G22.2590 - Natural Language Processing - Spring 2003 Prof. Grishman
Lecture 13 Outline
April 24, 2003
Asgn 9 and pattern learning: people were able to get most
examples to work (except those with a name recognition error), but the
process of developing all these patterns is time consuming. Many
extraction systems now learn patterns (or probabilistic extraction
models) from annotated text, or in some cases even from unannotated text..
Such learning methods will be a prime focus of the Advanced NLP class next
Name recognition: An improved Jet HMM model for names is
available as name_hmm_2. This may be helpful for people using a name
recognizer as part of their term project. It improves tagging accuracy
for news text from about 75% to about 82%, primarily by using a larger
corpus (300 articles instead of 100), and also by small changes to the
HMM model. Further improvement (to the low 90's%) would require a
more elaborate model than currently provided by Jet (one which uses more
features and more context in computing transition and emission probabilities).
Discourse. Until now we considered the structure and meaning
of sentences in isolation. We now turn to issues primarily connected
with multi-sentence text -- discourse.
Reference Resolution (J&M 18.1)
referent: real-world object being referred to
referring expression: a portion of text referring to that
object (we shall also refer to these as mentions of the object)
discourse entities: the set of objects referred to by a text
coreference: two expressions -- antecedent and anaphor
-- referring to the same thing
the first mention of an object in a discourse evokes this object
Types of referring expressions
definite pronouns (he, she, it, ...): generally anaphoric
- but 'it' has non-referring usages: "It is raining." "It
is unlikely that he will come."
indefinite pronouns (one): can be modified ('the green one')
definite NPs (the car): generally anaphoric
indefinite NPs (a car): generally evoke a new discourse entity
- may also be generic: "Giraffes are beautiful creatures."
names: named entities can be later referred to by portions of their
inferrables: sometimes the relation between anaphor and antecedent
is not one of identity ...
"I entered the room and looked at the ceiling."
zero anaphora: sometimes the anaphor is implicit
- many languages allow subject omission, and some allow omission of
other arguments (e.g., Japanese)
- some cases of inferrable anaphora can be described in terms of PPs
with zero anaphors:
"IBM announced the appointment of Fred as president [of
expressions can also refer to events, propositions, ...
- "Fred claimed that no one programs in Lisp. That
Resolving pronoun reference
constraints: number and gender agreement
preferences: recency, grammatical role (reference to subject preferred)
implementation: associate score with preferences; select antecedent
of highest score satisfying constraints
(can incorporate preferences into search order -- Hobbs' search order)
accuracy fairly good -- somewhere in 80's%
Resolving other referring expressions
names: generally quite straightforward -- look for prior name of
which this is a substring
common noun phrases: generally quite hard
- deciding if an NP is anaphoric
- deciding if an NP description is consistent with an antecedent
(for example, we may use different nouns to describe the
same entity -- "the soldier", "the Marine", etc.)
Anaphora resolution in Jet
assumes the only referring expressions are noun groups
generates entity annotations corresponding to discourse entities
anaphora resolution consists of linking each noun group to a new or existing
simple rule for common noun phrases -- only checks for matching head
performed by resolve operation (typically done after all pattern
results are displayed in a separate entity window (which appears,
along with the regular document viewer, if entities have been generated)
Using anaphora resolution for extraction: an example
In many cases, we want to be able to retrieve an argument from context
when it is not part of the immediate syntactic structure. A simple
way of doing this is to generate a zero anaphor (an ngroup constituent
not spanning any text) and then let reference resolution map it to an entity.
We have created a version of the AppointPatterns which uses this method
to collect organization names and, in some cases, people names.
Discourse Analysis: Analyzing Text Coherence (J&M 18.2)
Why are we interested in analyzing the structure of a discourse beyond
the sentence level?
How to analyze text coherence?
resolve ambiguities from earlier stages of processing (syntactic, semantic,
establish intersential connections:
contextual (inferrable) reference
Most of these connections are implicit, but they are an important part
of the information which the hearer/reader is expected to glean from the
discourse. A discourse lacking such connections is "incoherent".
I walked from the kitchen into the living room and looked at the ceiling.
Just before dawn, the Valian sighted the Zwiebel and fired two torpedos.
It sank swiftly, leaving few survivors.
Jack poisoned Sam. He died within a week. vs.
Jack poisoned Sam. He was arrested within a week.
First we need to define the criteria of text coherence
different types of text have different goals and organizing principles,
so across all text types it is possible to define only very general criteria
more specific criteria can be defined for particular genres: exposition,
most work has been done on analysis of narrative
In analyzing the connection between two sentences S1 and S2, we use 'world
knowledge' to generate expectations arising out of S1, and try to match
these against S2
if there are multiple interpretations of S2, we may succeed in matching
only some of them against the expectations, thus disambiguating S2 (e.g.,
the anaphora examples just above)
More generally, we can apply abductive methods: given the prior discourse
and world knowledge, what minimal set of additional assumptions need one
make to infer the current sentence? (J&M 697-704)