08_Summarizing

Reduce information overload by extracting relevant information

• Main Type Dimensions:
  • Single document vs. Multi-document summarization
  • Generic summarization vs. Query-focused summarization:
    • Generic summarization does not take into account the particular user or her information need, as opposed to query-focused summarization
  • Abstractive vs. Extractive summarization
    • Abstractive summarization results in a text that differs from the original
    • Extractive summarization consists of the original phrases and sentences
      • Most state-of-the-art systems focus on this since it is easier

• Other types:
  • Contrastive multiple-document summaries:
    • the common topics of all the documents as well as unique topics pertaining to each document
  • Update summaries:
    • only new information that has not been covered before

1. Content selection:
   • choose sentences to extract from the document, either with an unsupervised or supervised method
   • Baseline: use first k sentences
2. Information ordering:
   • choose an order for the sentences
   • Baseline: keep the order of the original text
3. Sentence realization:
   • clean up sentences, e.g. sentence simplification, sentence fusion, etc.
   • Baseline: do not perform any combination or clean-up

• Simplest approach:
  • select sentences that have more informative words
    • e.g. measured with the
      • maximum likelihood estimate
      • TF-IDF
  • Sentences are scored based on the score of the informative words they contain
  • All sentences are ranked by their score and the top-ranked sentences are kept for the summary

The SumBasic system

• Compute probability for each word
• n is the number of times the word appeared in the input and N is the total number of words in the input
• p(w) = n/N
• for each sentence average probability of the words in it
• then pick best scoring sentence
• update probability for words in chosen sentence p(w_new) = p(w_old)^2

Log-Likelihood ratio (LLR)

• quite common
• Document(s) to summarize is compared to Background corpus

LexRank

• graph-based content selection
• represents a cluster of documents as a network of related sentences
• sentences that are similar to a lot of others are more central
• uses TF-IDF applied to sentences

• group sets of words, esp. nouns, which are semantically related (same word - same sense, synonyms, hypernyms/hyponyms, co-hyponyms, collocations)
• Lexical chains can be used to identify important concepts from a document
• Each noun instance usually belongs to exactly one lexical chain
  • → it is necessary to perform word sense disambiguation
• Example: Picture

• For Summarization:
  • All terms representing the same concept occur in the same chain
    • → avoids repetition
  • The chain combines the weight (frequency) of its members, so that low frequency terms may still help identifying important concepts
  • Build lexical chains:
    • Extract nouns and noun phrases
    • Use a lexical resource or statistics over large corpora to determine word
  • relatedness
  • Identify strong chains based on their length
  • Extract significant sentences
    • one sentence for each chain (or more for larger chains)

• Define features to assess sentence saliency
• Training data: corpus where sentences are annotated as part of the extract summary (1) or not (0)
• Example Features:
  • Fixed-phrase feature:
    • “in conclusion” indicates summary
  • Position feature:
    • 1^st/ last paragraph, initial/ final sentence more likely to be important
  • Thematic word feature:
    • Repetition as indicator of importance
  • Important words:
    • Sentence with high salience (e.g. several words with TF-IDF weight)
  • Uppercase word feature:
    • Often indicates named entities

• Characteristics of manually simplified texts
  • Fewer and shorter phrases
  • Fewer adjectives, adverbs and coordinating conjunctions
• Syntactic simplification:
  • Removal of specific syntactic structures, using hand-built transformational rules applied to dependency trees

Greater amount of redundancy

→ Challenge: avoid redundant and nearly identical sentences

• Redundancy factor based on the similarity between a candidate sentence and the sentences already selected for the summary:
  • A sentence is penalized if it is too similar to the summary
  • Alternative Method: apply clustering algorithm to produce clusters of related sentences
    • -> select a single sentence from each cluster
    • Example: Topic Cluster
      • A single topic cluster contains the set of all entities (e.g. NPs) pertaining to a topic in the document collection

• Generic summaries
  • Identify the clusters that span all documents
  • Select one candidate noun phrase from each cluster
  • Select sentences containing these NPs
• Contrastive summaries
  • Topics covered only in a single document are contrastive topics
• (Query-)Focused summaries
  • Focused summaries only contain information relevant to the user current context.
  • E.g. returned snippets by web search engines

• Extrinsic evaluation
  • Measure how different summarization algorithms influence the performance of human subjects for specific tasks:
    • Gather facts (on a given topic)
    • Decide if a document is relevant or not to a query (based on the summary)
• Intrinsic evaluation
  • Score an automatically generated summary by comparing it to human generated reference summaries
    • Metrics usually measure the amount of n-gram overlap
    • E.g. The Pyramid Method
      • Measures the amount of units of meaning shared between the candidate and reference summaries
      • Human annotators label the units of meaning in each reference and candidate summary

• Document summarization like keyphrase extraction, document summarization hopes to identify the essence of a text.
• Why unsupervised summarization works: These methods work based on the idea that sentences “recommend” other similar sentences to the reader.
• Summarization of multimedia documents, e.g. pictures or movies, is also possible.
• Also, keyphrases can improve information retrieval—if documents have keyphrases assigned, a user could search by keyphrase to produce more reliable hits than a full-text search.
• Summarization systems are able to create both query relevant text summaries and generic machine-generated summaries depending on what the user needs.