Introduction - Methods for Mining and Summarizing Text Conversations - page 12

the task of detecting whether a contribution to the conversation is subjective or not. This can be
further refined by determining whether a contribution that has been identified to be subjective
expresses a positive vs. a negative sentiment. Finally, at an even more specific level, you may want
to determine the strength of the expressed opinion, for instance in the interval [-3, +3], with a zero
value for non-subjective (i.e., neutral) contributions. For illustration, if the sentence “I applaud the
new budget proposal.” was considered, it should be classified as expressing a positive opinion, with
strength +3. And if we look at our sample email conversation, sentences like “great idea!” , “I do not like
this assignments” , “Let's talk tomorrow at UBC after class” should be, respectively, classified as positive
(+3), negative (-2) and neutral (0).
Numerous techniques have been proposed in the literature to perform opinion mining from
generic text [ Pang and Lee , 2008 ]. Most of these techniques rely on lexical and syntactic features of
the text, meaning that they look for the presence (or absence) in the target text of particular words,
as well as of particular syntactic constructions. For instance, two features commonly used are the
presence of adjectives with a positive or negative orientation (e.g., interesting vs. boring ) and the
presence of syntactic patterns involving negation like not < intensifier >< adjective > , which
would match the phrase not very inspiring . Notice that machine learning methods are often used,
in order to determine what features are most useful for opinion mining, out of the large number of
candidates (e.g., [ Wilson et al. , 2006 ]).
In Chapter 3 , we will see how approaches to opinion mining, based on lexical and syntactic
features, can be directly applied to text conversations. That chapter will also discuss how the set
of features can be expanded to include conversational features. For instance, features related to
the speaker of the particular contribution, as well as the position of the contribution within the
conversation.

Extracting the Conversational Structure Conversations have properties that clearly distinguish
them from monologues. To have a conversation, you need two or more participants who exchange
information by talking or in writing. Each contribution to a conversation, technically each turn,
is performing one or more dialogue acts; for instance, making a statement, asking or answering dialogue

acts

a question, and making or accepting a request. These dialogue acts tend to occur in pairs, called
adjacency pairs , where the first turn from one participant is generating the following turn by another adjacency

pairs

participant. Common examples of these pairs are, for example, a question followed by an answer and
a request followed by and acceptance or a rejection. For illustration, in our sample email conversation
the request expressed in Email-1 “Would you like to join me?” is followed by an uncertain response in
Email-1.1, and by reject response in Email-1.2.
The structure of a conversation specifies how all the dialogue acts that comprises the con-
versation are connected to each other. In synchronous conversation like spoken ones and instant
messaging, the structure of the conversation can be expected by and large to be linear, as turns must
occur one after the other, with minimal delay and minimal overlap. In contrast, asynchronous written
conversations, like emails and blogs, often display a more complex structure, as consecutive turns