Speech Summarization Julia Hirschberg (thanks to Sameer Maskey for some slides) CS4706 Summarization Distillation • ‘…the process of distilling the most important information from a source (or sources) to produce an abridged version for a particular user (or users) and task (or tasks) [Mani and Maybury, 1999] • Why summarize? Too much data! Types of Summarization • Indicative – Describes the document and its contents • Informative – ‘Replaces’ the document • Extractive – Concatenate pieces of existing document • Generative – Creates a new document • Document compression [Salton, et al., 1995] Sentence Extraction Similarity Measures [McKeown, et al., 2001] SOME SUMMARIZATION TECHNIQUES BASED ON TEXT (LEXICAL FEATURES) Extraction Training w/ manual Summaries [Hovy & Lin, 1999] Concept Level Extract concepts units [Witbrock & Mittal, 1999] Generate Words/Phrases [Maybury, 1995] Use of Structured Data Sentence Extraction/Similarity measures [Salton, et al. 1995] • Extract sentences by their similarity to a topic sentence and their dissimilarity to sentences already in summary (Maximal Marginal Relativity) • Similarity measures – Cosine Measure – Vocabulary Overlap – Topic word overlap – Content Signatures Overlap Concept/content level extraction [Hovy & Lin, 1999] • Present key-words as summary • Builds concept signatures by finding relevant words in 30,000 WSJ documents, each categorized into different topics • Phrase concatenation of relevant concepts/content • Sentence planning for generation Feature-based statistical models [Kupiec, et al., 1995] • • • • Create manual summaries Extract features Train statistical model using various ML techniques Use the trained model to score each sentence in the test data • Extract N highest-scoring sentences k P( s S | F1 , F2, ...Fk ) P( F j 1 j |s S ) P( s S ) k P( F ) j 1 j • Where S is summary given k features Fj and P(Fj) & P(Fj|s of S) can be computed by counting occurrences Structured Database [Maybury, 1995] • Summarize text represented in structured form: database, templates – E.g. generation of a medical history from a database of medical ‘events’ Relative frequency of E # of occurrence s of event E Total # of all events • Link analysis (semantic relations within the structure) • Domain dependent importance of events Comparing Speech and Text Summarization • Alike – Identifying important information – Some lexical, discourse features – Extraction or generation or compression • Different – Speech Signal – Prosodic features – NLP tools? – Segments – sentences? – Generation? – Errors – Data size Text vs. Speech Summarization (NEWS) Speech Signal Speech Channels - phone, remote satellite, station Transcript- Manual Transcripts - ASR, Close Captioned Lexical Features Some Lexical Features Many Speakers - speaking styles Segmentation -sentences Story presentation style Error-free Text NLP tools Structure -Anchor, Reporter Interaction Prosodic Features -pitch, energy, duration Commercials, Weather Report Speech Summarization Today • Mostly extractive: – Words, sentences, content units • Some compression methods • Generation-based summarization difficult – Text or synthesized speech? Generation or Extraction? • • • • • • • • • • • • • • SENT27 a trial that pits the cattle industry against tv talk show host oprah winfrey is under way in amarillo , texas. SENT28 jury selection began in the defamation lawsuit began this morning . SENT29 winfrey and a vegetarian activist are being sued over an exchange on her April 16, 1996 show . SENT30 texas cattle producers claim the activists suggested americans could get mad cow disease from eating beef . SENT31 and winfrey quipped , this has stopped me cold from eating another burger SENT32 the plaintiffs say that hurt beef prices and they sued under a law banning false and disparaging statements about agricultural products SENT33 what oprah has done is extremely smart and there's nothing wrong with it she has moved her show to amarillo texas , for a while SENT34 people are lined up , trying to get tickets to her show so i'm not sure this hurts oprah . SENT35 incidentally oprah tried to move it out of amarillo . she's failed and now she has brought her show to amarillo . SENT36 the key is , can the jurors be fair SENT37 when they're questioned by both sides, by the judge , they will be asked, can you be fair to both sides SENT38 if they say , there's your jury panel SENT39 oprah winfrey's lawyers had tried to move the case from amarillo , saying they couldn't get an impartial jury SENT40 however, storythe judge moved against them in that matter … summary [Christensen et al., 2004] Sentence extraction with similarity measures [Hori C. et al., 1999, 2002] , [Hori T. et al., 2003] SPEECH SUMMARIZATION TECHNIQUES Word scoring with dependency structure [Koumpis & Renals, 2004] Classification [He et al., 1999] User access information [Zechner, 2001] Removing disfluencies [Hori T. et al., 2003] Weighted finite state transducers Content/Context sentence level extraction for speech summary [Christensen et al., 2004] Find sentences similar to the lead topic sentences Use position features to find the relevant nearby sentences after detecting a topic sentence where Sim is a similarity measure between two sentences or a sentence and a document (D) and E is the set of sentences already in the summary ^ Sk s arg max {Sim( s1, si )} si D / E ^ Sk s arg max {Sim( D, si )} si D / E Choose a new sentence which is most like D and most different from E Weighted finite state transducers for speech summarization [Hori T. et al., 2003] • Summarization includes speech recognition, paraphrasing, sentence compaction integrated into single Weighted Finite State Transducer • Decoder can use all knowledge sources in one-pass strategy R H C LG • Speech recognition using WFST – Where H is state network of triphone HMMs, C is triphone connection rules, L is pronunciation and G is trigram language model • Paraphrasing can be looked at as a kind of machine translation with translation probability P(W|T) where W is source language and T is Z H C LGS D the target language • If S is the WFST representing translation rules and D is the language model of the target language speech summarization can be looked at as the following composition Speech Translator H C L Speech recognizer G S Translator D User Access Identifies What to Include [He et al., 1999] • Summarize lectures or shows by extracting parts that have been viewed the longest • Needs multiple users of the same show, meeting or lecture for training • E.g. To summarize lectures compute the time spent on each slide • Summarizer based on user access logs did as well as summarizers that used linguistic and acoustic features – Average score of 4.5 on a scale of 1 to 8 for the summarizer (subjective evaluation) • Word level extraction by scoring/classifying words [Hori C. et al., 1999, 2002] Score each word in the sentence and extract a set of words to form a sentence whose total score is the product/sum of the scores of each word Example: Word Significance score (topic words) Linguistic Score (bigram probability) Confidence Score (from ASR) Word Concatenation Score (dependency structure grammar) M S (V ) {L(vm | ...vm1 ) I I (vm ) cC (vm ) T Tr(vm1,vm ) m 1 Where M is the number of words to be extracted, and I C T are weighting factors for balancing among L, I, C, and T r Segmentation Using Discourse Cues [Maybury, 1998] Discourse Cue-Based Story Segmentation Discourse Cues in CNN Start and end of broadcast Anchor/Reporter handoff, Reporter/Anchor handoff Cataphoric Segment (“still ahead …”) Time Enhanced Finite State Machine representing discourse states such as anchor segment, reporter segment, advertisement Other features: named entities, part of speech, discourse shifts “>>” speaker change, “>>>” subject change Source Precision Recall ABC 90 94 CNN 95 75 Jim Lehrer Show 77 52 CU: Summarization without Words: Does importance of ‘what’ is said correlates with ‘how’ it is said? • Hypothesis: “Speakers change their amplitude, pitch, speaking rate to signify importance of words, phrases, sentences.” – If so, then the prediction labels for sentences predicted using acoustic features (A) should correlate with labels predicted using lexical features (L) – In fact, this seems to be true (corr .74 between precitions of A and L Is It Possible to Build ‘good’ Automatic Speech Summarization Without Any Transcripts? Feature Set F-Measure ROUGE-avg L+S+A+D 0.54 0.80 L 0.49 0.70 S+A 0.49 0.68 A 0.47 0.63 Baseline 0.43 0.50 • Just using A+S without any lexical features we get 6% higher Fmeasure and 18% higher ROUGE-avg than the baseline Evaluation using ROUGE • F-measure too strict – Predicted summary sentences must match summary sentences exactly – What if content is similar but not identical? • ROUGE(s)… ROUGE metric • • • • • Recall-Oriented Understudy for Gisting Evaluation (ROUGE) ROUGE-N (where N=1,2,3,4 grams) ROUGE-L (longest common subsequence) ROUGE-S (skip bigram) ROUGE-SU (skip bigram counting unigrams as well) • Does ROUGE solve the problem? Next Class • Emotional speech • HW 4 assigned