April 3, 2008
Discuss patterns for appointment events: variables, concept hierarchy, write statement (Assignment 10)
Alternatives to hand-coded patterns ... supervised learning
Large scale corpus annotation is burdensome, because it must be repeated for each type of event we would like to extract. Can we learn from raw (unannotated) text?
Several approaches have been developed for such weakly supervised
learning. These often are 'bootstrapping' methods, which start with a few
'seed' patterns and gradually discover additional patterns. These procedures rely on the availability of
large amounts of text … no longer an obstacle because of the Web.
One such procedure, described by Sergei Brin, discovers patterns for binary relations. It works as follows
There is a risk that incorrect patterns can lead to incorrect pairs, so the errors can grow rapidly. Brin reduced this effect by using a functional relation (book / author) and rejecting patterns which did not produce functional relations. Agichtein conducted similar experiments for the company – headquarters location relation.
In some cases we may not know what sorts of relations are worth extracting. In such a case we can run an unsupervised relation learner. Hasegawa et al. describe such a learner, which clusters the word sequences appearing between two organization names (for example) based on the commonality of names and similarity of the word sequences.
Sergei Brin. Extracting Patterns and Relations from the World Wide Web. (Also available in PDF) In Proc. World Wide Web and Databases International Workshop, pages 172-183. Number 1590 in LNCS, Springer, March 1998.
Eugene Agichtein and Luis Gravano. Snowball: extracting relations from large plain-text collections. ACM Digital Libraries 2000.
Takaaki Hasegawa, Satoshi Sekine, and Ralph Grishman. Discovering relations among named entities from large corpora. Proc. Annual Meeting Assn. Computational Linguistics (ACL 04) 2004; Barcelona, Spain.
Some applications of NLP, such as grammar checking, depend only on the syntactic form of an input. Most applications, however, are dependent on what a sentence 'means'. For example, for information extraction, we want to find mentions of a type of an event, however expressed. For question answering, we need to connect the question to a body of knowledge which can provide the answer. In fact, as we shall see when we consider the processing of extended (multi-sentence) texts, we generally must integrate the sentence with a great deal of 'background knowledge' in order to understand what it means. How should we express the meaning of an utterance?
Clearly we can express the meaning in natural language. Why is this not satisfactory for analyzing the meaning? The components (the words) are ambiguous; even sentences in isolation are ambiguous. Resolving these ambiguities can be difficult. Even though syntactic analysis has made some relationships explicit, others are still implicit; for example, it does not indicate the quantificational structure of a sentence. Furthermore, the rules for inferring new facts from given facts in natural language may be very complicated.
In order to analyze and manipulate the meaning of sentences, we will transform the sentences into a meaning representation language. We want to transform the sentences into a language which
[J&M also mention the characteristic of having a canonical form for each meaning. This is an ultimate goal rather than an easily achieved criterion. However, we can see both syntactic analysis and semantic analysis as moving in this direction -- reducing paraphrase, i.e., reducing the variation in form for a given meaning.]
These are the properties of the languages of logic. Actual systems may use different representations, but they are generally equivalent to the formal language (extensions of predicate calculus) we will use for presentation.
The simplest form is propositional logic, but it is not powerful enough for our purposes. Predicate logic combines predicates and their arguments. Basic elements:
Predicate logic has simple rules of inference, such as modus ponens (from A and A==>B, infer B).
Predicate calculus is intended for representing “eternal truths” (like the facts of mathematics). We face several problems when we try to use it for representing events. First, how many arguments does an event have (consider J&M example of eating, p. 524)? In natural language, the same type of event may be described with many different sets of arguments and modifiers (time, location, speed, ...). We can use meaning postulates to relate these, but that requires many such postulates and may make commitments we do not intend. Second, we need to individuate events (say that two events are the same or different; count events).
We can address this problem by reification -- treating events as objects (J&M p. 527).
There are many other issues which we may need to address in our meaning representation language:
Information extraction applications are generally concerned with identifying specific, individual events and relations between entities. They need to capture event modifiers such as time and location, but not quantification. Such applications typically use a frame or slot-filler representation. For each type of event (or set of event types taking the same arguments) we define a frame (template), with one slot containing a unique identifier of the event, and one slot for each possible argument/modifier. Similarly, a frame is defined for each type of entity. Slots may be filled with constants or the identifiers of other events or entities.
Mapping Syntax to Semantics (J&M Chapter 15))
We want to compute the semantic representation of a sentence from the parse tree. Because the parse tree provides a structural framework, we will use a compositional, syntax-driven translation process. This means that we will associate a (partial) semantic interpretation with some or all of the nodes of the parse tree, to be computed (using a rule) from the interpretations of its children.
We could embed this translation in a procedure associated with each type of node. Alternatively, one can formalize this by a set of rules associated with the productions of the grammar (J&M sec. 15.1). The grammar will be extended to add a SEM feature, representing the semantic interpretation of a node. Each production will then incorporate the rule for computing its SEM value, and the SEM of the root will be the interpretation of the sentence. (J&M p. 549).
The semantics of a verb phrase is essentially the semantics of a clause, with one argument (the subject) missing … a predicate with one unbound argument. We can represent this by a lambda expression (p. 551). Lambda expressions are commonly used to capture the rules for composing the semantics.
For the process of translating syntactic to semantic forms, it is convenient to introduce restricted quantifiers, of the form
(exists x: C(x))
These do not add any power to predicate calculus; they can be rewritten
x: C(x)) P(x) = (forall x) (C(x) => P(x))
(exists x: C(x)) P(x) = (exists x) (C(x) & P(x))
Roughly speaking, a noun phrase can be translated to a constant or a restricted quantifier.
One source of ambiguity is quantifier scope:
A woman gives birth in the United States every five minutes.
We can represent the two readings in conventional predicate calculus using different quantifier scopes. If we explicitly represent all the semantic ambiguities in a sentence in this way, we may have very many readings. It is therefore practical to initially produce (from the parse) a representation which captures multiple readings … which encodes (some of) the ambiguity. (And hope that this ambiguity can be resolved at a later stage of semantic analysis.)
In particular, we can use complex terms (J&M p. 555)
<Quantifier x: C(x)>
with the understanding that
P(<Quantifier x: C(x)>) = (Quantifier x: C(x)) P(x)
an expression contains several complex terms, the scope of the quantifiers is
indeterminate. Semantic analysis will generate such quasi-logical forms,
with a separate step then determining the quantifier scope and generating a
predicate calculus expression.