<?xml version="1.0" standalone="yes"?>
<Paper uid="W03-0611">
  <Title>Learning the Meaning and Usage of Time Phrases from a Parallel Text-Data Corpus</Title>
  <Section position="8" start_page="0" end_page="0" type="evalu">
    <SectionTitle>
6 Current and Future Work
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.1 Verb Choice
</SectionTitle>
      <Paragraph position="0"> We would like to use our corpus to learn choice rules for verbs which are near-synonyms (Edmonds and Hirst, 2002). We are currently attempting to learn rules which predict which of three possible verbs - decreasing, easing, and falling - are used when the wind speed decreases. null We have conducted two experiments. The first was a semantic analysis, where we attempted to learn a choice rule based on features extracted from the numerical data. To do this, we used our aligned corpus to extract semantic features which we thought could be relevant to this decision (such as the amount by which the wind speed has decreased), and then analysed this with Ripper (Cohen, 1995). This gave the rules shown in Figure 2; these  word is variable, by forecaster (mode in bold) casters. These rules are mildly effective; 10-fold cross validation error is 25%, compared to a baseline error rate of 33% from always choosing the most common verb (easing). These rules suggest that at least for some forecasters, decreasing suggests a larger change in the wind speed than easing; this is the sort of near-synonym connotational difference that we expected to find. More surprisingly (at least to us), the presence of forecast date in some of the rules suggests that forecasters change how they write over time. Perhaps in retrospect this should not have been a surprise, because we have also observed changes over time in how people write in a previous project (Reiter et al., 2000).</Paragraph>
      <Paragraph position="1"> We also analysed collocation effects, that is whether we could predict verb choice based purely on the words immediately preceding and following the verb (and hence ignoring the numerical prediction data). This was done on the complete corpus (not just verbs that were part of successfully aligned phrases). It is difficult to directly compare the collocation analysis with the semantic one due to differences in the corpora texts used, but in general terms the reduction in baseline error rate seems comparable to the semantic analysis. Some of the collocation effects were both strong and forecaster-dependent. For example, Table 7 shows the choice of wind decrease verb when the word following the verb was variable (indicating wind direction was variable). In this context, forecasters F1 and F3 usually used falling; F2 always used easing; and F4 always used decreasing (F5 never used variable in his forecasts). Similar individual differences were observed in other collocations. For example, when the word preceding the verb was gradually, F3, F4, and F5 preferred decreasing, but F2 always used easing (F1 never used gradually in his forecasts).</Paragraph>
      <Paragraph position="2"> In summary, it seems that the choice between the near synonyms decreasing, easing, and falling depends on a2 semantics: how much the actual wind speed has changed; a2 collocation: immediately preceding and following words in the sentence; a2 author: which forecaster wrote this particular text; a2 date: when the text was written.</Paragraph>
      <Paragraph position="3"> time F1 F2 F3 F4 F5 all 0000 later later by late evening by midnight in evening later  0300 later soon soon soon tonight soon 0600 later overnight soon by morning later in period later 0900 soon soon soon by midday in morning soon 1200 by midday soon by midday by midday in morning by midday by by mid by mid early by mid 1500 afternoon soon afternoon afternoon afternoon afternoon 1800 by evening by evening by late afternoon by evening by evening by evening in evening/ later/ 2100 later later by evening later by evening by evening bold font means this phrases was at least twice as common as the second-most common term. X/Y means X and Y were equally common  All of these factors are important, and in particular the kind of semantic differences investigated by (Edmonds and Hirst, 2002) are only one factor among many, and do not dominate the choice decision. We plan to continue working on this and other analyses of near-synonyms, and obtain a better idea of how these factors interact.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.2 Other corpora
</SectionTitle>
      <Paragraph position="0"> In addition to the weather corpus, the SUMTIME project also has access to a parallel text-data corpus of doctors describing signal data from a neonatal intensive care unit (Alberdi et al., 2001). We would like to analyse this corpus to determine the meanings of words such as steady and oscillations. However, a preliminary analysis has suggested to us that we cannot conduct such an analysis until we remove non-physiological events from the data (Sripada et al., 2003a). For example, a doctor may describe a signal as steady even when it contains a large spike, if the doctor believes the spike is due to a non-physiological event (such as a sensor falling off the baby and then being replaced by a nurse). Hence non-physiological events (known in this domain as 'artifacts') must be removed from the data before it is possible to analyse word meaning. We are currently working on artifact removal, and once this is complete we will start our analysis of word meanings.</Paragraph>
      <Paragraph position="1"> SUMTIME is also working on generating textual summaries of gas turbine sensor data (Yu et al., 2003). Unfortunately in this domain, as in many other NLG applications (Reiter et al., 2003), there is no existing corpus of manually written texts describing the input data. We have explicitly asked two experts to write descriptions of 38 signal fragments. This very small corpus showed that once again there were major differences between individuals (Reiter and Sripada, 2002a), but the corpus is too small to allow meaningful statistical analysis of word meanings.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>