Title: Data Collection and Normalization for the Scenario-Based Lexical Knowledge Resource of a Text-to-Scene Conversion System
1Data Collection and Normalization for the
Scenario-Based Lexical Knowledge Resource of a
Text-to-Scene Conversion System
2Who I Am
- Rising senior at Reed College in Portland
- Linguistics major, concentration in Russian
3Overview
- WordsEye Scenario-Based Lexical Knowledge
Resource (SBLR) - Use of Amazons Mechanical Turk (AMT) for data
collection - Manual normalization of the AMT data and
definition of semantic relations - Automatic normalization techniques of AMT data
with respect to building the SBLR - Future automatic normalization techniques
4WordsEye Text-to-Scene Conversion
- the humongous white shiny bear is on the
american mountain range. the mountain range is
100 feet tall. the ground is water. the sky is
partly cloudy. the airplane is 90 feet in front
of the nose of the bear. the airplane is facing
right.
5Scenario-Based Lexical Knowledge Resource (SBLR)
- Information on semantic categories of words
- Semantic relations between predicates (verbs,
nouns, adjectives, prepositions) and their
arguments - Contextual, common-sense knowledge about the
visual scenes various actions and items occur in
6How to build the SBLR efficiently?
- Manual construction of the SBLR is time-consuming
and expensive - Past methods have included mining information
from external semantic resources (e.g. WordNet,
FrameNet, PropBank) information extraction
techniques from other corpora
7Amazons Mechanical Turk (AMT)
- Online marketplace for work
- Anyone can work on AMT, however
- It is possible to screen workers by various
criteria. We screened ours by - Located in the USA
- 99 approval rating
8AMT Tasks
- In each task, we asked for up to 10 responses. A
comment box was provided for gt10 responses. - Task 1 Given the object X, name 10 locations
- where you would find X. (Locations)
- Task 2 Given the object X, name 10 objects
found near X. (Nearby Objects) - Task 3 Given the object X, list 10 parts of X.
(Part- Whole)
9AMT Task Results
Target Words User Inputs Reward
Locations 342 6,850 0.05
Objects 245 3,500 0.07
Parts 342 6,850 0.05
- 17,200 total responses
- Spent 106.90 for all three tasks
- It took approximately 5 days to complete each task
10- Goal How to automatically normalize data
collected from AMT in such a way that AMT would
be useful for building the Scenario-Based Lexical
Knowledge Resource (SBLR)?
11Manual Normalization of AMT Data
- Removal of uninformative target item-response
item pairs between which no relevant semantic
relationship was held - Definition of the semantic relations held between
the remaining target item-response item pairs - This manually normalized set of data was used as
the standard against which we measured various
automatic normalization techniques.
12Rejected Target-Response Pairs
- Misinterpretation of ambiguous target item (e.g.
mobile) - Viable interpretation of target item was not
contained within the SBLR (e.g. crawfish as food
rather than a living animal) - Too generic responses (e.g. store in response to
turntable)
13Examples of Approved AMT Responses
- Locations
- mural - gallery
- lizard - desert
- Nearby Objects
- ambulance - stretcher
- cauldron - fire
- Part-Whole
- scissors - blade
- monument - granite
14Semantic Relations
- Defined a total of 34 relations
- Focused on defining concrete, graphically
depictable relationships - Generic relations accounted for most of the
labeled pairs (e.g. containing.r, next-to.r) - Finer distinctions were made within these generic
semantic relations (e.g. habitat.r, residence.r
within the overarching containing.r relation)
15Example Semantic Relations
- Locations
- mural - gallery - containing.r
- lizard - desert - habitat.r
- Nearby Objects
- ambulance - stretcher - next-to.r
- cauldron - fire - above.r
- Part-Whole
- scissors - blade - object-part.r
- monument - granite - stuff-object.r
16Semantic Relations within Locations Task
Relation Number of occurrences Percentage of total scored pairs
containing.r 1194 38.01
habitat.r 346 11.02
on-surface.r 333 10.6
geographical-location.r 306 9.74
group.r 183 5.83
- We collected 6850 locations for 342 target
objects from our 3D library.
17Semantic Relations within Nearby Objects Task
Relation Number of occurrences Percentage of total scored pairs
next-to.r 4988 75.66
on-surface.r 375 5.69
containing.r 293 4.44
habitat.r 243 3.69
object-part.r 153 2.32
- We collected 6850 nearby objects for 342 target
objects from our 3D library.
18Semantic Relations within Part-Whole Task
Relation Number of occurrences Percentage of total scored pairs
object-part.r 2675 79.12
stuff-object.r 552 16.33
containing.r 50 1.48
habitat.r 36 1.06
stuff-mass.r 17 0.5
- We collected 3500 parts of 245 objects.
19Automatic Normalization Techniques
- Collected AMT data was classified into
higher-scoring versus lower-scoring sets by - Log-likelihood and log-odds of sentential
co-occurrences in the Gigaword English corpus - WordNet path similarity
- Resnik similarity
- WordNet average pair-wise similarity
- WordNet matrix similarity
- Accuracy evaluated by comparison against manually
normalized data
20Precision Recall
- AMT data is quite cheap to collect, so we were
concerned predominantly with precision (obtaining
highly accurate data) rather than recall
(avoiding loss of some data). - In order to achieve more accurate data (high
precision), we will lose a portion of our AMT
data (low recall)
21Locations Task
Base- line Log- likel. Log- odds WN Path Sim. Resnik WN Avg. PW WN Matrix Sim.
Precision 0.5527 0.7502 0.7715 0.5462 0.5562 0.6014 0.4782
Recall 1.0 0.7945 0.6486 0.9649 0.9678 0.3454 1.0
- Achieved best precision with log-odds.
- Within high-scoring set, responses that were too
general (e.g. turntable - store) were rejected. - Within low-scoring set, extremely specific
locations that were unlikely to occur within a
corpus or WordNets synsets were approved (e.g.
caliper - architects briefcase)
22Nearby Objects Task
Base- line Log- likel. Log- odds WN Path Sim. Resnik WN Avg. PW WN Matrix Sim.
Precision 0.8934 0.8947 0.9048 0.9076 0.9085 0.9764 0.8795
Recall 1.0 1.0 0.8917 1.0 1.0 0.2659 1.0
- Relatively few target-response pairs were
discarded, resulting in high recall. - High precision due to open-ended nature of task
responses often fell under a relation, if not
next-to.r.
23Part-Whole Task
Base- line Log- likel. Log- odds WN Path Sim. Resnik WN Avg. PW WN Matrix Sim.
Precision 0.7887 0.7832 0.8231 0.7963 0.7974 0.8823 0.8935
Recall 1.0 0.4129 0.4622 1.0 1.0 0.2621 0.2367
- Rejected target-response pairs from the
high-scoring set were often due to responses that
named attributes, rather than parts, of the
target item (e.g. croissant - flaky) - Approved pairs from the low-scoring set were
mainly due to obvious, common sense responses
that would usually be inferred, not explicitly
stated (e.g. bunny - brain)
24Future Automatic Normalization Techniques
- Computing word association measures on much
larger corpora (e.g. Googles 1 trillion word
corpus) - WordNet synonyms and hypernyms
- Latent Semantic Analysis to build word similarity
matrices
25In Summary
- WordsEye Scenario-Based Lexical Knowledge
Resource (SBLR) - Amazons Mechanical Turk our tasks
- Manual normalization of AMT data
- Automatic normalization techniques used on AMT
data and results - Possible future automatic normalization methods
26Thanks to
- Richard Sproat
- Masoud Rouhizadeh
- All the CSLU interns
27Questions?