<?xml version="1.0" standalone="yes"?>
<Paper uid="W98-1216">
  <Title>An Attempt to Use Weighted Cusums to Identify Sublanguages</Title>
  <Section position="6" start_page="0" end_page="0" type="concl">
    <SectionTitle>
6 Discussion
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.1 Does the WQsum test identify
</SectionTitle>
      <Paragraph position="0"> different sublanguages? Let us consider first the results as shown in Table 6. Our main concern of course is to see whether the WQsurn test can identify and distinguish the groups. Taking 1.65 as our threshold, we can rank the groups according to the number of other groups with which each gets a pairwise score average above this threshold. In addition, since the 'ideal' situation as far as our hypothesis goes would be for a low within-group average score suggesting homogeneity, and a high average score for comparisons with other groups, suggesting distinctiveness, as a further, informal measure of the extent to which the groups meet this condition, we can divide the average comparison score by the within-group score. Table 7 shows a ranking of the groups along these lines.</Paragraph>
      <Paragraph position="1"> The groups seem to divide into roughly four types.</Paragraph>
      <Paragraph position="2"> The first type, groups which support our hypothesis the best, have a low within-group average, a high pairwise average, and can easily be distinguished from most of the other groups. In this group are 'xwords', 'univs', 'blurbs', 'BMJ', and 'economy'. At the other end of the scale, at the bottom of Table 7, are those groups which can have a low 'Sig.' score. This group is not necessarily marked by a low pairwise average or a high within-group score: the 'obits' group for example has the second lowest within-group average, and scores quite highly on our informal ratio score. Yet the WQsum test cannot distinguish it from six of the other groups.</Paragraph>
      <Paragraph position="3"> A third type is where the 'Sig.' score is high despite a high within-group average which would suggest lack of homogeneity. The 'recipes' group, for example, stands out as a distinct sublanguage, with highly significant scores compared to all other groups. Despite the fact that the within-group score is above the Somers 135 Use Weighted Cusums to Idenn'fy Sublanguages</Paragraph>
      <Paragraph position="5"> pairwise averages above the 1.65 threshold ('Sig.'), and secondly according to the informal score described in the text. An asterisk indicates a group where the within-group average is above the 1.65 threshold.</Paragraph>
      <Paragraph position="6">  1.65 threshold, suggesting lack of homogeneity among the recipes, the average of the scores for pairwise comparisons with other groups is sufficiently high to compensate this: as Table 6 shows, the average scores for recipes are consistently high, and ot~en the highest in any row. This can be contrasted with the case of the 'church' group, where the within-group average is below the 1.65 threshold, but so are nearly half the scores for pairwise comparisons. But the situation can also be contrasted with the 'TVseripts' and 'childrens' groups: pairwise scores with all the other grofaps indicate significant differences, but so does the within-group average. This means that each TV script or children's story seems significantly different from all the other samples, including the other TV scripts or children's stories. For the 'tourism' group, too, the scores for pairwise comparison are about the same as the within-group score. It so happens that these scores are a bit nearer the threshold, so we get a 10-4 'Sig.' score rather than 14--0, but the conclusion is the same: the WQsum earmot distinguish these sublanguages.</Paragraph>
      <Paragraph position="7"> Finally we have the case of the 'lawreps' and 'emails', which are internally homogenous, and can be distinguished from some, but not all of the other groups. Let us now summarize these observations, and categorize the four types: A Good result. Homogeneous and distinctive sublanguage: 'xwords', 'univs', 'blurbs', 'BMJ', and 'economy'.</Paragraph>
      <Paragraph position="9"> This suggests that the WQsum test is able to quantify the similarity of individual groups, as well as to distinguish sublanguages. In this experiment we have taken groups of texts and compared them, but in fact the WQsum algorithm is designed to work on the basis of individual texts. In principle, we could simply take a pair of texts and use the algorithm to determine to what extent they are the same sublanguage. It must be said however that it seems to make more sense to use the test in the comparative manner illustrated here, for example comparing three texts to see which pair is most similar. It also seems important to have a baseline score for an established group of texts belonging to the same sublanguage.</Paragraph>
    </Section>
    <Section position="2" start_page="0" end_page="0" type="sub_section">
      <SectionTitle>
6.2 Reservations and future directions
</SectionTitle>
      <Paragraph position="0"> A very short time before the final version of this paper was due to be delivered, a further possibility came to our notice. Tweedie &amp; Donnelly (1996) describe an alternative, multivariate test using weighted eusurns to compare more than two texts. Although we have not had a chance to study this proposal, it claims to give more accurate results than the pairwise application of the WQsum formula that has been reported in this paper. An obvious next step is to try their proposal.</Paragraph>
      <Paragraph position="1"> Also, a further step that we might take would be to answer the criticism that the scale of our investigation is too small. The fact that we have taken only three 25sentence samples of each sublanguage obviously means Use Weighted Cusums to ldentify Sublanguages  that our conclusions must be somewhat limited. An anonymous reviewer commented that &amp;quot;the texts were so different, that it shouldn't be hard at all to discriminate between them&amp;quot;. The results in Tables 4 and 6 show that this is not the ease at all: the groups that the test failed to distinguish are not necessarily those which to the human eye are most similar (see Table 8), nor are the successfully identified groups necessarily the most dissimilar. Perhaps this finding is not so surprising when we consider that the linguistic features that are used in the test are so superficial: there is no reason to expect that the incidence of words beginning with a vowel, for example, would correlate highly with sublanguage type. And therein lies the real interest of this technique: because the linguistic features are superficial, it seems that there is no intuition that we can appeal to here.</Paragraph>
      <Paragraph position="2"> Finally, throughout this paper we have referred to 'sublanguage', and the possibility that bur WQsum algorithm can identify different sublanguages. It seems that the algorithm can distinguish texts, but it is by no means clear what aspect of their difference it is capturing. It could for example be merely genre, or some other aspect of sublanguage, that it is capturing though again intuitions are difficult to appeal to because of the superficiality of the linguistic features used. We need to look more closely at the differences between the text pairs it fails to distinguish and those where it succeeds, in order to try to get a feel for what, exactly, the test is capturing. Nevertheless, we feel that it is an interesting avenue to explore, the more so as it seems to be quite unlike the other methods described in this field.</Paragraph>
    </Section>
  </Section>
class="xml-element"></Paper>