<?xml version="1.0" standalone="yes"?>
<Paper uid="W02-0702">
  <Title>Topic Detection Based on Dialogue History</Title>
  <Section position="3" start_page="0" end_page="0" type="metho">
    <SectionTitle>
2 Topic detection
</SectionTitle>
    <Paragraph position="0"> Among conventional topic detection methods, one uses compound words that features certain topic as trigger information for detecting a topic (Hatori et al., 2000), and another uses domain-dependant dictionaries and thesauruses to construct knowledge applicable to a certain topic (Tsunoda et al., 1996). In the former method, a scene-dependant dictionary provides the knowledge relevant to the scene and compound words in the dictionary are used for detecting a topic. In the latter method, words appearing in a scene are defined as the knowledge relevant to the scene and superordinate/subordinate relation and synonyms provided by thesauruses are used to enhance the robustness.</Paragraph>
    <Paragraph position="1"> These conventional methods are suitable for written texts but not for dialogue utterances in a speech translation system. The following two major constraints make the topic detection for dialogue utterances more difficult.</Paragraph>
    <Paragraph position="2">  (1) Constraint due to single sentence process - Sentences in a dialogue are usually short with few keywords.</Paragraph>
    <Paragraph position="3"> - In a dialogue, the frequency values of the word in a sentence are mostly one, making it difficult to apply a statistical method.</Paragraph>
    <Paragraph position="4"> (2) Constraint due to the nature of spoken dialogue - In a dialogue, one topic is sometimes expressed with two or more sentences.</Paragraph>
    <Paragraph position="5"> - The words appearing in a sentence are sometimes replaced by anaphora or omitted by ellipsis in the next sentence.</Paragraph>
    <Paragraph position="6"> - Topics frequently change in a dialogue.</Paragraph>
    <Paragraph position="7"> On the other hand, a speech translation system requires the following: - Topic detection for each utterance in a dialogue; - Prompt topic detection in real time processing; - Dynamic tracking of topic transition.</Paragraph>
    <Paragraph position="8">  To make topic detection adaptive to the speech translation system, we propose a method applicable to one utterance in a dialogue as an input, which can be used for tracking the topic transitions dynamically and outputting most appropriate topic for the latest utterance. The k-nearest neighbor method (Yang, 1994) is used with the clustering method linked with the dialogue history as a topic detection algorithm for dialogue utterance. The k-nearest neighbor method is known to have high precision performance with less restriction in the field of document categorization. This method is frequently used as a baseline in the field and also applied to topic detection for story but not for a single sentence (Yang et al., 1999). This paper incorporates two new methods to the k-nearest neighbor method to overcome two constraints mentioned above.</Paragraph>
    <Paragraph position="9"> To overcome the first constraint, we cluster a set of sentences in training data into subsets (called subtopics) based on similarity between the sentences. A topic is detected by calculating the relevance between the input sentence and these subtopics. Clustering sentences on the same subtopic increases number of characteristic words to be compared with input sentence in calculation.</Paragraph>
    <Paragraph position="10"> To overcome the second constraint, we group an input sentence with other sentences in the dialogue history. A topic is detected by calculating the relevance between this group and each possible topic. Grouping the input sentence with the preceding sentences increases number of characteristic words to be compared with topics in calculation. We consider the order of the sentences in the dialogue in calculating the relevance to avoid the influence of topic change in the dialogue.</Paragraph>
  </Section>
  <Section position="4" start_page="0" end_page="0" type="metho">
    <SectionTitle>
3 Topic detection algorithm
</SectionTitle>
    <Paragraph position="0"> This section explains three methods used in the proposed topic detection algorithm: 1) k-nearest neighbor method, 2) the clustering method using TF-IDF, and 3) the application of the dialogue history.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.1 k-nearest neighbor method
</SectionTitle>
      <Paragraph position="0"> We denote the character vector for a given sentence in the training data as D j , and that for a given input sentence as X. Each vector has a TF-IDF value of the word in the sentence as its element value (Salton 1989).</Paragraph>
      <Paragraph position="1"> The similarity between the input sentence X and the training data D j is calculated by taking the inner product of the character vectors. The conditional probability of topic C  The relevance score between the input sentence X and each topic C l is calculated as the sum of similarity for k sentences taken from the training data in descending order of similarity.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.2 Topics clustering method
</SectionTitle>
      <Paragraph position="0"> This method clusters topics into smaller subtopics. The word &amp;quot;topic&amp;quot; used in this method consists of several subtopics representing detailed situations. The topic &amp;quot;Hotel&amp;quot; consists of subtopics such as &amp;quot;Checking In&amp;quot; and &amp;quot;Room Service&amp;quot;. Sentences in training data categorized under the same topic are further grouped into subtopics based on their similarity.</Paragraph>
      <Paragraph position="1"> Calculating the relevance between the test data input and these subsets of training data provides more keywords in detecting topics. Our method to create the subtopics identifies a keyword in a sentence set, and then recursively divides the set into two smaller subsets, one that includes the keyword and one that does not.</Paragraph>
    </Section>
  </Section>
  <Section position="5" start_page="0" end_page="0" type="metho">
    <SectionTitle>
TF-IDF Clustering Method
</SectionTitle>
    <Paragraph position="0"> (1) Find the word that has the highest TF-IDF value among the words in the sentence set; (2) Divide the sentence set into two subsets; one that contains the word obtained in step (1) and one that does not; (3) Repeat steps (1) and (2) recursively until TF-IDF value reaches the threshold.</Paragraph>
    <Paragraph position="1">  Subtopics created by this method consist of keywords featuring each subtopic and their related words.</Paragraph>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
3.3 Application of the dialogue history
</SectionTitle>
      <Paragraph position="0"> The proposed method applies the dialog history in topic detection. The method interprets a current input sentence and the sentences prior to the current input as a dialogue history subset, and detects topics by calculating the relevance score between the dialogue history subset and the each topic. The method increases number of keywords in the input for calculation. We assign a weight to each sentence in the dialogue history subset to control the effect of time-sequence in sentences.</Paragraph>
      <Paragraph position="1"> The relevance score combined with the dialog history is calculated as:</Paragraph>
      <Paragraph position="3"> Here the similarity is calculated with the input sentence X and the sentence in the dialog history</Paragraph>
      <Paragraph position="5"> as the weights for the input sentences and the sentences in the dialogue history, respectively.</Paragraph>
    </Section>
  </Section>
  <Section position="6" start_page="0" end_page="22" type="metho">
    <SectionTitle>
4 Evaluation
</SectionTitle>
    <Paragraph position="0"> To evaluate the proposed method, we prepared training data and test data from a travel conversation corpus. We also prepared three</Paragraph>
    <Paragraph position="2"> types of clusters with different thresholds and two types of dialogue history with different weight values.</Paragraph>
    <Section position="1" start_page="22" end_page="22" type="sub_section">
      <SectionTitle>
4.1 Training data
</SectionTitle>
      <Paragraph position="0"> In the evaluation, we used approximately 25,000 sentences from our original travel conversation corpus as our training data. The sentences are manually classified into four topics: 1) Hotel, 2) Restaurant, 3) Shopping, and 4) Others. The topic &amp;quot;Others&amp;quot; consists of sentences not categorized into the remaining three. Topics such as &amp;quot;Transportation&amp;quot; or &amp;quot;Illnesses and injuries&amp;quot; are placed into this &amp;quot;Others&amp;quot; in this evaluation.</Paragraph>
    </Section>
    <Section position="2" start_page="22" end_page="22" type="sub_section">
      <SectionTitle>
4.2 Test data
</SectionTitle>
      <Paragraph position="0"> We prepared two sets of test data. One set consists of 62 typical travel dialogues comprising 896 sentences (hereafter called &amp;quot;typical dialogue data&amp;quot;). The other set consists of 45 dialogues comprising 498 sentences, which may include irregular expressions but closely representing daily spoken language (hereafter called &amp;quot;real situation dialogue data&amp;quot;). Sentences in &amp;quot;typical dialogue data&amp;quot; are often heard in travel planning and travelling situations, and form a variety of initiating dialogues as the travel conversation unfolds. The data includes words and phrases often used in the topics listed above, and each sentence is short with little redundancy. On the other hand, &amp;quot;real situation dialogue data&amp;quot; consists of spoken dialogue phrases which are likely to appear in user-specific situations in the travel domain.</Paragraph>
      <Paragraph position="1"> Some phrases may be typically used, while others may consist of colloquial expressions and words and phrases that are redundant. Some of the words may not appear in the training data.</Paragraph>
    </Section>
    <Section position="3" start_page="22" end_page="22" type="sub_section">
      <SectionTitle>
4.3 Clustering the topics
</SectionTitle>
      <Paragraph position="0"> We applied the clustering with the aforementioned method to 8,457 sentences from training data which are categorized into one or more of the three topics: 1) Hotel, 2) Restaurant, and 3) Shopping. Clusters are created on three different thresholds: 8,409 clusters (small-sized cluster), 3,845 clusters (medium-sized cluster) and 2,203 clusters (large-sized cluster). In carrying out clustering, we set one sentence as one cluster if the sentence does not contain a word whose TF-IDF value is not equal to or greater than the threshold. We excluded data that falls only under the topic &amp;quot;Others&amp;quot; and data that falls under all four topics, which are considered to be general conversation.</Paragraph>
      <Paragraph position="1"> Variations of these topics produce 13 probable combinations.</Paragraph>
      <Paragraph position="2"> The number of clusters is smallest (13) when we set one topic as one cluster and largest (8,457) when we set one sentence as one cluster.</Paragraph>
    </Section>
    <Section position="4" start_page="22" end_page="22" type="sub_section">
      <SectionTitle>
4.4 Use of the dialogue history
</SectionTitle>
      <Paragraph position="0"> To evaluate the effect of the dialogue history, we use an input sentence, the most preceding and the next preceding sentence (hereafter &amp;quot;sentence 0&amp;quot;, &amp;quot;sentence -1&amp;quot;, and &amp;quot;sentence -2&amp;quot;) as a dialogue history. Two types of sentence weights are applied to these three sentences, one of equal weights and one of weights based on a time series. These sets are:</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>