Information Extraction and Automatic Summarisation

1
Information Extraction and Automatic Summarisation

2
How IE fits in with IR

• IR selects a few relevant documents from many
• IE starts with one or a few relevant documents
• IE pulls out the words and phrases most central
  to the meaning of that/those documents to produce
  an extract.

3
Two process associated with information extraction

• determination of facts to go into structured
  fields in a database.
• extraction of text that can be used to summarise
  an item.
• In the first case only a subset of the important
  facts in an item may be identified and extracted.
  The term slot is used to define a particular
  category of information to be extracted. Slots
  are organised into semantic frames.

4
What do we most want to know from a journal
article about agriculture?

• AGENT chemical agent applied
• CV cultivar (e.g. King Edward)
• HLP high level property (e.g. yield)
• INF influence (e.g. drought)
• LAB site of test (e.g. laboratory)
• LLP low level property (e.g. root mass)
• LOC location
• PEST pest or disease
• SOIL soil
• SPEC crop species (e.g. potato)

5
Automatic Abstracting

• In the second case, rather than trying to
  determine specific facts, the goal of document
  summarisation is to extract a summary of an item
  maintaining the most important ideas while
  significantly reducing its size. For journal
  articles, this is called automatic abstracting.
  The abstract is a way for the user to determine
  the utility of an article without having to read
  the whole item.

6
Kupieks heuristics

• Sentence length feature that requires the
  sentence to be over five words in length.
• Fixed phrase feature that looks for the existence
  of phrase cues, e.g. in conclusion.
• Paragraph feature that places emphasis on the
  first ten and the last five paragraphs in an item
  and also the location of the sentences within the
  paragraph.
• Thematic word feature that uses word frequency.
• Uppercase word feature that places emphasis on
  proper names and acronyms.
• discovered that location based heuristics give
  better results than the frequency based features.

7
Paices rules

• Frequency Keyword Approach First find a set of
  index terms for the document (manually,
  mid-frequency, tf idf, words occurring in the
  title, etc.). Then choose the sentences which
  contain most keywords.
• Location The first sentence in a paragraph is
  most central to the theme of a text. The last
  sentence is the next most central.
• Cue method Not actually keywords, but their
  presence in a document show that the sentence is
  (or is not) important. These may be bonus words,
  e.g. greatest, significant, or stigma words, e.g.
  hardly, impossible.
• Indicator phrases, e.g. The main aim of our
  paper is to describe , Our investigation has
  shown that .

8
Hoey method cohesion in text.

• The most important sentences in a document are
  those which are related to the largest number of
  other sentences. Find how many concepts in each
  sentence are related to concepts in other
  sentences. Concepts may be related by
• Exact match, e.g. computer and computer
• Grammatical variants e.g. computer, computing
• Synonyms e.g. sedate, tranquilise, drug
• Antonymy e.g. cold, hot
• General-specific e.g. scientists, biologists

9
Hoey (2)

• Form a repetition net, with entries in the form s
  ( a , b) such as 26 ( 6, 4) meaning sentence no.
  26 is bonded to 6 earlier sentences and 4 later
  sentences.
• If a b is high, the sentence is central to the
  topic
• If only b is high, the sentence is a topic
  opener
• If only a is high, the sentence is topic closing.

10
Hoey (3)

• Cohesion in text is concerned with explicit
  references within a sentence which can only be
  understood by reference to material elsewhere in
  the text.
• Anaphora come after their explicit mention in the
  text, e.g. Marie Curie was born in Warsaw. She
  devoted her life to the study of radioactivity.
• Cataphora come before their explicit mention in
  the text, e.g. He was to become the best known
  physicist of his generation. His name was Albert
  Einstein.

11
Generating Canned Text

• This paper studies the effect of AGENT on the HLP
  of SPEC
• OR
• This paper studies the effect of INF on the HLP
  of SPEC
• when it is infested by PEST.
• An experiment was undertaken
• using cultivars CV
• in, at LOC
• where the soil was SOIL.
• The HLP is, are measured by analysing the LLP.

12
Extracts vs. Abstracts (Mani, p6)

• An extract is a summary consisting entirely of
  material copied from the input
• An abstract is a summary at least some of whose
  material is not present in the input.