<?xml version="1.0" standalone="yes"?>
<Paper uid="I05-2030">
  <Title>Opinion Extraction Using a Learning-Based Anaphora Resolution Technique</Title>
  <Section position="4" start_page="173" end_page="175" type="metho">
    <SectionTitle>
3 Method for opinion extraction
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="173" end_page="174" type="sub_section">
      <SectionTitle>
3.1 Analogy with anaphora resolution
</SectionTitle>
      <Paragraph position="0"> We consider the task of extracting opinion tuples &lt;Product, Attribute, Value&gt; from review sites and message boards on the Web dedicated to providing and exchanging information about retail goods. On these Web pages, products are often specified clearly and so it is frequently a trivial job to extract the information for the Product slot.</Paragraph>
      <Paragraph position="1"> We therefore in this paper focus on the problem of extracting &lt;Attribute, Value&gt; pairs.</Paragraph>
      <Paragraph position="2"> In the process of attribute-value pair identification for opinion extraction, we need to deal with the following two cases: (a) both a value and its corresponding attribute appear in the text, and (b) a value appears in the text while its attribute is missing since it is inferable form the value expression and the context. The upper half of Figure 1 illustrates these two cases in the automobile domain. In (b), the writer is talking about the &amp;quot;size&amp;quot; of the car, but the expression &amp;quot;size&amp;quot; is not explicitly mentioned in the text. In addition, (b) includes the case where the writer evaluates the product itself. For example, &amp;quot;I'm very satisfied with my car!&amp;quot;: in this case, a value expression &amp;quot;satisfied&amp;quot; evaluates the product as a whole, therefore a corresponding attribute does not exists. null For the case (a), we first identify a value expression (like in Figure 1) in a given text and then look for the corresponding attribute in the text.</Paragraph>
      <Paragraph position="3"> Since we also see the case (b), on the other hand, we additionally need to consider the problem of whether the corresponding attribute of the identified value expression appears in the text or not. The structure of these problems is analogous to that of anaphora resolution; namely, there are exactly two cases in anaphora resolution that have a clear correspondence with the above two cases as illustrated in Figure 1: in (a) the noun phrase (NP) is anaphoric; namely, the NP's antecedent appears in the text, and in (b) the noun phrase is non-anaphoric. A non-anaphoric NP is either ex- null denotes the word sequence corresponding to the Attribute. Likewise, we also use &lt;&gt; v for the Value.</Paragraph>
      <Paragraph position="4">  and anaphora resolution ophoric (i.e. the NP has an implicit referent) or indefinite. While the figure shows Japanese examples, the similarity between anaphora resolution and opinion extraction is language independent.</Paragraph>
      <Paragraph position="5"> This analogy naturally leads us to think of applying existing techniques for anaphora resolution to our opinion extraction task since anaphora resolution has been studied for a considerably longer period in a wider range of disciplines as we briefly review below.</Paragraph>
    </Section>
    <Section position="2" start_page="174" end_page="174" type="sub_section">
      <SectionTitle>
3.2 Existing techniques for anaphora
</SectionTitle>
      <Paragraph position="0"> resolution Corpus-based empirical approaches to anaphora resolution have been reasonably successful. This approach, as exemplified by (Soon et al., 2001; Iida et al., 2003; Ng, 2004), is cost effective, while achieving a better performance than the best-performing rule-based systems for the test sets of MUC-6 and MUC-7  .</Paragraph>
      <Paragraph position="1"> As suggested by Figure 1, anaphora resolution can be decomposed into two subtasks: anaphoricity determination and antecedent identification. Anaphoricity determination is the task of judging whether a given NP is anaphoric or nonanaphoric. Recent research advances have provided several important findings as follows: * Learning-based methods for antecedent identification can also benefit from the use of linguistic clues inspired by Centering Theory (Grosz et al., 1995).</Paragraph>
      <Paragraph position="2"> * One useful clue for anaphoricity determination is the availability of a plausible candidate for the antecedent. If an appropriate candidate for the antecedent is found in the preceding discourse context, the NP is likely to be anaphoric.</Paragraph>
      <Paragraph position="3"> For these reasons, an anaphora resolution model performs best if it carries out the following pro- null tify the best candidate antecedent for it, and (2) Anaphoricity determination: Judge whether the candidate really stands for the true antecedent of the NP.</Paragraph>
    </Section>
    <Section position="3" start_page="174" end_page="175" type="sub_section">
      <SectionTitle>
3.3 An opinion extraction model inspired by
</SectionTitle>
      <Paragraph position="0"> analogy with anaphora resolution As illustrated in Figure 2, an opinion extraction model derived from the aforementioned analogy with anaphora resolution as follows:  1. Initialization: Identify attribute and value candidates by dictionary lookup 2. Attribute identification: Select a value and identify the best candidate attribute corresponding to the value 3. Pairedness determination: Decide whether the candidate attribute stands for the real attribute of the value or not (i.e. the value has no explicit corresponding attribute in the text) 4. Opinionhood determination: Judge whether the obtained attribute-value pair   expresses an opinion or not Here, the attribute identification and pairedness determination processes respectively correspond to the antecedent identification and anaphoricity determination processes in anaphora resolution. Note that our opinion extraction task requires an additional subtask, opinionhood determination -- an attribute-value pair appearing in a text does not necessarily constitute an opinion. We elaborate on the notion of opinionhood in section 4.1. From the above discussion, we can expect that the findings for anaphora resolution mentioned in 3.2 stated above apply to opinion extraction as well. In fact, the information about the candidate  For simplicity, we call a value both with and without an attribute uniformly by the term attribute-value pair unless the distinction is important.</Paragraph>
      <Paragraph position="1">  attribute is likely to be useful for pairedness determination. We therefore expect that carrying out attribute identification before pairedness determination should outperform the counterpart model which executes the two subtasks in the reversed order. The same analogy also applies to opinionhood determination; namely, we expect that opinion determination is bet performed after attribute determination. Furthermore, our opinion extraction model also can be implemented in a totally machine learning-based fashion.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="175" end_page="176" type="metho">
    <SectionTitle>
4 Evaluation
</SectionTitle>
    <Paragraph position="0"> We conducted experiments with Japanese Web documents to empirically evaluate the performance of our opinion extraction model, focusing particularly on the validity of the analogy discussed in the previous section.</Paragraph>
    <Section position="1" start_page="175" end_page="175" type="sub_section">
      <SectionTitle>
4.1 Opinionhood
</SectionTitle>
      <Paragraph position="0"> In these experiments, we define an opinion as follows: An opinion is a description that expresses the writer's subjective evaluation of a particular subject or a certain aspect of it.</Paragraph>
      <Paragraph position="1"> By this definition, we exclude requests, factual or counter-factual descriptions and hearsay evidence from our target opinions. For example, The engine is powerful is an opinion, while a counter-factual sentence such as If only the engine were more powerful is not regarded as opinion.</Paragraph>
    </Section>
    <Section position="2" start_page="175" end_page="175" type="sub_section">
      <SectionTitle>
4.2 Opinion-tagged corpus
</SectionTitle>
      <Paragraph position="0"> We created an opinion-tagged Japanese corpus consisting of 288 review articles in the automobile domain (4,442 sentences). While it is not easy to judge whether an expression is a value or an attribute, we asked the annotator to identify attribute and value expressions according to their subjective judgment.</Paragraph>
      <Paragraph position="1"> If some attributes are in a hierarchical relation with each other, we asked the annotator to choose the attribute lowest in the hierarchy as the attribute of the value. For example, in a sound system with poor sound, only sound is annotated as the attribute of the value poor.</Paragraph>
      <Paragraph position="2"> The corpus contains 2,191 values with an attribute and 420 values without an attribute. Most of the attributes appear in the same sentence as their corresponding values or in the immediately preceding sentence (99% of the total number of pairs). Therefore, we extract attributes and their corresponding values from the same sentence or from the preceding sentence.</Paragraph>
    </Section>
    <Section position="3" start_page="175" end_page="176" type="sub_section">
      <SectionTitle>
4.3 Experimental method
</SectionTitle>
      <Paragraph position="0"> As preprocessing, we analyzed the opinion-tagged corpus using the Japanese morphological analyzer ChaSen  and the Japanese dependency structure analyzer CaboCha</Paragraph>
      <Paragraph position="2"> models for attribute identification, pairedness determination and opinionhood determination. We used the 2nd order polynomial kernel as the kernel function for SVMs. Evaluation was performed by 10-fold cross validation using all the data.</Paragraph>
      <Paragraph position="3">  We use dictionaries for identification of attribute and value candidates. We constructed a attribute dictionary and a value dictionary from review articles about automobiles (230,000 sentences in total) using the semi-automatic method proposed by Kobayashi et al. (2004). The data used in this process was different from the opinion-tagged corpus. Furthermore, we added to the dictionaries expressions which frequently appearing in the opinion-tagged corpus. The final size of the dictionaries was 3,777 attribute expressions and 3,950 value expressions.</Paragraph>
      <Paragraph position="4">  To examine the effects of appropriately choosing the order of model application we mentioned in the previous section, we conducted four experiments using different orders (AI indicates attribute identification, PD indicates pairedness determination and OD indicates opinion determination): null  Note that Proc.4 is our proposed ordering.</Paragraph>
      <Paragraph position="5"> In addition to these models, we adopted a base-line model. In this model, if the candidate value and a candidate attribute are connected via a dependency relation, the candidate value is judged to have an attribute. When none of the candidate attributes have a dependency relation, the candidate value is judged not to have an attribute. We adopted the tournament model for attribute identification (Iida et al., 2003). This model implements a pairwise comparison (i.e. a match) between two candidates in reference to the given value treating it as a binary classification problem, and conducting a tournament which consists of a series of matches, in which the one that prevails through to the final round is declared the  winner, namely, it is identified as the most likely candidate attribute. Each of the matches is conducted as a binary classification task in which one or other of the candidate wins.</Paragraph>
      <Paragraph position="6"> The pairedness determination task and the opinionhood determination task are also binary classification tasks. In Proc.1, since pair identification is conducted before finding the best candidate attribute, we used Soon et al.'s model (Soon et al., 2001) for pairedness determination. This model picks up each possible candidate attribute for a value and determines if it is the attribute for that value. If all the candidates are determined not to be the attribute, the value is judged not to have an attribute. In Proc.4, we can use the information about whether the value has a corresponding attribute or not for opinionhood determination. We therefore create two separate models for when the value does and does not have an attribute.</Paragraph>
      <Paragraph position="7">  We extracted the following two types of features from the candidate attribute and the candidate value: (a) surface spelling and part-of-speech of the target value expression, as well as those of its dependent phrase and those in its depended phrase(s) (b) relation between the target value and candidate attribute (distance between them, existence of dependency, existence of a co-occurrence relation) We extracted (b) if the model could use both the attribute and the value information. Existence of a co-occurrence relation is determined by reference to a predefined co-occurrence list that contains attribute-value pair information such as &amp;quot;height of vehicle - low&amp;quot;. We created the list from the 230,000 sentences described in section 4.3.1 by applying the attribute and value dictionary and extracting attribute-value pairs if there is a dependency relation between the attribute and the value. The number of pairs we extracted was about 48,000.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>