<?xml version="1.0" standalone="yes"?>
<Paper uid="P06-1113">
  <Title>Question Answering with Lexical Chains Propagating Verb Arguments</Title>
  <Section position="4" start_page="897" end_page="898" type="metho">
    <SectionTitle>
2 VerbNet Syntactic Patterns
</SectionTitle>
    <Paragraph position="0"> The algorithm for propagating verb arguments uses structures for representing them. Several choices were considered for retrieving verbs' argument structure. Verb syntactic patterns from WordNet (called frames) could not be used because some tokens in the patterns (like &amp;quot;PP&amp;quot; or &amp;quot;CLAUSE&amp;quot;) cannot be mapped to arguments.</Paragraph>
    <Paragraph position="1"> FrameNet (Baker et al., 1998) and PropBank (Kingsbury and Palmer, 2002) contain verb syntactic patterns, but they do not have a mapping to WordNet. Finally VerbNet (Kipper et al., 2000b) represents a verb lexicon with syntactic and semantic information. This resource has a mapping to WordNet and therefore was considered the most suitable for propagating predicate arguments along lexical chains.</Paragraph>
    <Section position="1" start_page="897" end_page="897" type="sub_section">
      <SectionTitle>
2.1 VerbNet description
</SectionTitle>
      <Paragraph position="0"> VerbNet is based on classes of verbs. Each verb entry points to a set of classes and each class represents a sense of a verb. The classes are organized hierarchically. Each class contains a set of syntactic patterns corresponding to licensed constructions. Each syntactic pattern is an ordered list of tokens and each token represents a group of words.</Paragraph>
      <Paragraph position="1"> The tokens contain various information and constraints about the word or the group of words they represent. The name of the token can represent the thematic role of an argument, the verb itself, prepositions, adjectives, adverbs or plain words.</Paragraph>
      <Paragraph position="2"> VerbNet uses 29 thematic roles (presented in ta-</Paragraph>
    </Section>
    <Section position="2" start_page="897" end_page="897" type="sub_section">
      <SectionTitle>
Extent Value
</SectionTitle>
      <Paragraph position="0"> ble 1). VerbNet has a static aspect and a dynamic aspect. The static aspect refers to the organization of verb entries. The dynamic aspect refers to the lexicalized trees associated with syntactic patterns. A detailed description of VerbNet dynamic aspect can be found in (Kipper et al., 2000a).</Paragraph>
      <Paragraph position="1"> The algorithm for propagating predicate arguments uses the syntactic patterns associated with each sensekey. Each class contains a set of Word-Net verb sensekeys and a set of syntactic patterns. Therefore, syntactic patterns can be associated with verb sensekey from the same class. Since sensekeys represent word senses in WordNet, each verb synset can be associated with a set of VerbNet syntactic patterns. VerbNet syntactic patterns allow predicate arguments to be propagated along lexical chains. However, not all verb senses in WordNet are listed in VerbNet classes. For the remaining verb sensekeys that are not listed in Verb-Net, syntactic patterns were assigned automatically using machine learning as described in the following section.</Paragraph>
    </Section>
    <Section position="3" start_page="897" end_page="898" type="sub_section">
      <SectionTitle>
2.2 Associating syntactic patterns with new
</SectionTitle>
      <Paragraph position="0"> verb senses In order to propagate predicate arguments along lexical chains, ideally every verb in every synonym set has to have a set of syntactic patterns. Only a part of verb senses are listed in VerbNet classes. WordNet 2.0 has 24,632 verb sensekeys, but only 4,983 sensekeys are listed in VerbNet classes. For the rest, syntactic patterns were assigned automatically. In order to assign these syntactic patterns to the verb senses not listed in Verb-Net, training examples were needed, both positive and negative. The learning took place for one syntactic pattern at a time. A syntactic pattern can be listed in more than one class. All verb senses associated with a syntactic pattern can be considered positive examples of verbs having that syntactic pattern. For generating negative examples,  the following assumption was used: if a verb sense listed in a VerbNet class is not associated with a given syntactic pattern, then that verb sense represents a negative example for that pattern. 352 syntactic patterns were found in all VerbNet classes. A training example was generated for each pair of syntactic patterns and verb sensekeys, resulting in a total number of 1,754,016 training examples.</Paragraph>
      <Paragraph position="1"> These training examples were used to infer rules that would classify if a verb sense key can be associated with a given syntactic pattern. Training examples were created by using the following features: verb synset semantic category, verb synset position in the IS-A hierarchy, the fact that the verb synset is related to other synsets with CAUSATION relation, the semantic classes of all noun synsets derivationally related with the given verb synset and the WordNet syntactic pattern ids. A machine learning algorithm based on C5.0 (Quinlan, 1998) was run on these training examples. Table 2 presents the performance of the learning algorithm using a 10-fold cross validation for several patterns. A number of 20,759 pairs of verb senses with their syntactic patterns were added to the existing 35,618 pairs in VerbNet. In order to improve the performance of the question answering system, around 100 patterns were manually associated with some verb senses.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="898" end_page="900" type="metho">
    <SectionTitle>
3 Propagating Verb Arguments
</SectionTitle>
    <Paragraph position="0"> Given the argument structure of a verb in a sentence and a lexical chain between this verb and another, the algorithm for propagating verb arguments transforms this structure step by step, for each relation in the lexical chain. During each step the head of the structure changes its value and the arguments can change their position. The arguments change their position in a way that preserves the original meaning as much as possible.</Paragraph>
    <Paragraph position="1"> The argument structures mirror the syntactic patterns that a verb with a given sense can have. An argument structure contains the type of the pattern, the head and an array of tokens. Each token represents an argument with a thematic role or an adjective, an adverb, a preposition or just a regular word. The head and the arguments with thematic roles are represented by concepts. A concept is created from a word found in text. If the word is found in WordNet, the concept structure contains its surface form, its lemma, its part of speech and its WordNet sense. If the word is not found in WordNet, its concept structure contains only the word and the part of speech. The value of the field for an argument is represented by the concept that is the head of the phrase representing the argument. Because a synset may contain more than one verb and each verb can have different types of syntactic patterns, propagation of verb arguments along a single relation can result in more than one structure. The output of the algorithm as well as the output of the propagation of each relation in the lexical chain is the set of argument structures with the head being a verb from the set of synonyms of the target synset. For a given relation in the lexical chain, each structure coming from the previous step is transformed into a set of new structures. The relations used and the process of argument propagation is described below.</Paragraph>
    <Section position="1" start_page="898" end_page="900" type="sub_section">
      <SectionTitle>
3.1 Relations used
</SectionTitle>
      <Paragraph position="0"> A restricted number of WordNet relations were used for creating lexical chains. Lexical chains between verbs were used for propagating verb arguments, and lexical chains between nouns were used to link semantically related arguments expressed with different words.</Paragraph>
      <Paragraph position="1"> Between verb synsets the following relations were used: HYPERNYM, TROPONYM, ENTAILMENT and CAUSATION. These relations were selected because they reveal patterns about how they propagate predicate arguments.</Paragraph>
      <Paragraph position="2"> The HYPERNYMY relation links one specific verb synset to one that is more general. Most of the time, the arguments have the same thematic roles for the two verbs. Sometimes the hypernym  synset has a syntactic pattern that has more thematic roles than the syntactic pattern of the start synset. In this case the pattern of the hypernym is not considered for propagation.</Paragraph>
      <Paragraph position="3"> The HYPONYMY relation is the reverse of HYPERNYMY and links one verb synset to a more specific one. Inference to a more specific verb requires abduction. Most of the time, the arguments have the same thematic roles for the two verbs.</Paragraph>
      <Paragraph position="4"> Usually the hyponym of the verb synset is more specific and have less syntactic patterns than the original synset. This is why a syntactic pattern of a verb can be linked with the syntactic pattern of its hyponym that has more thematic roles. These additional thematic roles in the syntactic pattern of the hyponym will receive the value ANY-CONCEPT when verb arguments are propagated along this relation. null ENTAILMENT relation links two verb synsets that express two different events that are related: the first entails the second. This is different than HYPERNYMY or HYPONYMY that links verbs that express the same event with more or less details. Most of the time the subject of these two sentences has the same thematic role. If the thematic role of subjects is different, then the syntactic pattern of the target verb is not considered for propagation. The same happens if the start pattern contains less arguments than the target pattern. Additional arguments can change the meaning of the target pattern. null A relation that is the reverse of the ENTAILMENT is not coded in WordNet but, it is used for a better connectivity. Given one sentence a0a2a1 with a verb a3 a1 that is entailed by a verb a3a5a4 , the sentence a0a6a1 can be reformulated using the verb a3a7a4 , and thus creating sentence a0 a4 . Sentence a0 a1 does not imply sentence a0 a4 but makes it plausible. Most of the time, the subject of these two sentences has the same thematic role. If the thematic role of subjects is different, then the pattern of the target verb synset is not considered for propagation. The same happens if the start pattern has less arguments than the target pattern. Additional arguments can change the meaning of the target pattern. null The CAUSATION relation puts certain restrictions on the syntactic patterns of the two verb synsets. The first restriction applies to the syntactic pattern of the start synset: its subject must be an Agent or an Instrument and its object must be a Patient.</Paragraph>
      <Paragraph position="5"> The second restriction applies to the syntactic pattern of the destination synset: its subject must be a Patient. If the two syntactic patterns obey these restrictions then an instance of the destination synset pattern is created and its arguments will receive the value of the argument with the same thematic role in the pattern belonging to start synset.</Paragraph>
      <Paragraph position="6"> The reverse of the CAUSATION relation is not codified in WordNet database but it is used in lexical chains to increase the connectivity between synsets. Similar to causation relation, the reverse causation imposes two restrictions on the patterns belonging to the start and destination synset. First restriction applies to the syntactic pattern of the start synset: its subject must have the thematic role of Patient. The second restriction applies to the syntactic pattern of the destination synset: its subject must be an Agent or an Instrument and its object must be a Patient. If the two syntactic patterns obey these restrictions then an instance of the destination synset pattern is created and its arguments will receive the value of the argument with the same thematic role in the pattern belonging to start synset.</Paragraph>
      <Paragraph position="7"> When deriving lexical chains for linking words from questions and correct answers in TREC 2004, it was observed that many chains contain a pair of DERIVATION relations. Since a pair of DERIVATION relations can link either two noun synsets or two verb synsets, the pair was concatenated into a new relation called SIM DERIV. The number of SIM-DERIV relations is presented in table 3. For example the verb synsets emanate#2 and emit#1 are not synonyms (not listed in the same synset) but they are linked by a SIM-DERIV relation (both have a DERIVATION relation to the noun synset (n-emission#1, emanation#2) - nominalizations of the two verbs are listed in the same synset). There are no restrictions between pairs of patterns that participate in argument propagation.</Paragraph>
      <Paragraph position="8"> The arguments in the syntactic pattern instance of the destination synset take their values from the arguments with the same thematic roles from the  The VERBGROUP and SEE-ALSO relations were not included in the experiment because it is not clear how they propagate arguments.</Paragraph>
      <Paragraph position="9"> A restricted set of instances of DERIVATION relation was used to link verbs to nouns that describe their action. When arguments are propagated from verb to noun, the noun synset will receive a set of syntactic patterns instances similar to the semantic instances of the verb. When arguments are propagated from noun to verb, a new created structure for the verb sense takes the values for its arguments from the arguments with similar thematic roles in the noun structure.</Paragraph>
      <Paragraph position="10"> Between the heads of two argument structures there can exist lexical chains of size 0, meaning that the heads of the two structures are in the same synset. However, the type of the start structure can be different than the type of the target structure. In this case, the arguments still have to be propagated from one structure to another. The arguments in the target structure will take the values of the arguments with the same thematic role in the start structure or the value ANY-CONCEPT if these arguments cannot be found.</Paragraph>
      <Paragraph position="11"> Relations between nouns were not used by the algorithm but they are used after the algorithm is applied, to link the arguments from a resulted structure to the arguments with the same semantic roles in the target structure. If such a link exists, then the arguments are considered to match. From the existing WordNet relations between noun synsets only HYPERNYM and HY-PONYM were used.</Paragraph>
    </Section>
    <Section position="2" start_page="900" end_page="900" type="sub_section">
      <SectionTitle>
3.2 Assigning weights to the relations
</SectionTitle>
      <Paragraph position="0"> Two synsets can be connected by a large number of lexical chains. For efficiency, the algorithm runs only on a restricted number of lexical chains.</Paragraph>
      <Paragraph position="1"> In order to select the most likely lexical chains, they were ordered decreasingly by their weight.</Paragraph>
      <Paragraph position="2"> The weight of a lexical chain is computed using the following formula inspired by (Moldovan and</Paragraph>
      <Paragraph position="4"> where n represents the number of relations in the lexical chain. The formula uses the weights a0a29a28 (a30 a1a32a31a17a19a17a34a33 ) of the relations along the chain (presented in table 4) and coefficients for pairs of relationsa6 a28a8a28a36a35 a1 (some of them presented in table 5, the rest having a weight of 1.0). This formula resulted from the observation that the relations are not equal (some relations like HYPERNYMY are stronger than other relations) and that the order of relations in the lexical chain influences its fitness (the order of relations is approximated by the weight given to pairs of relations). The formula uses the &amp;quot;measure of generality&amp;quot; of a concept defined as:</Paragraph>
      <Paragraph position="6"/>
    </Section>
  </Section>
  <Section position="6" start_page="900" end_page="901" type="metho">
    <SectionTitle>
CAUSATION ENTAILMENT 1.25
CAUSATION R-ENTAILMENT 0.8
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="900" end_page="901" type="sub_section">
      <SectionTitle>
3.3 Example
</SectionTitle>
      <Paragraph position="0"> In the test set from the QA track in TREC 2004 we found the following question with correct answer: Q 28.2: (Abercrombie &amp; Fitch) When was it established? A: ... Abercrombie &amp; Fitch began life in 1982 ... The verb establish in the question has sense 2 in WordNet 2.0 and the verb begin in the answer  has also sense 2. The following lexical chain can be found between these two verbs:  where the argument with the thematic role of Agent has the value ANY-CONCEPT, and the Patient argument has the value Abercrombie &amp; Fitch. From the answer, an argument structure is created for verb begin#2 using the pattern:a0</Paragraph>
      <Paragraph position="2"> where the Patient argument has the value Abercrombie &amp; Fitch and the Theme argument has the value n-life#2. This structure is propagated along the lexical chain, each relation at a time. First for the R-CAUSATION relation links the verb begin#2  event though it is changing its syntactic role from subject of the verb begin#2 to the object of the verb begin#3. The Theme argument is lost along this relation, instead the new argument with the thematic role of Agent receives the special value ANY-CONCEPT.</Paragraph>
      <Paragraph position="3"> The second relation in the chain, SIM-DERIV links two verbs that have the same syntactic pat-</Paragraph>
      <Paragraph position="5"> Therefore a new structure is created for the verb establish#2 using this pattern and its arguments take their values from the similar arguments in the argument structure for verb begin#3. This new structure exactly matches the argument structure from the question therefore the answer is ranked the highest in the set of candidate answer. Figure 1 illustrates the argument propagation process for this example.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>