<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1008">
  <Title>Identifying Semantic Roles Using Combinatory Categorial Grammar</Title>
  <Section position="4" start_page="0" end_page="3" type="metho">
    <SectionTitle>
3 Predicate-argument relations in CCG
</SectionTitle>
    <Paragraph position="0"> Combinatory Categorial Grammar (CCG) (Steedman, 2000), is a grammatical theory which provides a completely transparent interface between surface syntax and underlying semantics, such that each syntactic derivation corresponds directly to an interpretable semantic representation which includes long-range dependencies that arise through control, raising, coordination and extraction.</Paragraph>
    <Paragraph position="1"> In CCG, words are assigned atomic categories such as NP, or functor categories like (S[dcl]nNP)=NP (transitive declarative verb) or S=S (sentential modifier). Adjuncts are represented as functor categories such as S=S which expect and return the same type. We use indices to number the arguments of functor categories, eg.</Paragraph>
    <Paragraph position="2">  , and indicate the word-word dependencies in the predicate-argument structure as tupleshw</Paragraph>
    <Paragraph position="4"> is the lexical category of the head word w</Paragraph>
    <Paragraph position="6"> Long-range dependencies can be projected through certain types of lexical categories or through rules such as coordination of functor categories. For example, in the lexical category of a relative pronoun, (NPnNP</Paragraph>
    <Paragraph position="8"> ), the head of the NP that is missing from the relative clause is unified with (as indicated by the indices i) the head of the NP that is modified by the entire relative clause.</Paragraph>
    <Paragraph position="9"> Figure 1 shows the derivations of an ordinary sentence, a relative clause and a right-node-raising construction. In all three sentences, the predicate-argument relations between London and denied and plans and denied are the same, which in CCG is expressed by the fact that London fills the first (ie. subject) argument slot of the lexical category of de- null , and plans fills the second (object) slot. The relations extracted from the CCG derivation for the sentence &amp;quot;London denied plans on Monday&amp;quot; are shown in Table 1.</Paragraph>
    <Paragraph position="10"> The CCG parser returns the local and long-range word-word dependencies that express the predicate-argument structure corresponding to the derivation. These relations are recovered with an accuracy of around 83% (labeled recovery) or 91% (unlabeled recovery) (Hockenmaier, 2003). By contrast, standard Treebank parsers such as (Collins, 1999) only return phrase-structure trees, from which non-local dependencies are difficult to recover.</Paragraph>
  </Section>
  <Section position="5" start_page="3" end_page="3" type="metho">
    <SectionTitle>
3 Monday
</SectionTitle>
    <Paragraph position="0"> sentence &amp;quot;London denied plans on Monday&amp;quot; The CCG parser has been trained and tested on CCGbank (Hockenmaier and Steedman, 2002a), a treebank of CCG derivations obtained from the Penn Treebank, from which we also obtain our training data.</Paragraph>
  </Section>
  <Section position="6" start_page="3" end_page="3" type="metho">
    <SectionTitle>
4 Mapping between PropBank and
</SectionTitle>
    <Paragraph position="0"> CCGbank Our aim is to use CCG derivations as input to a system for automatically producing the argument labels of PropBank. In order to do this, we wish to correlate the CCG relations above with PropBank arguments. PropBank argument labels are assigned to nodes in the syntactic trees from the Penn Treebank. While the CCGbank is derived from the Penn Treebank, in many cases the constituent structures do not correspond. That is, there may be no constituent in the CCG derivation corresponding to the same sequence of words as a particular constituent in the Treebank tree. For this reason, we compute the correspondence between the CCG derivation and the PropBank labels at the level of head words. For each role label for a verb's argument in PropBank, we first find the head word for its constituent according to the the head rules of (Collins, 1999). We then look for the label of the CCG relation between this head word and the verb itself.</Paragraph>
  </Section>
  <Section position="7" start_page="3" end_page="3" type="metho">
    <SectionTitle>
5 The Experiments
</SectionTitle>
    <Paragraph position="0"> In previous work using the PropBank corpus, Gildea and Palmer (2002) developed a system to predict semantic roles from sentences and their parse trees as determined by the statistical parser of Collins (1999). We will briefly review their probability model before adapting the system to incorporate features from the CCG derivations.</Paragraph>
    <Section position="1" start_page="3" end_page="3" type="sub_section">
      <SectionTitle>
5.1 The model of Gildea and Palmer (2002)
</SectionTitle>
      <Paragraph position="0"> For the Treebank-based system, we use the probability model of Gildea and Palmer (2002). Probabilities of a parse constituent belonging to a given semantic role are calculated from the following features: null The phrase type feature indicates the syntactic type of the phrase expressing the semantic roles: examples include noun phrase (NP), verb phrase (VP), and clause (S).</Paragraph>
      <Paragraph position="1"> The parse tree path feature is designed to capture the syntactic relation of a constituent to the predicate. It is defined as the path from the predicate through the parse tree to the constituent in question, represented as a string of parse tree nonterminals linked by symbols indicating upward or downward movement through the tree, as shown in Figure 2.</Paragraph>
      <Paragraph position="2"> Although the path is composed as a string of symbols, our systems will treat the string as an atomic value. The path includes, as the first element of the string, the part of speech of the predicate, and, as the last element, the phrase type or syntactic category of the sentence constituent marked as an argument.</Paragraph>
      <Paragraph position="3">  in the parse tree and#downward movement.</Paragraph>
      <Paragraph position="4"> The position feature simply indicates whether the constituent to be labeled occurs before or after the predicate. This feature is highly correlated with grammatical function, since subjects will generally appear before a verb, and objects after. This feature may overcome the shortcomings of reading grammatical function from the parse tree, as well as errors in the parser output.</Paragraph>
      <Paragraph position="5"> The voice feature distinguishes between active and passive verbs, and is important in predicting semantic roles because direct objects of active verbs correspond to subjects of passive verbs. An instance of a verb was considered passive if it is tagged as a past participle (e.g. taken), unless it occurs as a descendent verb phrase headed by any form of have (e.g. has taken) without an intervening verb phrase headed by any form of be (e.g. has been taken).</Paragraph>
      <Paragraph position="6"> The head word is a lexical feature, and provides information about the semantic type of the role filler. Head words of nodes in the parse tree are determined using the same deterministic set of head word rules used by Collins (1999).</Paragraph>
      <Paragraph position="7"> The system attempts to predict argument roles in new data, looking for the highest probability assignment of roles r</Paragraph>
      <Paragraph position="9"> to all constituents i in the sentence, given the set of features F</Paragraph>
      <Paragraph position="11"> g at each constituent in the parse tree, and the predicate p:</Paragraph>
      <Paragraph position="13"> We break the probability estimation into two parts, the first being the probability P(r</Paragraph>
      <Paragraph position="15"> a constituent's role given our five features for the consituent, and the predicate p. Due to the sparsity of the data, it is not possible to estimate this probability from the counts in the training data. Instead, probabilities are estimated from various subsets of the features, and interpolated as a linear combination of the resulting distributions. The interpolation is performed over the most specific distributions for which data are available, which can be thought of as choosing the topmost distributions available from a backoff lattice, shown in Figure 3.</Paragraph>
      <Paragraph position="17"/>
      <Paragraph position="19"> gjp) for a set of roles appearing in a sentence given a predicate, using the following formula:</Paragraph>
      <Paragraph position="21"> This approach, described in more detail in Gildea and Jurafsky (2002), allows interaction between the role assignments for individual constituents while making certain independence assumptions necessary for efficient probability estimation. In particular, we assume that sets of roles appear independent of their linear order, and that the features F of a constituents are independent of other constituents' features given the constituent's role.</Paragraph>
    </Section>
    <Section position="2" start_page="3" end_page="3" type="sub_section">
      <SectionTitle>
5.2 The model for CCG derivations
</SectionTitle>
      <Paragraph position="0"> In the CCG version, we replace the features above with corresponding features based on both the sentence's CCG derivation tree (shown in Figure 1) and the CCG predicate-argument relations extracted from it (shown in Table 1).</Paragraph>
      <Paragraph position="1"> The parse tree path feature, designed to capture grammatical relations between constituents, is replaced with a feature defined as follows: If there is a dependency in the predicate-argument structure of the CCG derivation between two words w and w</Paragraph>
      <Paragraph position="3"> the path feature from w to w  is defined as the lexical category of the functor, the argument slot i occupied by the argument, plus an arrow ( or!) to indicate whether w or w  is the categorial functor. For example, in our sentence &amp;quot;London denied plans on Monday&amp;quot;, the relation connecting the verb denied with plans is (S[dcl]nNP)=NP.2. , with the left arrow indicating the lexical category included in the relation is that of the verb, while the relation connecting denied with on is ((SnNP)n(SnNP))=NP.2.!, with the right arrow indicating the the lexical category included in the relation is that of the modifier. If the CCG derivation does not define a predicate-argument relation between the two words, we use the parse tree path feature described above, defined over the CCG derivation tree. In our training data, 77% of PropBank arguments corresponded directly to a relation in the CCG predicate-argument representation, and the path feature was used for the remaining 23%. Most of these mismatches arise because the CCG parser and PropBank differ in their definition of head words. For instance, the CCG parser always assumes that the head of a PP is the preposition, whereas PropBank roles can be assigned to the entire PP (7), or only to the NP argument of the preposition (8), in which case the head word comes from the NP: (7) ... will be offered [PP ARGM-LOC in the U.S].</Paragraph>
      <Paragraph position="4"> (8) to offer ...[PP to [NP ARG2 the public]].</Paragraph>
      <Paragraph position="5"> In embedded clauses, CCG assumes that the head is the complementizer, whereas in PropBank, the head comes from the embedded sentence itself. In complex verb phrases (eg. &amp;quot;might not have gone&amp;quot;), the CCG parser assumes that the first auxiliary (might) is head, whereas PropBank assumes it is the main verb (gone). Therefore, CCG assumes that not modifies might, whereas PropBank assumes it modifies gone. Although the head rules of the parser could in principle be changed to reflect more directly the dependencies in PropBank, we have not attempted to do so yet. Further mismatches occur because the predicate-argument structure returned by the CCG parser only contains syntactic dependencies, whereas the PropBank data also contain some anaphoric dependencies, eg.:  the adhesive] is designed to...</Paragraph>
      <Paragraph position="6"> Such dependencies also do not correspond to a relation in the predicate-argument structure of the CCG derivation, and cause the path feature to be used. The phrase type feature is replaced with the lexical category of the maximal projection of the Prop-Bank argument's head word in the CCG derivation tree. For example, the category of plans is N, and the category of denied is (S[dcl]nNP)=NP.</Paragraph>
      <Paragraph position="7"> The voice feature can be read off the CCG categories, since the CCG categories of past participles carry different features in active and passive voice (eg. sold can be (S[pt]nNP)=NP or S[pss]nNP).</Paragraph>
      <Paragraph position="8"> The head word of a constituent is indicated in the derivations returned by the CCG parser.</Paragraph>
      <Paragraph position="9">  are annotated with PropBank roles ARG0, ARG1 and ARGM-TMP.</Paragraph>
    </Section>
    <Section position="3" start_page="3" end_page="3" type="sub_section">
      <SectionTitle>
5.3 Data
</SectionTitle>
      <Paragraph position="0"> We use data from the November 2002 release of PropBank. The dataset contains annotations for 72,109 predicate-argument structures with 190,815 individual arguments (of which 75% are core, or numbered, arguments) and has includes examples from 2462 lexical predicates (types). Annotations from Sections 2 through 21 of the Treebank were used for training; Section 23 was the test set. Both parsers were trained on Sections 2 through 21.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>