BEGIN:VCALENDAR VERSION:2.0 PRODID:ILLC Website X-WR-TIMEZONE:Europe/Amsterdam BEGIN:VTIMEZONE TZID:Europe/Amsterdam X-LIC-LOCATION:Europe/Amsterdam BEGIN:DAYLIGHT TZOFFSETFROM:+0100 TZOFFSETTO:+0200 TZNAME:CEST DTSTART:19700329T020000 RRULE:FREQ=YEARLY;BYMONTH=3;BYDAY=-1SU END:DAYLIGHT BEGIN:STANDARD TZOFFSETFROM:+0200 TZOFFSETTO:+0100 TZNAME:CET DTSTART:19701025T030000 RRULE:FREQ=YEARLY;BYMONTH=10;BYDAY=-1SU END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:/NewsandEvents/Archives/2022/newsitem/13273/11 -January-2022-Computational-Linguistics-Seminar-Ar thur-Bražinskas DTSTAMP:20211221T161302 SUMMARY:Computational Linguistics Seminar, Arthur Bražinskas ATTENDEE;ROLE=Speaker:Arthur Bražinskas (Universit y of Edinburgh) DTSTART;TZID=Europe/Amsterdam:20220111T160000 LOCATION:Zoom DESCRIPTION:Opinion summarization is the automatic creation of text reflecting subjective informatio n expressed in multiple documents, such as user re views of a product. These short summaries can help users make better purchasing decisions by condens ing useful information in hundreds or even thousan ds of reviews. However, due to the high cost of su mmary production, datasets large enough for superv ised learning were absent until recently. This lea d to a variety of extractive methods that construc t summaries from review sentences. However, these methods often produce incoherent summaries with un important details. This presentation will focus on abstractive approaches that generate summaries us ing a free vocabulary and thus can yield more cohe rent texts. We will discuss summarizers trained in unsupervised, few-shot, and supervised regimes. T hese models combine principles of latent probabili stic models, variational inference, and reinforcem ent learning. In our unsupervised model (Copycat), we treat the product and review representations a s latent continuous variables. At test time, we in duce summarizing representations and map them to s ummarizing texts. In the supervised model (SelSum) , we decompose the system into a selector (posteri or) and summarizer. The selector treats reviews as latent categorical variables and selects a summar y-relevant subset in training. Only the small subs et is passed to the summarizer, which results in c omputational and memory savings. The system is tra ined end-to-end using variational inference and re inforcement learning. Finally, we fit another sele ctor (prior) that selects subsets of informative r eviews to summarize in test time. X-ALT-DESC;FMTTYPE=text/html:\n

Opinion summar ization is the automatic creation of text reflecti ng subjective information expressed in multiple do cuments, such as user reviews of a product. These short summaries can help users make better purchas ing decisions by condensing useful information in hundreds or even thousands of reviews. However, du e to the high cost of summary production, datasets large enough for supervised learning were absent until recently. This lead to a variety of extracti ve methods that construct summaries from review se ntences. However, these methods often produce inco herent summaries with unimportant details. This pr esentation will focus on abstractive approaches th at generate summaries using a free vocabulary and thus can yield more coherent texts. We will discus s summarizers trained in unsupervised, few-shot, a nd supervised regimes. These models combine princi ples of latent probabilistic models, variational i nference, and reinforcement learning. In our unsup ervised model (Copycat), we treat the product and review representations as latent continuous variab les. At test time, we induce summarizing represent ations and map them to summarizing texts. In the s upervised model (SelSum), we decompose the system into a selector (posterior) and summarizer. The se lector treats reviews as latent categorical variab les and selects a summary-relevant subset in train ing. Only the small subset is passed to the summar izer, which results in computational and memory sa vings. The system is trained end-to-end using vari ational inference and reinforcement learning. Fina lly, we fit another selector (prior) that selects subsets of informative reviews to summarize in tes t time.

\n URL:https://projects.illc.uva.nl/LaCo/CLS/ END:VEVENT END:VCALENDAR