Summarization and Personal Information Management Carolyn Penstein Ros

Summarization and Personal Information Management Carolyn Penstein Rosé Language Technologies Institute/ Human-Computer Interaction Institute

Today’s topic…. Summarizing “nasty spoken dialogues”

Announcements Questions? n Plan for Today n ¨ Discussion from group activity ¨ Zechner paper ¨ Student presentation by Nguyen Bach

New Direction: Conversation Summarization Why is conversation summarization different? n Not just about what content was communicated – it’s about what happened n

In Class Activity Students discussed the readings of the week in synchronous chat n Students produced a summary of their conversation n

Comments from Activity n Very specific in scope ¨A bit more like query based summarization than generic summarization ¨ In both cases large portions of the text are irrelevant and should not be included ¨ But the query is less topic oriented – more like “what was finally agreed on” ¨ How would the relevance criteria be different?

Cues people mentioned n Phrases that indicated that an opinion was being expressed: ¨ "The authors were cautious about how. . . ¨ ". . . recommendation are simple yet profound. . . “ ¨ ". . . systems had an interesting correlary. “ ¨ "I believe. . . ", "I agree. . . ", "I stated that. . . ", "yes they. . . ". ¨ Inclusion of page numbers n Roles: opinion of one of the “group leaders” had more influence

Nature of the data n More stream of consciousness ¨ Bouncing ideas back and forth and refining them before inserting into the summary ¨ In that case, the summary did not include any mention of what happened to produce that consensus n n n In some cases almost every point in the summary was discussed Sometimes students just focused on coordination – you don’t actually see the evolution of ideas In other cases, one student wrote the summary, and the other students just made small modifications

Goals of the summary Would instructors want to know about the process? n Would instructors want the students to have a real exchange of ideas rather than just get the summary done? n Not just content compression – the process is really important – techniques would have to be really different n

Connection with Zechner paper… n Maybe divide data into topic segments first ¨ Summarize each topic separately? ¨ Remove irrelevant topics entirely? n Organize into different “flows” or “rounds” ¨ Like n thread recovery? Identify units, like question/answer pairs, so you can get complete propositions out ¨ Much more is implicit in dialogue than in expository text

Zechner paper…

Can you tell what the conversation was about? se e h ft o ich h W t? s be s m ee s

What would your summary look like?

Quote from journal version… We want to stress beforehand, though, that the originality of our system should not be seen in the particular implementation of its individual components, but rather in their selection and specific composition to address the issues at hand in an effective and also efficient way.

Discourse structure important sometimes the crucial information is found in a sequence of turns from several speakers One of the properties of multiparty dialogues is that shared information is created between dialogue participants. The most obvious interactions of this kind are question-answer (Q-A) pairs.

Not trivial to match questions and answers

Analysis of Knowledge Building * Used to predict how effectively students are learning together (Joshi & Rosé, 2007).

Examples

New types of cues Prosodic cues can help with relevance ranking n Prior work related to broadcast news shows that prosody helps with identifying key portions of reports n Possible to see what is being emphasized n Maybe possible to identify points of tension in conversations, and then when the tension dissipates… n

Questions?