<?xml version="1.0" standalone="yes"?>
<Paper uid="A00-2019">
  <Title>An Unsupervised Method for Detecting Grammatical Errors</Title>
  <Section position="6" start_page="145" end_page="146" type="concl">
    <SectionTitle>
5 Conclusion
</SectionTitle>
    <Paragraph position="0"> The unsupervised techniques that we have presented for inferring negative evidence are effective in recognizing grammatical errors in written text.</Paragraph>
    <Paragraph position="1">  Preliminary results indicate that ALEK's error detection is predictive of TOEFL scores. If ALEK accurately detects usage errors, then it should report more errors in essays with lower scores than in those with higher scores. We have already seen in Table 1 that there is a negative correlation between essay score and two of ALEK's component measures, the general corpus n-grams. However, the data in Table 1 were not based on specific vocabulary items and do not reflect overall system performance, which includes the other measures as well.</Paragraph>
    <Paragraph position="2"> Table 6 shows the proportion of test word occurrences that were classified by ALEK as containing errors within two positions of the target at each of 6 TOEFL score points. As predicted, the correlation is negative (rs = -1.00, n = 6, p &lt; .001, two-tailed). These data support the validity of the system as a detector of inappropriate usage, even when only a limited number of words are targeted and only the immediate context of each target is examined.</Paragraph>
    <Paragraph position="3">  score point, classified as containing an error by ALEK and by a human judge For comparison, Table 6 also gives the estimated proportions of inappropriate usage by score point based on the human judge's classification. Here, too, there is a negative correlation: rs = -.90, n = 5, p &lt; .05, two-tailed.</Paragraph>
    <Paragraph position="4"> Although the system recognizes a wide range of error types, as Table 6 shows, it detects only about one-fifth as many errors as a human judge does. To improve recall, research needs to focus on the areas identified in section 3.2 and, to improve precision, efforts should be directed at reducing the false positives described in 3.3. ALEK is being developed as a diagnostic tool for students who are learning English as a foreign language. However, its techniques could be incorporated into a grammar checker for native speakers.</Paragraph>
  </Section>
class="xml-element"></Paper>