Transcript Slide 1

Research Overview
Research Focus
Learning and Inference Paradigms for Natural Language Understanding
Our research focuses on the computational foundations of
intelligent behavior. We develop theories and systems
pertaining to intelligent behavior using a unified methodology, at
the heart of which is the idea that learning has a central role in
intelligence. We investigate the role of learning in supporting
intelligent inference and use this understanding to develop
systems that learn and make intelligent inferences in complex
domains. Such systems must acquire the bulk of their
knowledge from raw, real world data, and behave robustly when
presented with new, previously unseen, situations.
We have focused on the problems of natural language
understanding and intelligent access to textual information,
developing the underlying machine learning theory needed to
make progress in these challenging areas hand in hand with
developing state-of-the-art software tools to solve a range of
Natural Language Processing tasks.
Research Expertise
Our work addresses foundational questions in learning,
knowledge representation and reasoning, experimental
paradigms and large scale system development, drawing on
methods from theoretical computer science, probability and
statistics, artificial intelligence, linguistics, and experimental
computer science. We are driven both by longer term goals to
understand and develop capabilities for natural language
comprehension, and by challenging shorter term applications in
the area of information extraction and knowledge access.
In terms of "traditional" research areas our work falls mostly into
Natural Language Processing and Machine Learning, as
well as Information Extraction and Data Analytics in general.
A collection of Classifiers; Log-linear
models (HMM, CRF) or a combination
Transliteration Target Task –
alignment structure
Penalty for
violating
the constraint.
(Soft) constraints
component
Weight Vector for “local” models
How far y is from
a “legal”
assignment
Making complex decisions in real world problems often involves assigning values to
sets of interdependent variables where the expressive dependency structure
influences what assignments are possible. Constrained Conditional Models (CCMs;
a.k.a. Integer Linear Programming formulation of NLP problems) is a learning and
inference framework that augments the learning of conditional (probabilistic or
discriminative) models with declarative constraints as a way to support decisions
in an expressive output space while maintaining modularity and tractability of
training and inference and facilitating the use of background knowledge.
Predicted
Negative examples cannot
have a good structure,
Correct
The feasible structures
of an example
Negative examples restrict
the space of hyperplanes
supporting the decisions for x
Tracking Entities across Documents
 Dan Roth
Postdoctoral
Researchers
 James Clarke
 Yee Seng Chan
 Shankar Vembu
Research Scientist
 Mark Sammons
Research
Programmers
 Joshua Gioja
 Michael Paul
 Ming-Wei Chang
 Michael Connor
 Quang Do
 Dan Goldwasser
 Gourab Kundu
 Jeff Pasternack
 Lev Ratinov
 Nick Rizzolo
 Alla Rozovskaya
 Vivek Srikumar
 Yuancheng Tu
 V.G. Vinod Vydiswaran
Undergraduate Students
 Nikhil Johri
I
t a l y
‫ה י ל ט י א‬
‫ה י ל ט י א‬
Find Obama in the Hebrew Wikipedia
Learning to predict structure requires highly specialized supervision – a major
challenge in scaling machine learning techniques to real world natural language
processing problems. We investigate methods that allow learning models for
structured prediction using an indirect supervision signal which is considerably
easier to obtain and suggest indirect learning protocols for common NLP learning
scenarios. We achieve this by identifying companion problems, which allow
straightforward induction or generation of binary labels for the same inputs as
the structured prediction problem. This binary signal is used in our LCLR learning
algorithm as feedback for a structured prediction component, using the intuitive
constraint that positive labels from the binary signal must correspond to wellformed structures in the structured prediction problem, while negative binary
labels must correspond to poorly-formed structures. Our approach achieves
improved performance on problems such as transliteration, paraphrasing,
recognizing textual entailment, and semantic parsing (inducing a logical form
from natural language text).
Modeling Children's Language Acquisition
The BabySRL Project: Using NLP and Machine Learning techniques, we are
investigating models of language acquisition by children. We propose that
shallow but abstract representations of sentence-structure guide early sentence
interpretation, and have built computational models that mirror key observations
of child language learning. Our joint work with a team of psycholinguists at the
University of Illinois focuses on the acquisition of verb-argument structures by
investigating simple language features involving the number and order of nouns in
a sentence.
We Are...
Head
t a l y
Transliteration Companion
Task– binary supervision
Yes/No
Some of our more recent foci include the development of
models that bridge our work on Semantic Parsing and
Psycholinguistics research on Natural Language Acquisition,
and a new line of work on Natural Language Understanding
in Context – transforming natural language instructions into
actionable models. A key example of the latter direction is a
program that learns how to play strategic games by following
written instructions.
Graduate Students
I
Assessing Trustworthiness
Identifying entities in text and linking different mentions of those entities to the
correct individuals they could represent, is an open research problem. We have
applied our machine learning expertise to develop applications for recognizing
when a given word or phrase represents an entity of interest (NER), and to
resolving co-referent mentions within and across documents and disambiguate
mentions of entities with similar or identical names (co-reference resolution).
Our Wikifier maps phrases in text documents to the corresponding entries on
Wikipedia.
Language experiments with children support the proposal that toddlers build
partial structures that preserve the number and order of nouns in a sentence,
but leave open the possibility that children's behavior stems from different
underlying representations. In our experiments with a computational model of
semantic-role labeling (SRL), whose representations of sentence-structure are
under our control, we simulate the learning process using unlabeled data and a
seed set of concrete nouns. By using only the shallow structures consistent with
the proposed language acquisition model, we can assess the model’s predictions
on unseen data (sentences with verbs not used in the training process). With the
ordered set of noun representation, our system reproduces errors seen in child
experiments, providing an independent source of evidence for the
psycholinguistic analysis.
Textual Inference and Knowledge
Representation
Our research into textual inference, a core capability required by many NLP applications, has led us
to develop intermediate knowledge representations that facilitate learning and inference in
complex domains. In addition to multi-view data representations supporting flexible run-time
inference, we have formulated a model that encapsulates specialized knowledge capabilities as
metrics. These compare arbitrary constituents of text based on standard NLP analytics, simplifying
development and integration. The diagram to the right shows a multi-view representation of a
textual entailment pair, together with color-coded edges corresponding to the outputs of
specialized metrics. The green edges linking SRL constituents indicate a positive match, while the
red edges connecting Numerical Quantity constituents indicate that the metric reports a mis-match.
Although much work in NLP has focused on determining what a
document means, we also must know whether or not to believe it. We
have started to develop techniques to assess both the reliability of
sources of information on the internet – organizations and their
associated web sites, or the writings of individual people – and to assess
the reliab-ility of individual statements in a domain of interest (such as
the medical domain). Our framework makes use of CCMs and allows us
to incorporate users’ prior knowledge into TruthFinders algorithms.
This work complements research on
Evidence Search and on detecting
sensitive information in large
document collection, combining
analyses of the text content and of
the link structure connecting these
documents on the internet to assess
the trustworthiness of different
sources, and even of individual
statements.
Machine Learning Support for NLP and
Information Extraction
We have developed Learning-Based Java to make programming with
Machine Learning paradigms easier for non-expert users. We have also
developed state of the art NLP and Information Extraction tools that are
widely used. You can download our software and see real-time demos at:
http://L2R.cs.uiuc.edu/~cogcomp/
Learning from Environment Response
Our Reading in Context work investigates a new approach to semantic
interpretation. We assume an external actionable context such as the real
world or simulated environment and text containing instructions for a task
in that actionable environment. We evaluate the ability of a reasoning
agent to interpret the text by observing the actions it takes in that context,
that its performance after reading the instructions.
Natural Language Instruction “...there are only a few
rules, like for example, you can play any card on an
empty freecell. You don't want to do this too much or
you'll run out of freecells and get stuck..."
Knowledge Base
Game API:
source of
Labeled
solitaire
moves
value(x,
y) homecell(x
) suit(x,y)
freecell(x) top(x,y)
empty(x)
Task Decisions
Learner (TDL)
Language
Learner(LL)
Interpretation
card(x)
 empty(y)
 freecell(y
)
A straight forward approach to this problem is to convert the textual
instructions into complete logical formulas that can be used as a rule
based system in the relevant world context. Unfortunately semantic
parsing, the process of converting text into its logical interpretation, is a
very difficult task, and providing the supervision required to train a
semantic parser for each domain is infeasible. When a semantic parser is
available, it will generate some fraction of incorrect formulae, meaning
that the induced knowledge will be noisy and/or incomplete.
We introduce a new approach in which the interpretation is not used
directly in decision making but rather integrated into a real world
learning task, thus allowing the interpretation to change according to
real world response to the actions of the learned model. Our
experimental setting is an automated Freecell Solitaire player, which
augments its knowledge resources using rules induced from natural
language text instructions available on the world wide web. The system
converts text to logical formulae fragments, which are used as features for
task learning, yielding a representation of the rules of Freecell solitaire.
We evaluate our model by observing the task improvement due to the
generated features, and observe significant improvement due to the use
of natural language instructions.
Our Research is supported by NSF, NIH, DARPA, DHS, ONR as well as Boeing, Motorola, and Google.