<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-1210">
  <Title>Automatic Detection of Causal Relations for Question Answering</Title>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
4 Automatic detection of causation
</SectionTitle>
    <Paragraph position="0"> relationships In this section we describe a method for automatic detection of lexico-syntactic patterns that express causation.</Paragraph>
    <Paragraph position="1"> The algorithm consists of two major procedures. The first procedure discovers lexico-syntactic patterns that can express the causation relation, and the second procedure presents an inductive learning approach to the automatic detection of syntactic and semantic constraints on the constituent components.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.1 Automatic discovery of lexico-syntactic
</SectionTitle>
      <Paragraph position="0"> patterns referring to causation One of the most frequent explicit intra-sentential patterns that can express causation is a0a2a1a4a3a6a5a8a7a10a9a12a11a14a13 a1a4a3a16a15a18a17 . In this paper we focus on this kind of patterns, where the verb is a simple causative.</Paragraph>
      <Paragraph position="1"> In order to catch the most frequently used lexico-syntactic patterns referring to causation, we used the following procedure (Hearst 1998): Discovery of lexico-syntactic patterns: Input: semantic relation R Output: lexico-syntactic patterns expressing R  STEP 1. Pick a semantic relation R (in this paper, CAUSATION) STEP 2. Pick a pair of noun phrases a19a6a20 , a19a22a21 among which R holds.</Paragraph>
      <Paragraph position="2"> Since CAUSE-TO is one of the semantic relations explicitly used in WordNet, this is an excellent re- null source for picking a19a6a20 and a19a22a21 . The CAUSE-TO relation is a transitive relation between verb synsets. For example, in WordNet the second sense of the verb develop is &amp;quot;causes to grow&amp;quot;. Although WordNet contains numerous causation relationships between nouns that are always true, they are not directly mentioned. One way to determine such relationships is to look for all patterns a0a2a1a4a3 a5a22a23a25a24a27a26a29a28 a9 a7a10a13a30a1a31a3a32a15a18a17 that occur between a noun entry and another noun in the corresponding gloss definition. One such example is the causation relationship between a0 bonyness a1 and a0 starvation a1 . The gloss of a0 bonyness (#1/1) a1 is (extreme leanness (usually caused by starvation or disease)). null WordNet 1.7 contains 429 such relations linking nouns from different domains, the most frequent being medicine (about 58.28%).</Paragraph>
      <Paragraph position="3"> STEP 3. Extract lexico-syntactic patterns that link the two selected noun phrases by searching a collection of texts.</Paragraph>
      <Paragraph position="4"> For each pair of causation nouns determined above, search the Internet or any other collection of documents and retain only the sentences containing the pair. From these sentences, determine automatically all the patterns a0a2a1a4a3 a5 verb/verb expression a1a4a3 a15 a17 , where a0a2a1a4a3 a5 - a1a31a3 a15 a17 is the pair considered. null The result is a list of verbs/verbal expressions that refer to causation (see Table 1). Some of these verbs are always referring to causation, but most of them are ambiguous, as they express a causation relation only in a particular context and only between specific pairs of nouns. For example, a0 a1a31a3 a5 produces a1a4a3a16a15a18a17 . In most cases, the verb produce has the sense of manufacture, and only in some particular contexts it refers to causation.</Paragraph>
      <Paragraph position="5"> In this approach, the acquisition of linguistic patterns is done automatically, as the pattern is predefined (a0a2a1a4a3 a5 verb a1a4a3a32a15a18a17 ). As described in the next subsections, the relationships are disambiguated and only those referring to causation are retained.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
4.2 Learning Syntactic and Semantic
</SectionTitle>
      <Paragraph position="0"> Constraints for causal relation The learning procedure proposed here is supervised, for the learning algorithm is provided with a set of inputs along with the corresponding set of correct outputs. Based on a set of positive and negative causal training examples provided and annotated by the user, the algorithm creates a decision tree and a set of rules that classify new data. The rules produce constraints on the noun constituents of the lexical patterns.</Paragraph>
      <Paragraph position="1"> For the discovery of the semantic constraints we used C4.5 decision tree learning (Quinlan 1999).</Paragraph>
      <Paragraph position="2"> The learned function is represented by a decision tree, or a set of if-then rules. The decision tree learning searches a complete hypothesis space from simple to complex hypotheses until it finds a hypothesis consistent with the data. Its bias is a preference for the shorter tree that places high information gain attributes closer to the root.</Paragraph>
      <Paragraph position="3"> The error in the training examples can be overcome by using different training and a test corpora, or by cross-validation techniques.</Paragraph>
      <Paragraph position="4"> C4.5 receives in general two input files, the NAMES file defining the names of the attributes, attribute values and classes, and the DATA file containing the examples.</Paragraph>
      <Paragraph position="5">  Since a part of our constraint learning procedure is based on the semantic information provided by WordNet, we need to preprocess the noun phrases (NPs) extracted and identify the cause and the effect. For each NP we keep only the largest word sequence (from left to right) that is defined in WordNet as a concept.</Paragraph>
      <Paragraph position="6"> For example, from the noun phrase &amp;quot;a 7.1 magnitude earthquake&amp;quot; the procedure retains only &amp;quot;earthquake&amp;quot;, as it is the WordNet concept with the largest number of words in the noun phrase.</Paragraph>
      <Paragraph position="7"> We did not consider those noun phrases in which the head word had other part of speech than noun.</Paragraph>
    </Section>
    <Section position="3" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
Test Corpus
</SectionTitle>
      <Paragraph position="0"> In order to learn the constraints, we used the LA TIMES section of the TREC 9 text collection. For each of the 60 verbs generated with the procedure described in section 4.1, we searched the text collection and retained 120 sentences containing the verb. Thus, a training corpus &amp;quot;A&amp;quot; of 6,000 sentences, and respectively, a test corpus of 1,200 sentences were automatically created. Each sentence in these corpora was then parsed using the syntactic parser developed by Charniak (Charniak 1999).</Paragraph>
      <Paragraph position="1"> Focusing only on the sentences containing relations indicated by the pattern considered, we manually annotated all instances matched by the pattern as referring to causation or not. Using the training corpus, the system extracted 6,523 relationships of the type a0 a1a31a3 a5 verb a1a4a3a32a15a18a17 , from which 2,101 were Causal verbs give rise (to) stir up create start induce entail launch make produce contribute (to) develop begin generate set up bring rise effect trigger off stimulate bring about commence call forth provoke set off unleash arouse set in motion effectuate elicit bring on kick up lead (to) conduce (to) give birth (to) trigger educe derive (from) originate in call down associate (with) lead off put forward  causal relations, while 4,422 were not.</Paragraph>
      <Paragraph position="2">  The next step consists of detecting the constraints necessary on nouns and verb for the pattern a0a2a1a4a3a6a5 verb a1a31a3a32a15a18a17 such that the lexico-syntactic pattern indicates a causation relationship.</Paragraph>
      <Paragraph position="3"> The basic idea we employ here is that only some categories of noun phrases can be associated with a causation link. According to the philosophy researcher Jaegwon Kim (Kim 1993), any discussion of causation implies an ontological framework of entities among which causal relations are to hold, and also &amp;quot;an accompanying logical and semantical framework in which these entities can be talked about&amp;quot;. He argues that the entities that represent either causes or effects are often events, but also conditions, states, phenomena, processes, and sometimes even facts, and that coherent causal talk is possible only within a coherent ontological framework of such states of affairs.</Paragraph>
      <Paragraph position="4"> Many researchers ((Blaheta and Charniak 2000), (Gildea and Jurafsky 2000), showed that lexical and syntactic information is very useful for predicate-argument recognition tasks, such as semantic roles. However, lexical and syntactic information alone is not sufficient for the detection of complex semantic relations, such as CAUSE.</Paragraph>
      <Paragraph position="5"> Based on these considerents and on our observations of the English texts, we selected a list of 19 features which are divided here into two categories: lexical and semantic features.</Paragraph>
      <Paragraph position="6"> The lexical feature is represented by the causation verb in the pattern considered. As verb senses in WordNet are fine grained providing a large list of semantic hierarchies the verb can belong to, we decided to use only the lexical information the verb provides. The values of this feature are represented by the 60 verbs detected with the procedure described in section 4.1. This feature is very important, as our intention here is to capture the semantic information brought by the verb in combination with the subject and object noun phrases that attach to it. As we don't use word sense disambiguation to disambiguate each noun phrase in context, we have to take into consideration all the WordNet semantic hierarchies they belong to according to each sense. For each noun phrase representing the cause, and respectively the effect, we used as semantic features the 9 noun hierarchies in WordNet: entity, psychological feature, abstraction, state, event, act, group, possession, and phenomenon. Each feature is true if it is one of the semantic classes the noun phrase can belong to, and false otherwise.</Paragraph>
      <Paragraph position="7">  Input: positive and negative causal examples Output: lexical and semantic constraints Step 1. Generalize the training examples Initially, the training corpus consists of examples that contain only lexical features in the following format: a0 cause NP; verb; effect NP; targeta17 , where target can be either &amp;quot;Yes&amp;quot; or &amp;quot;No&amp;quot;, depending whether or not an example encodes cause.</Paragraph>
      <Paragraph position="8"> For example, a0 earthquake; generate; Tsunami; Yesa17 indicates that between the noun &amp;quot;earthquake&amp;quot; and the noun &amp;quot;Tsunami&amp;quot; there is a cause relation.</Paragraph>
      <Paragraph position="9"> From this intermediate corpus a generalized set of training examples was built, by expanding each intermediate example with the list of semantic features using the following format:</Paragraph>
      <Paragraph position="11"> For instance, the initial example becomes a0 f, f, f, f, f, f, f, f, t, generate, f, f, f, f, f, t, f, f, f, yesa17 , as the noun phrase earthquake belongs only to the a0 phenomenona1 noun hierarchy and the noun phrase Tsunami is only in the a0 eventa1 noun hierarchy in WordNet.</Paragraph>
      <Paragraph position="12"> Step 2. Learning constraints from training examples For the examples in the generalized training corpus (those that are either positive or negative), constraints are determined using C4.5. In this context, the features are the characteristics that distinguish the causal relation, and the values of the features are either specific words (e.g., the verb) or their Word-Net corresponding semantic classes (the furthest ancestors in WordNet of the corresponding concept).</Paragraph>
      <Paragraph position="13"> On this training corpus we applied C4.5 using a 10-fold cross validation. The output is represented by 10 sets of rules generated from the positive and negative examples.</Paragraph>
      <Paragraph position="14"> The rules in each set were ranked according to their frequency of occurrence and average accuracy obtained for that particular set. In order to use the best rules, we decided to keep only the ones that had a frequency above a threshold (occur in at least 7 of the 10 sets of rules) and with an average accuracy greater than 60a0 .</Paragraph>
      <Paragraph position="15">  Table 2 summarizes the constraints learned by the program.</Paragraph>
      <Paragraph position="16"> As we can notice, the constraints combine information about the semantic classes of the noun phrases representing the cause and effect with the lexical information about the verb.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>