Please enable JavaScript.
Coggle requires JavaScript to display documents.
Watson IBM knowledge base (PRISMATIC) (2: Terminology (Frame - the basic…
Watson IBM knowledge
base (PRISMATIC)
1
: Method to acquire shallow syntactic knowledge and ist implied semantics
2 stages
Inferring of additional semantics from aggregate statistics of the automatically extracted shallow knowledge
Automatic extraction of shallow knowledge from large collection of documents
Specifically - representing syntactic relations within sentential context as frames and slots
Syntactic relations don't capture all of the knowledge contained in text, but redundancy of a large corpus helps to increase the coverage of the knowledge base
Existing knowledge bases (VerbNet, FrameNet etc.) are manually built => tend to specify type constraints at a very high level
they are not suitable for the purposes of Watson
Goal: infer more fine-grained knowledge for predicates automatically from a large amount of data + no limitation to a specific type system (verbs, binary semantic relations, specific type hierarchy etc.)
2
: Terminology
Slot
- a binary relation. Most of them are dependency relations extracted from the parse tree (e.g. subject, object)
Slot value
- either the lemma form of a term from the sentence or a type annotated by NER
PRISMATIC can be formally defined as a bag of frames made of set of slot - slot value pairs
Frame projection
- a portion of a frame (or the operation that produces the portion) that occurs with regularity in many frames (e.g. S-V-O)
Frame
- the basic semantic unit representing a set of entities and their relations in a piece of text
Frame is made of a set of slot and value pairs
Can be
intentional
(its values are types) and
extensional
(its values are instances)
Intentional: {verb,"annex", object-type,"Region"}
Extensional: {subject,"United States", verb,"annex",
object,"Texas"}
3:
System
NLP tools: a dependency parser, a rule-based NER, and a co-reference resolution component
3 phases
Corpus processing (by NLP tools)
A Slot Grammar - based
parser
(ESG) is used to fill in the slots
A simple rule-based
co-reference resolution
component helps to enhance the coverage of the frame projection (valuable in case of sparse data)
Relation detectors
to fill frame slots not captured by ESG (e.g.
isa
slot)
Semantic annotation by a
rule-based NER
(identifies the type of slot-fillers) => this info is then used in the frame extraction stage to construct intentional frames
Frame extraction
Is performed on the basis of the dependency parses and associated annotations
Elements are limited to those that represent
the participant information
of a predicate
2 levels deep at most
Reasons: to
reduce the chance of parse error
in each frame, to focus on immediate participants of a predicate
=> a single tree may result in multiple frames
Additional info:
semantic relations from a syntactic pattern-based relation detector (e.g.
isa
)
nonparser info (e.g. type annotation from NER)
a flag to indicate whether a word is a proper noun (allows to separate intentional and extensional parts of PRISMATIC)
Frame projection
Frame projections specify a projection operation on a frame
e,g, an N-P-OT frame only keeps noun, preposition and object-type slots and discards all the rest
other examples: S-V-O, S-V-O-IO (indirect obj), S-V-P-O, N-Isa, N-Isa-Mod (modifier) etc.
they dissect frames along different dimensions
examples of usefulness
learning a predominant object type for a verb (e.g. annex - region)
such intentional frames can be used to learn entailment rules
to find the lexical type of nouns (e.g. Bill Clinton - politician)
to make frame projections more flexible - allow them to specify optional value constraints for slots (e.g. S-V-O where both S and O are proper nouns)
creating strictly extensional frames which can be used to provide factual evidence in question answering
induce interesting knowledge patterns from the redundancy in the data
4:
Aggregate statistics (2 examples)
Normalized PMI
- takes into consideration the popularities of different subsets of a frame and scores them accordingly. Ranges from -1 (no co-occurrence) to 1 (complete co-occurrence)
E.g. {subj, "Einstein", verb, "win"} and {obj, "award"}
Conditional probability
- the probability of a specific set of fillers given a subset of such slot-filler pairs. Is less affected by the corpus size
E.g. {subj, "Einstein", verb, "win", obj, "award"} | {subj, "Einstein", verb, "win"}
Frequency
- the number of frames whose slot values match with all the slot-value pairs in a given frame
Gives an estimate of popularity of fillers for a given set of slots
E.g. {subj, "Einstein", verb, "win", obj, "award"} = 142 signifies a set of popular fillers in the corpus
Is affected by the size of the corpus and the popularity of individual fillers
E.g. if Einstein appears often, Einstein winning Award will also appear more often
5:
Applications in Watson
Candidate generation - uses Isa slot and modifier slot to generate up to 20 candidates
type coercion - to determine whether a candidate answer is of the right lexical answer type - LAT (Isa slot)
Missing link - to restore the semantic relations btw entities. Allows to find missing links and estimate the degree of closeness btw 2 concepts
Type inference and its related uses
for question analysis - if LAT in question is meaningless (it, this) allows to infer a meaningful LAT by looking at the question. Problem: lack of context-aware knowledge in PRISMATIC
for answer scoring: checking whether a given candidate fits into the frame of question
for co-reference resolution
for word-sense disambiguation
6:
Limitations
does not take context into consideration when producing the aggregate statistics
gaps in coverage
confidence can be improved to help create better aggregate statistics