Lexical Attraction Analysis using History and Structure
Humans are capable of learning languages without any previous knowledge of grammar, syntax, et cetera. Deniz Yuret's model of language acquisition using lexical attraction is a first step toward modeling this human process, but its proof-of-concept nature sabotages its accuracy to an untenable level. By extending the lexical attraction model to include probabilistic analysis of sentence structure and word usage histories, I hope to increase the accuracy of the model in order to show the validity of lexical attraction as a means of modeling human language acquisition.
A complete research project (idea, implementation, testing, analysis)
Steps
Implement a probability model for sentence structures using lexical attraction.
Implement a way of using usage history for words in the lexical attraction model, possibly by implementing different link types for lexical attraction.
Implement lexical analysis using the probabilistic structure-space and a probabilistic history-space in conjunction with a similarity metric (Coen 2006).
Test the modified model for improved accuracy (compared to the original model), and analyze any results (or lack thereof).
Contributions
Proposal of a method for improving the accuracy of lexical attraction analysis using both sentence structure and word usage history.
Implementation of probability spaces for sentence structures and word usage histories.
Implementation of a similarity metric between probability spaces to help determine the likelihood of lexical attraction between words.
Testing to determine the effects of the new methods in the lexical attraction model.
Analysis to specify the root causes of any non-negligible effects of new methods.
Conclusion concerning the viability of lexical attraction in light of any changes in accuracy in the modified model.
Raymond Cheng
Story Representation in Gauntlet
The Gauntlet system shows promise in becoming a general method for communicating and storing knowledge in language.
However, current implementations lack a method of storing higher-level stories.
My project seeks to produce a form of representation for stories and a memory that can easily recall stories using intermediate features.
A complete research project with proof-of-concept demonstration
Steps
Define a representation of stories and story memory that is conducive to search and recollection.
> br/> Demonstrate this representation with a number of stories.
Contributions
Will propose a representation of stories and story memory using intermediate features.
Will demonstrate this system in the Gauntlet system by feeding in simple stories and recalling stories.
Harold Cooper
Verb Classes for Trajectories
Jackendoff argues that syntactic regularities reveal deeper semantic regularities. I will search for syntactic regularities in a corpus of trajectories, focusing on verb classes based on trajectory structure. Since trajectories capture primarily the spatial aspects of verbs, these classes may help lead to spatial — classes. This new trajectory knowledge could also bootstrap better trajectory parsing.
A complete research project.
Steps
Build a corpus of trajectories, probably using Mark Seifter's trajectory parser.
Analyze the consistency of existing verb classes, such as Levin, LCS, and/or FrameNet, within the trajectory corpus.
Implement a clustering algorithm for verbs based on similar trajectory occurrences. If structure alone seems insufficient, use lexical attraction information as well.
Compare learned classes to existing verb classes.
Time permitting, try to use verb classes to improve trajectory parser.
Contributions
A new classification of verbs.
An analysis of verb use among trajectories.
A clustering algorithm on trajectories.
A corpus of trajectories.
Possibly, an improved trajectory parser.
Shirley Fung
Case Based Reasoning in Copyright Law
Lawyers often have to construct arguments for a case using precedents from past rulings.
Finding the precedents require in-depth knowledge of a large database of cases. I am
interested in Copyright Law, especially when it has been known that globalization and
advancement in technology have make it harder to predict how rulings will turn out. I
would like to use a past project done on Trade Secrets Law, to create a similar system
that works in the domain of Copyright Law. In turn, I wish to understand the dynamics
and changes in Copyright Law, but studying how decision patterns may have changed over
the years.
A complete research project with proof-of-concept demonstration
Steps
Create a database of case briefs in Copyright Law
Define a frame for the domain of Copyright Law
Present a system design of the search algorithm used for finding best matches
Show that it is possible to find good matches of precedents by validating test cases
Implement a working system
Analyze where the system fails, and why it did
Contributions
Will apply case based reasoning in a domain that has not been done before.
Will develop an appropriate frame for the domain of Copyright Law.
Will develop a proof-of-concept demonstration of the case based reasoning system.
Will provide a way for lawyers to better understand Copyright Law.
Aseem Kishore
Exploring the efficiency of thread memory
A re-implementation.
For their original paper on thread memory, Vaina and Greenblatt implemented their theory as a LISP program. They stated that the program was able to conceptualize nominals, to answer questions about them, to make deductions and to remember them. However, they make no mention about how efficient thread memory is in terms of space and time. Specifically, how much information is a machine able to practically learn in this way, while tying existing information together? I am interested in exploring the performance of thread memory in this and similar aspects.
Steps
Implement a crude but functional architecture for thread memory, using the approach detailed by Vaina and Greeblatt.
Implement a program that uses this architecture to store and access information.
Train this program in a similar manner as Vaina and Greenblatt, using children's books and other elementary examples.
Profile the program's time and space efficiency for each new input (e.g. sentence) as the number of inputs grows.
Profile the program's time efficiency in answering questions about information it has learned, as a function of the question's difficulty.
Possibly let the program crawl the web, selectively choosing to take as input elementary sentences it can understand, and profile how much information it's able to generate on its own relative to how much it explicitly learns through input.
Contributions
Will provide an alternate implementation of Vaina and Greenblatt's thread memory theory, in a more up-to-date language (either Python or Java).
Will shed light on thread memory's practicality in terms of efficiency, both in learning and in answering questions.
Reid Kleckner
Experiments with thread memory in Gauntlet
Gauntlet claims to use thread memory to support the organization of its
thought, but it does not have a proper bundling mechanism to allow it to
make any of the inferences suggested by Vaina and Greenblatt in their
original paper. In order to extend Gauntlet to make those inferences, it
needs a mechanism for bundling threads.
Steps
Write an implementation of the bundling algorithm for Gauntlet's
threads.
Extend Gauntlet so that it can make inferences from this information.
Verify or dispute the findings of Vaina and Greenblatt.
Contributions
Will add bundling implementation to Gauntlet.
Will allow Gauntlet to use bundling to make inferences.
Will demonstrate the extent of the usefulness of thread memory.
Duks Koschitz
Curved Crease Origami
Little work has been done since Dr David Huffman a great pioneer of
computational origami began to investigate the possibilities of curved
creases. This work is a study of curved crease origami and focuses on
quadratic curves. The examples will be based on 2-dimensional crease
patterns that can be folded into 3-dimensional shapes.
A research project with hands-on digitally manufactured case studies
Steps
Produce 2-d drawings of curved crease patterns.
Create a catalog of pattern types and test all drawn examples.
Contributions
Hands-on experimentation can yield interesting results for this field.
A structured attempt of doing this kind of expermentation can evolve into a
collection of interesting case studies for the field.
The fabrication techniques explored for the experiments might be
scalable and useful for furniture scale production.
Thomas Larsen
Vision-Based Word Sense Disambiguation
When faced with the sentences
1. He ran into the bank
2. He ran along the bank
a human knows that the word bank is probably being used as a financial institution in the first sentence, and as the side of a river in the second sentence. If
we are to understand human-level intelligence, we must understand how the language and vision interact to disambiguate among possible word sentences.
Sufficient vision technology does not yet exist, so my system will use a human surrogate instead of a vision system.
A complete research project with proof-of-concept demonstration
Steps
Understand existing Gauntlet framework (done)
Identify the types of questions a computer would ask a vision system in order to disambiguate among word senses.
Implement these questions into Gauntlet, with a human surrogate supplying answers.
Contributions
Will demonstrate the power of vision in disambiguating among word senses.
Using the Gauntlet framework, will demonstrate vision-based disambiguation on example sentences.
Jimmy Li
Tracking trajectories using log-polar plots
We shift the focus of our eyes rapidly from point to point in order to
observe a scene. Yet when we track a moving object or when we trace a
contour, we are able to steadily train our focal point along the
trajectory of the path or contour. How are we able to do this? A
log-polar transformation of the images we perceive may hold the
answer.
A research proposal
Steps
Understand log-polar mapping
Understand how Cartesian translation maps to log-polar form
Develop a technique for tracking Cartesian translation using log-polar coordinates
Test the model by asking it to track the motion of a ball
Contributions
Will propose a model that explains how the human visual system
is able to trace trajectories
Will develop a proof-of-concept program that demonstrates this model
Val Morash
Attention as an amodal representation of space
It has been proposed that human representation of space is amodal
(without a link to any specific sensory modality), because blind
persons are able to perform as well as sighted persons on spatial
tasks. Rao proposed that patterns of attention could be used as an
amodal spatial representation. I will examine to what extent patterns
of attention can be used to identify spatial configurations within and
across senses.
Experiment with human subject participation
Steps
Create stimuli consisting of objects in various spatial configurations.
Record eye and hand movements as subjects explore these objects
visually and haptically.
Identify patterns of exploration (from object to object) that
are consistent within and across senses.
Determine the percentage of each configuration captured by
consistent attention patterns.
Contributions
Propose a model, based on patterns of attention, for
representing spatial representations.
Determine the extent to which this model is effective for within
and across sense representation.
Lev Popov
Incorporating Lexical Attraction into PCFG Parsing
Lexical sentence parsing is one of the central problems in NLP, with a number of various
approaches created to solve it. Lexical Attraction concept proposed by Yuret is an
interesting alternative solution to this problem, achieving surprising results even in the
absence of any grammar representation. I propose to create a hybrid parser, utilizing both
traditional PCFG techniques and lexical attraction information to study whether these
methods can be effectively combined to boost parsing performance.
A complete research project with proof-of-concept demonstration
Steps
Implement or reuse a Lexical Attraction parser as described by Yuret
Implement or reuse a basic PCFG parser (from nltk)
Design a hybrid parser utilizing both of the techniques
Compare the individual performance of the separate parsers and test them agains the hybrid parser.
Contributions
Will design and implement a new, modified approach to lexical sentence parsing based on work of Yuret and basic PCFG parsers
Will study the interaction of the individual parsers, and study the effectiveness of combining them into a hybrid parser
Russell Ryan
Gait Identification with Cross-Modal Clustering
Cohen exposes the utility of cross-modal clustering for identifying how vision
and audio cues are combined in the learning process. In what other problem
spaces would this technique be effective? Different gaits (e.g. running or
walking) have similar characteristics. Running is simply a fast version of
walking. I aim to investigate whether the use of Cohen's cross-modal clustering
on two different aspects of gait, the movement of the upper and lower body,
will allow a computer to differentiate between different kinds of gaits.
A research proposal with proof of concept implementation
Steps
Define the two 'modalities' of gait to focus on clustering.
Codify approximation techniques for measuring these modalities from data.
Acquire recorded data of various different types of human gaits.
Develop a proof-of-concept program to attempt to cluster the two modalities based on the evaluation methods developed.
(Optional) Use a visualization library (e.g. Processing, Prefuse, Pyx, etc.) to graphically represent the results.
Contributions
Will analyze different charactistics of gait and present two 'modalities' which can be potentially clustered.
Will codify an determinstic method of numerically evaluating the two modalities.
Will develop a proof-of-concept program that evaluates the two modalities
in a data set and attempts to cluster the data.
Raphael Rush
Project Metaphor
Many definitions are implicit. For example, in the statement, "we need to hammer this point home," the meaning of “hammer” in this context (to reinforce something) is inferred from the fact that its object (a “point”) is an abstract object. I will use the thread memory model of word memory to demonstrate how metaphorical (i.e. abstract) meanings can be inferred from knowledge of concrete definitions.
A complete research project with proof-of-concept demonstration. If I can't get it working in time, I will present a progress report.
Steps
Define
Show how it is possible for a given state to hallucinate the next state, in an iteration of memory accesses. Develop a notion of transition.
Contributions
Will propose a model, based on thread memory, for metaphor recognition.
Will explain the process of inferring metaphor meaning within the context of that model.
Will develop a proof-of-concept program that infers the meaning of a metaphor from context and prior knowledge.
Adam Seering
Constructing Linguistic Trajectories
Jackendoff's linguistic trajectories are a good way of analyzing sentences. I believe that they can also be used to analyze physical situations where an object is moving.
Research project with fully-functional implementation code
Steps
Create computer data structures that can store Jackendoff trajectories and physical trajectories of sets of objects
Design an algorithm that can transform physical trajectories into Jackendoff trajectories
Implement this algorithm
Write a nice UI around this algorithm, so that it can be used by nonprogrammers
[PENDING TIME] Design and implement an algorithm that produces a sentence based on a Jackendoff trajectory
[PENDING TIME] Design and implement an algorithm that, given an array of physical trajectories and sentences describing those trajectories, will develop its own Jackendoff-esque frame and method for converting both physical trajectories and sentences to and from such a frame.
Contributions
Algorithm that can transform a physical trajectory into a Jackendoff trajectory
Implementation of this algorithm; some tests to verify that the algorithm works
User Interface around this algorithm, so that nontechnical users can try it out and so that it can be demo'ed more easily
