Pre-symposium research session, November 8

Without a good foundation, there can be no good building. Sentiment analysis, while growing in popularity, and undoubtedly a useful technology, is a large and complicated building set upon a rickety foundation. Sentiment is not, as is generally assumed, a simple distinction between "positive" and "negative" (and sometimes "neutral"). It is not hard at all to find many real-world counter-examples which show the limitations of such a view, convenient as such simplifying assumptions may be for developing computer models. As I will discuss, these limitations have very real (negative) effects on the development and evaluation of sentiment analysis systems; unless the real complexities of sentiment are taken into account, estimations of analytic accuracy are nearly meaningless. The news is not all bad, however ? I will also sketch a new structured approach to defining the problem which can help address sentiment analysis more fully. (Joint work with Dr. Kenneth Bloom, now at Google Cambridge.)

Our research explores the challenges and opportunities of working with multi-dimensional sentiment data. We begin with a broad data-set from the Experience Project. Users chose one of over 130 emotions to describe their mood and accompanied the choice with a short description. We leverage the fine-grained, multi-dimensional labels and over 1 million data instances to train a multi-dimensional emotive scoring system that can be applied to any text. Our research evaluates methods for reducing sets of labels to a smaller number of defined emotive axes along which to score text and explores the benefits of scoring social media data along nuanced, emotive axes.

Sentiment analysis techniques are being used in finance, marketing and survillance. However, little is known about how sentiment articulated by an individual or group impacts another. It is well known in financial trading that sentiment is 'absorbed' in prices and in the volume of trading. This presentation is based on academic research (Ahmad) and experience in financial markets (Sciro) which relates to the behaviour of stakeholders in financial markets. We will explore new evidence from neuropsychology, and its exotic progeny neurofinance, which suggests that there is rational herding and behavioural herding in financial markets. Herding typically involves a collective behaviour which has evolved for species protection and is seen also in risk averse/risk seeking behaviours. In this lecture we will describe key findings in the research literature (Ahmad) and observations based of many years of equity and commodity trading (Sciro) which appears to link sentiment with herding behvaiour.

The explosion of Web opinion data has made essential the need for automatic tools to analyze and understand people's sentiments toward different topics. In most sentiment analysis applications the sentiment lexicon plays a central role. However, it is well known that there is no universally optimal sentiment lexicon since the polarity of words is sensitive to the topic domain. Even worse, in the same domain the same word may indicate different polarities with respect to different aspects. In this talk, we focus on the problem of learning a sentiment lexicon that is not only domain specific but also dependent on the aspect in context given an unlabeled opinionated text collection. We propose a novel optimization framework that provides a unified and principled way to combine different sources of information for learning such a context-dependent sentiment lexicon. Experimental results show that our approach can not only identify new sentiment words specific to the given domain but also determine the different polarities of a word depending on the aspect in context. In further quantitative evaluation, our method is proved to be effective in constructing a high quality lexicon by comparing with a human annotated gold standard. In addition, using the learned context-dependent sentiment lexicon has led to improvement in the accuracy in aspect-level sentiment classification tasks.

More and more companies are leveraging sentiment analysis tools to track and dissect how consumers feel about products and services of their own and also of the competition. These tools tend to be based on decades of semantic and Natural Language Processing research from top universities around the world. While algorithms are typically tested against well-known and studied corpuses that experts have agreed upon as ground truth, challenges arise as the tools transition from the lab to the start-up. At Argus Insights we have been a customer of one of the best-in-class toolsets for the past year. After accepting the positive response rate as a point of fact, we began to notice discrepencies based on our core dataset of consumer responses to their user experiences for products and services. Here we will outline our approach to testing the performance of these sentiment tools and detail the startling results that have implications for the entire industry.

As online communities grow and the volume of user generated content increases, the need for community management also rises. Community management has three main purposes: to create a positive experience for existing participants, to promote appropriate, socionormative behaviors, and to encourage potential participants to make contributions. Research indicates that the quality of content a potential participant sees on a site is highly influential; off-topic, negative comments with malicious intent are a particularly strong boundary to participation or set the tone for encouraging similar contributions. A problem for community managers, therefore, is the detection and elimination of such undesirable content. As a community grows this undertaking becomes more daunting. Can an automated system aid community managers in this task? We address this question through a machine learning approach to automatic detection of inappropriate negative user contributions. Our training corpus is a set of comments from a news commenting site that we tasked Amazon Mechanical Turk workers with labeling. Each comment is labeled for the presence of profanity, insults, and the object of the insults. Support vector machines trained on this data are combined with relevance and sentiment analysis systems in an ensemble approach to the detection of inappropriate negative user contributions. The system shows great potential for automated community management tools.

A new theory of emotions is derived from the semantics of the language of emotions. The sound structures of 36 Old Arabic word roots that express specific emotions are converted into abstract models. By substitution from two tables, abstract models are converted into concrete theories about the nature of the specific emotions that are likely to be validated. Theories confirmed by the author's own emotional experience (self report) and by previously corroborated theories, are considered corroborated. These theories about specific emotions are woven together into an integrated theory of all emotions. The theory models emotions and emotional mechanisms, dimensions and polarities in ways amenable to affective computing. The findings are supported by clinical psychology. Old Arabic is chosen because its words, sounds and meanings are consistent and have not changed for at least 1,400 years. The theory can be expanded by incorporating additional emotional word roots from Arabic and other alphabetical languages.