Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 71

Figure 3.13: Extracting action item information.
unigram features with higher-order n-grams, concluding that higher-order n-grams are superior for
these tasks. The best reported F-score for document-level classification is 0.78.
Corston-Oliver et al. [ 2004 ] also aim to detect action items in emails in order to create an
email “to-do ” list, but with a much richer set of features and larger corpus. They train SVM classifiers
using three feature types: message features such as the number of recipients and the message size,
superficial features such as n-grams, names and special characters, and linguistic features including
part-of-speech n-grams and features of logical form. Like Purver et al. [ 2007 ], Corston-Oliver et
al. subsequently try to summarize the action items in order to create a succinct to-do list. This
post-processing step involves reformulating the sentences by identifying the clauses containing the
task, deleting extraneous words, replacing certain deictic expressions with non-deictic expressions,
and replacing all temporal expressions with absolute dates.
A very recent example of work on detecting action items in email is Lampert et al. [ 2010 ],
where the authors show that determining whether an email contains a request for action can benefit
from an initial segmentation of the email into nine functional zones. These zones include: content
written by the current sender, greetings, sign offs, quoted reply content, forwarded content, etc. A
competitive F-score of 0.84 is reported for experiments run on an annotated subset of the Enron
corpus.
Both action item detection and decision classification can be considered a type of focused
summarization, and we discuss summarization in much more detail in Chapter 4 . For now, we
simply note that decisions and action items are the types of information that many people will desire
to know about when they have missed a meeting or want to digest a long email conversation. They
want to know what decisions have been made, and what they are required to do.