Mining Text Conversations - Methods for Mining and Summarizing Text Conversations - page 70

responsible for the task. In this case, Baba and David Jordan are the owners (the you in Sentence 2
refers back to Baba and David Jordan in Sentence 1).
As with decision sentences, action items are more common in meetings and emails than
in blogs and discussion fora, for the same reasons. Here, we describe work on those two types of
conversations.
In the preceding section, we described how the AMI and ICSI abstract summaries contain
decision subsections, and that one can exploit the abstract-extract links to determine which utterances
from the meeting are relevant to decisions. The AMI corpus also contains action subsections, and
one can similarly exploit the abstract-extract links to obtain a gold-standard labeling of meeting
utterances concerning action items.
Murray and Renals [ 2008 ] train logistic regression classifiers on AMI data labeled in such a
manner, using prosodic, lexical, structural, length and speaker features. On their test set, the highest
AUROC score was 0.93 (out of 1), indicating that these features are very effective for identifying
action items utterances. Structural features alone matched the performance achieved when using all
features, the simple explanation being that action items are much more likely to occur at the end of
a meeting. Lexical features were also effective; note that both Sentences 1 and 2 above are signaled
by the pattern you will , and such cues are common.
Purver et al. [ 2006a ] take a finer-grained approach to action item detection, using the ICSI
and CALO corpora. Rather than just classifying sentences as being related to an action item, they
aim to detect subclasses of action item utterances. The four classes are description , time-frame , owner
and agreement . Owner and time-frame are described above. Description utterances are those that description

describe the task to be carried out. Agreement utterances are those that indicate acceptance or

agreement

agreement with an assigned action item. The authors found that this finer-grained classification
is quite challenging, with the Agreement class being the easiest to detect of the four (F-score of
0.40), while the owner class being the most difficult (F-score of 0.17). Ownership of action items
in meetings is so difficult because the person doing the delegating may not explicitly mention to
whom the item is being assigned; rather they will simply speak directly to that person. In that sense,
ownership detection is closely related to the addressing problem of determining which speakers are
addressing one another. In subsequent work, Purver et al. [ 2007 ] investigate summarization of action
items by extraction of useful phrases such as those including explicit time-frames, and find that this
condensing stage can in some cases yield more informative results than simply giving the utterance
transcriptions.
Using such fine-grained analysis of action items, we could extract useful, specific information
from Sentences 1 and 2 above, as shown in Figure 3.13 . We can now present the user with descriptions
of the action items, as well as the owners and time-frames, and index the sentences from which these
pieces of information were taken.
On email data, Bennett and Carbonell [ 2005 ] aim to classify each message according to
whether the sender requires a response of some kind, as well as identifying the particular sentences
that mention the action items. They employ several types of supervised classifiers and compare