Research & Innovation in Sentiment Analysis
May 7, 2013, 8:45 am - 12:30 pm, New York
Research & Innovation is a special, half-day session that proceeds the Sentiment Analysis Symposium. It is intended for very advanced users and consultants, developers, and academic and industry researchers.
View presentation videos from the similar fall, 2011 research session online.
The session will take place in the Park Avenue Room, 6th floor of the Lighthouse International conference center, 111 East 59th Street, New York, NY.
Welcome: 8:45 am - 8:50 am
Session 1: 8:50 am - 10:30 am
- Vasudeva Varma, Professor/Chief Scientist, IIIT Hyderabad/SETU Software
Social Media Search and Analytics: Dealing with Surface and Semantic Noise
(mouseover here for description)
Social media has become a very good proxy for the real world. In many domains, the opinions and sentiments expressed in social media capture the pulse of people in the
society. Thus, we are increasingly able to rely on the voices of the people in the social media to form opinions and take actions. We need accurate social listening tools and applications
to capture, interpret and summarize the social media comments.
Building powerful social listening, search and analytics tools that go beyond keywords and co-occurrences involves solving various research challenges. In this talk, I will focus
specifically on social media noise at the surface level and at contextual or semantic level, which impact accuracy, recall of search and mining for opinions and sentiments from social
media. I will elaborate on Sentiment Expression Axis that ranges from literal expressions to smileys to sarcasm to spam and various sub problems in dealing with social media noise.
Surface noise and semantic noise are two different beasts. Sometimes simple statistical models and machine learning techniques may be sufficient to solve surface level noise, which
includes spelling variations, sentence fragments, and concept variations. These techniques typically help in canonicalization of surface forms. However, semantic noise is hard because to
filter the signal from the noise, the context and the world knowledge are necessary. Linguistic approaches to deal with sarcasm, double negation, and metaphorical usage may not be
sufficient (and robust enough) to deal with social media noise. I present an analysis of these significant real world problems that we uncovered from huge real world social media data.
- Samuel R Dismond III, Chief Scientist, 7INSIDE, Inc.
The Terma: Grounding Sentiment beyond Polarity
(mouseover here for description)
The Terma analyzes utterances in terms of underlying related canonical inner experiences. Terma canonical inner experiences are map-able to generally-accepted biological
functions/tissues. This mapping grounds the Terma result making it universally understandable. The Terma sentiment analysis system is comprised of an Inner Experience Semantic Network
(IESN) and a custom parser that processes personal or social unstructured English text. The Terma provides a depiction of sentiment that is easily recognized as a personal visceral
experience. Therefore, intuitive straightforward actions are easier to identify. This is because one or more specific actions are always fundamentally related to all canonical inner
experiences in the IESN.
- Claudio Pinhanez, Manager, Service Systems Research, IBM Research - Brazil
Modeling and Simulation of Sentient Messages in Social Media
(mouseover here for description)
In this session we will present the Social Media Simulator (SMS), a novel tool that allows the exploration of the impact and result of social media marketing analytics and
actions in online social networks like Twitter or Facebook. The SMS samples an egocentric social network and models the behavior of its nodes with regard to sentiments, topics and users
actions. With this fine-grained social media network model, we explore usage behaviors such as the effect and propagation of positive and negative messages and message decay prior to
action in the real network. The tool is demonstrated using data from the Obama/Romney Twitter networks of the 2012 elections.
- Jennifer Carlson, Scientist, Decisive Analytics Corporation
A Simple Method for Intersecting Affective and Topical Space
(mouseover here for description)
In this talk, we present a simple, flexible, and efficient method for measuring the degree to which topics of discussion are expressed with different dimensions of affect.
Our analysis technique combines statistical topic modeling with a collection of methods of extending, combining, and constructing specialized lexica specific to different categories of
affect. Statistical topic modeling algorithms, such as LDA, are unsupervised methods of discovering the distinct threads of discourse existing in text. We combine and extend existing
lexica by building a multi-faceted affect lexicon that expresses finer levels of affect than simple positive or negative sentiment. Our experiments use a lexicon that expresses
specialized concepts such as fear and anger. By examining how the distribution of words in topics aligns with types of affect defined in our lexicon, we can build a fine-grained model of
the distribution of affect across topics. This technology is inherently unsupervised and does not rely on preconceived ontologies or rating systems. Our lexicon can incorporate third
party lexica or can be extended to model a particular domain or problem of interest. Unlike more complicated approaches of topic modeling that require customized probability models to
extend the topic models into a new dimension, we are able to adapt to new problems spaces without the need for changes to the topic modeling engine. Because of limited commitments to
language specific resources and techniques, our approach is language agnostic. We will present results of experiments on Chinese, Arabic, and English corpora.
Session 2: 10:50 am - 12:30 pm
- Zornitsa Kozareva, Research Assistant Professor, University of Southern California, Information Sciences Institute
Multilingual Sentiment Analysis of Metaphor-Rich Texts
(mouseover here for description)
Understanding metaphor-rich texts like "Her lawyer is a shark.", "This idea is old hat.", "We need to construct a strong argument." among others and the affect associated
with them is a challenging problem, which has been of interest to the research community for a long time. The questions I will address are: "How can we build an automated system that can
identify the polarity and valence associated with metaphor-rich texts?" and "Is it possible to create such a system for multiple languages and domains?"
In this talk, I will introduce the task of multilingual sentiment analysis of metaphor-rich text and will present novel algorithms that integrate cogitative, affective, perceptual and
social processes with stylistic and lexical information. Finally, by running evaluations on datasets in English, Spanish, Russian and Farsi, I will show that the developed sentiment
analysis technology is portable and works equally well when applied to different languages.
- Kyle Dent, Senior Researcher, Palo Alto Research Center
Subjectivity and Relativism in Opinion Mining
(mouseover here for description)
This presentation will demonstrate an aspect of sentiment detection that makes it especially difficult for any rule-based approach. Understanding opinions can depend on a
particular point-of-view, and even more subtly should be considered relative to a particular question for a given analysis. In the most obvious case, when classifying a comment that
compares my product to a competitor's, my positive assessment will be exactly the opposite of my competitor's. The more subtle case is demonstrated by a sentence like, "I use boost
mobile because I pay my own bill." If reviewing this sentence for brand analysis, I would probably consider this a negative comment since if the customer had more money, he would select
a different service. On the other hand, if I were reviewing comments for price, I would likely classify it as positive.
The presentation will start with a brief review of better known difficulties in performing sentiment analysis with several samples showing hard to anticipate language affecting polarity.
Then I will demonstrate that polarity of opinion is perspective-sensitive using the example showing how one company's classification would be exactly the opposite of their competitor's;
then moving to the more subtle examples where an individual analyst could view the same comment as positive or negative depending on the question being asked of the data. If time, I can
also show part of a system we have developed that provides an interface for users to train a machine learning algorithm for different types of anaylses to answer different questions.
- Alessandro Oltramari, Scientific Consultant, B-Sm@rk
MySmark: A Tool for Emotional Semantic Tagging
(mouseover here for description)
Standard computational languages (RDF, OWL, etc.), vocabularies and ad-hoc specifications like EmotionML provide a powerful infrastructure to represent opinions and
sentiments in a machine-understandable format, but current state-of-the-art rating systems and tools such as the "+1" or "Like" buttons are far from providing a comprehensive
cognitive model of user's experience. Next generation tools need to go beyond the simplistic positive/negative dichotomy, allowing users to express a wide range of feelings and easily
associate them to knowledge contents. MySmark is the B-Sm@rk B2B and B2C solution to this problem: it aims at filling the gap between commercial semantic services and user's
evaluations, enabling emotional tagging of contents in the Semantic Web framework. Based on Plutchik's "wheels of emotion" and Mehrabian's PAD models, MySmark is a cognitive-grounded
evaluation platform that improves users's emotional feedbacks with semantic technologies.To design on-line business services that can dynamically adapt to user's experience, feeling
and knowledge, encoding semantic contents is a necessary but not sufficient condition: beliefs, goals and emotions come into play when consumers decide what to search, purchase and
possibly broadcast diversifying among the circles of friends, acquaintances or just other consumers. Business services need to integrate models of human cognition with semantic
technologies to capture this level of complexity (in general, this is the requirement for achieving a "Cognitive Semantic Web"). In this presentation we focus on these topics,
claiming that MySmark represents a good example of a platform that integrates cognition in the semantic web framework.
- Richard K. Merritt, Professor, Luther College
Cognitive Biases, Sentiment Analysis, and Influence Engineering: Mathematical and Visual Models for Persuasion Design
(mouseover here for description)
Cognitive Biases, Sentiment Analysis and Influence Engineering: Mathematical and Visual Models for Persuasion Design.
Description= Social, cognitive, decision, behavioral and emotional biases form the strongest set of data paradigms through which we can form identity resolutions and predict mindset and
motivation of subjects. These seemingly ethereal entities are the substance of designing the tools of influence. Similarly discerning sentiment across disparate data sources, social
networks and social trends is inextricably linked to understanding and predicting core biases.
Cognitive Biases, Sentiment Analysis and Influence Engineering: Mathematical and Visual Models for Persuasion Design addresses a model through which one can influence engineer for
increased probability of optimal results. Using Bayesian probability, Markov chaining, and research on cognitive biases the author establishes a mathematical and visual model for analysis
of data hewn from, social networks, geopolitical data, and demographics. Fundamental to this enterprise is limiting the most salient and influential behavioral, social and decision biases
to subject motivation.
Using contemporary research on persuasion, psychology and manipulation, this study defines the underlying dynamic design structures of influence. From the modern political campaign,
social networks and advertisement, human actions as sociopolitical agents and consumers can be intermediated by a contemporary influence engineering practice. It is the contention of this
author that language, image, and public relations can be more effectively inter-engineered to produce desired outcomes in consumers through the integrated and applied use of sentiment
analysis and specific psychological, mnemonic and aesthetic strategies.
Registration
You may register for the research session alone or in combination with the tutorial and/or the symposium. Visit the registration page.
Register Today!
|
 |
|
|
