Workshop on Generative Approaches to the Lexicon. </i>Ekaterina Shutova and Simone Teufel. 2009. Logical Metonymy: Discovering Classes of Meaning. In <i>Pro­ceedings of the CogSci 2009 Workshop on Semantic Space Models. </i>Ekaterina Shutova. 2009. Sense-based interpretation of logical metonymy using a statistical method. In <i>Pro­ceedings ofthe ACL 2009 Student Workshop. </i>Marc Vilain, John Burger, John Aberdeen, Dennis Con­nolly, and Lynette Hirschman. 1995. A model-theoretic coreference scoring scheme. In <i>Proceedings fo the 6th Message Understanding Conference, </i>pages 45-52.</p></references></body></article>