The program of Wednesday, March 5, 2014 Human Analytics workshops is outlined on this page. The workshops are your way to learn from experts about the solutions and technologies you need to compete in today's high-velocity, hyper-competitive business climate.
Visit the Symposium Agenda page for information on the March 6, 2014 conference presentations and panels. Reminder: Attend either day or both.
Insider's Guide to Social Media MeasurementMorning workshop: 9:00 am to 12:30 pm
Welcome to the world of "humetrics." Workshop instructor Stephen Rappaport coined that term to describe the big shift from our industry's age-old preoccupation with media measurement to understanding people by gauging and interpreting their digital lives. Rappaport's forth-coming "Digital Metrics Field Guide" (from which this workshop description is derived) helps us recognize data points not merely as impersonal dots on a trend line, percentage changes, or ratios. They are, in fact, personal -- capturing what people say, do and feel in real time. Once we view digital metrics as reflecting individuals, they become characters we employ to craft compelling narratives about people and brands that we later share with our colleagues in and outside of our areas. Those narratives fortify brands with a common understanding that increases the potential to act in the best interests of customers and prospects, and to create and execute successful marketing strategies.
The Insider's Guide to Social Media Measurement will have four sections:
- introduction and overview of social media measurement
- case studies illustrating the use of social media measures (with audience participation and comments)
- hands-on workshop. We will present three typical business scenarios and have the attendees work through them, including developing an analytic strategy, measurement plan, etc.
- wrap-up and final Q&A
Stephen Rappaport will be joined in presenting the workshop by Vincent Santino, Associate Director of Digital Insights & Analytics at Kaplan Test Prep; Peter Fontana, Research & Insights Director at the social media agency We Are Social; and Maribel Lara, Account Director at M80.
The Road to Customer IntelligenceAfternoon workshop: 1:30 pm to 5:00 pm
The Road to Customer Intelligence: Data, Analytics, Insight
Customer intelligence is the key to deeper and more profitable customer relationships, an essential asset for strategic, competitive decision making. How do you get there? The Road to Customer Intelligence: Data, Analytics, Insight, taught by industry leader Steve Ramirez, will teach you how.
The Customer Intelligence workshop will offer a thorough, practical look at elements that business analysts, managers, and executives must master to compete in today’s hyper-competitive, high-velocity, Big Data world: Unstructured Data, Voice of the Customer, Social Customer Insights, and Predictive Analytics.
How much of your data is your organization using to create insights about your customers? For many companies, an explosion of unstructured data sources like support email, call center notes, and social media, hasn't resulted in a corresponding windfall of better customer understanding, much less an improved customer experience or increase in revenues.
The phenomenon of Big Data has captured the imaginations of executives across industries. This workshop will examine a data-science case study. This cutting-edge Big Data analysis married text analytics of unstructured customer feedback, large-scale processing of transaction data, and the application of machine learning algorithms to identify which customer experience issues were most likely to drive churn, and for whom.
Yet text analytics and data analysis tools are only one piece of the puzzle. Without the right Voice of the Customer strategies in place, you’ll wind up with a handful of reports… and little else.Learn from the best practices of some of the most mature and successful Voice of the Customer programs. As a workshop participant, you will be introduced to a straightforward methodology you can deploy immediately.
Key takeaways include:
- How to benchmark your customer experience, marketing, and social insights efforts
- How to accelerate the maturity of your program, helping you to deliver greater value to your internal business partners
- How to prioritize your data acquisition and data management efforts
- What types of analyses you can perform to obtain the richest insights
If you've been challenged with how to derive real, demonstrable ROI from your VOC program, the Customer Insight Analytics workshop is for you.
Practical Sentiment AnalysisMorning tutorial: 9:00 am to 12:30 pm
The Practical Sentiment Analysis tutorial will introduce the concepts of sentiment analysis and opinion mining from unstructured text, looking at why they are useful and what tools and techniques are available. The tutorial is designed for advanced users, developers, consultants, and others who seek to understand the technology behind the tools they're using (or hope to build).
People use language to communicate not just ideas, but also to selectively reveal their internal mental and psychological life. By expressing opinions about ideas, people, and products, a person conveys secondary information about their personality, their values, and their background. Some of this secondary information is explicitly coded in the message, e.g. saying "I love hiking" sends the clear message that the speaker is an outdoorsy sort; however, other properties of the speaker can be gleaned from the general topics they discuss and from their sub-conscious linguistic expression, such as introversion and status.
Opinion mining is (by now) a well-known natural language processing technique that generally focuses on the explicit portion of opinion expression. Given the great volume of text created and readily accessible online, tremendous value can be derived from this level of analysis, especially for marketers, political campaigns, and the like. Opinion mining itself takes many forms depending on the granularity of analysis desired, from the most basic determination of whether a given document is generally positive or negative to much more specific questions such as whether a given individual is strongly in favor of a given political position based on texts they've authored, their online behaviors and their social network.
This tutorial will dive below the surface of opinion mining in three primary ways. First, we focus on some of the underlying algorithms and the opportunities and challenges for the varied kinds of inputs and outputs involved. In particular, we will discuss semi-supervised learning techniques and their relevance for entity and topic extraction in combination with opinion mining. We will also cover the difference between features used for topic classification and sentiment analysis and those used for stylistic analysis (such as authorship determination). Second, we look at what additional information might be determined from non-explicit components of linguistic expression, as well as non-textual aspects of the input, such as social networks, sensors, and other secondary data. Third, we will also consider issues in the empirical evaluation of automated opinion mining tools.
- Introduction. Brief overview of opinion mining, including motivation and many of the core challenges.
- Machine learning basics. Intuitive overview annotation and learning for text classifiers.
- Aspect-based opinion mining. Identifying the entities and topics that opinions have been expressed about, and building more granular models to analyze them.
- Implementation. From simple bag-of-words algoritms based on sentiment lexicons to full text processing pipelines.
- Evaluation and visualization. Different measures of performance, inter-annotator agreement, and caveats when interpreting performance claims. Visualization techniques for opinion mining.
- Building models with less human supervision. Learning from clusters, features and instances and active learning.
- Going beyond text: social networks, sensors, behaviors, third-party data, audio, and video.
- Further analysis and issues. Predicting authorship, demographics, psychographics, and location. Characterizing and measuring influence, relevance, veracity, advocacy, and volatility. New models for analysis.
Previous tutorials have been taught by Ronen Feldman of the Hebrew University (May 2013); Diana Maynard, research fellow at the Univ. of Sheffield, UK (November, 2012); Bing Liu of the University of Illinois at Chicago (May 2012; Christopher Potts of Stanford University (November 2011); and staff from eBay Research Labs (April 2011).
Technology & InnovationAfternoon presentations & panels: 1:30 pm to 5:00 pm
Session 1 (1:30 pm - 2:55 pm)
- Using Apache Hive for Sentiment Analysis
(mouseover here for description)While there are many ways to complete sentiment analysis, Hadoop is uniquely suited to synthesizing data from a variety of sources.It is this variety that characterizes Big Data, more so than the other famous V's, velocity and volume.Apache Hive is a fantastic tool for processing this data, and Qubole's Director of Product Management, Sadiq Shaik, has some secrets to share on the best practices.If you have experience with or interest in using Hadoop and Hive to leverage your data, this is the session for you!
Rajat Jain, Qubole
- Sequence Package Analysis and Sentimetrics
(mouseover here for description)Learning how customers feel about products and services requires a new natural language understanding method that can accurately measure customer sentiment appearing in call center recordings and social media posts. Given the fact that sentiments are often indirect and sometimes even a bit ambivalent, better methods are needed to perform effective sentiment analysis. Sequence Package Analyis, a new natural language understanding method which has been pointed to in the professional literature as useful in identifying early signs of customer frustation, presents a novel way of performing sentimetrics across different populations of speakers and bloggers. As a result, some of the most complex emotions can be detected and consequently fed back to the enterprise in order to aid in developing more suitable products.
Amy Neustein, Linguistic Technology Systems
- Evaluating Sentiment Analysis Evaluation
(mouseover here for description)The standard means of evaluating sentiment analyzers in academic research is the computation of classification accuracy measures (precision, recall, F-measure, accuracy, etc.) against a gold standard of human-subject judgements, elicited from a (usually small) sample of literate adults.That these subjects could reliably distinguish positive from negative sentiment within a text in principle, at least in non-technical domains, is surely above question.Recently, there has been a great deal of scepticism voiced as to whether the same can be said of sentiment scales that are more refined than a binary "thumbs-up/thumbs-down" classification.Our own work calls into question whether human-subject judgements are at all reliable even in the binary case.
Our starting point is the notion of ecological validity from human-computer interaction research.Humans can reliably perform tasks in a laboratory that are familiar to them outside the laboratory. But when those tasks diverge from everyday life - when the experiment no longer resembles the ecology of their behaviour - their performance becomes more inscrutable, with lower reliability, and lower intersubject agreements.
Using a standard textbook sentiment analysis strategy, and situating our task, for the sake of concreteness, within the financial domain, we have experimental evidence that even binary human-subject rankings may in fact be misleading.Although our observed classification accuracies are in line with published figures, small changes to this textbook strategy that improve accuracy can be shown to damage the applicability of this sentiment measure to stock-trading, when evaluated on historical price data.
Gerald Penn, University of Toronto
- Compositional Sentiment Analysis
(mouseover here for description)Almost all commercial - and many academic - sentiment analysis systems use some kind of supervised machine learning to classify text as positive, negative or neutral. A common complaint among users is that such systems are insufficiently fine-grained, not being able to distinguish, for example, that A has a positive attitude to B but a negative attitude to C, in the same sentence. They are also often found to be inaccurate when faced with sophisticated uses of language such as "I thought I couldn't fail to like this movie, but actually I could".
A line of research at Oxford, and recently also at Stanford University, has been trying to develop "compositional" approaches to sentiment analysis, that is, approaches that are sensitive to the grammatical structure of sentences. These techniques yield fine-grained sentiment profiles and are demonstrably more accurate than common classifier approaches.
This talk will give a non-technical overview of some of these recent research developments, and will demonstrate their utility in some practical applications of compositional sentiment analysis undertaken in Oxford, including tasks as varied as prediction of financial market indicators, bookmakers' odds on horse racing, and "man/woman of the match" decisions in various team games.
Stephen Pulman, Oxford University
Session 2 (3:20 pm - 5:00 pm)
- Emotions are mixed: identifying and visualizing the emotions of topics in public discourse
(mouseover here for description)Collective and individual emotions evolve and shift as people adapt to changes in their environment. These environmental changes - such as major events - are also in a state of flux and impossible to predict.Public discourse is rarely aligned with a singular emotion, but instead is a mixture of different emotional states from disparate groups that are shifting over time. An initial stage of a disaster might elicit public expressions of anxiety and fear, whereas the final stages could bring expressions of relief, joy, or grief.We will discuss ways of simultaneously detecting the topics of discussion in text and the mixtures of emotion that are associated with the topics. Results are displayed in an activation/arousal visualization, where each new topic of discussion will have an "emotional signature". This permits comparison of the emotional content of a large number of events and activities. We will demonstrate this capability over use cases in which emotions are mixed and shift as the event develops and evolves.
Jennifer Carlson, Decisive Analytics Corporation
- Sentiment Analysis by Ensemble Post-Processing
(mouseover here for description)An ensemble method for sentiment analysis is presented that combines three keyword-based methods and a Naive Bayes classifier. While any individual algorithm will struggle with certain text formats, the likelihood is lower that a majority of disparate ensemble members would predict a false sentiment. The benefit of an ensemble approach is exactly that a greater variety of text structures are more accurately analyzed, and the classification errors of each technique are mitigated when the ensemble members are combined into a final prediction.
However, the individual ensemble members are not combined according to standard voting. Instead, a subset of the training data was withheld, on which the conditional accuracy of all the methods was calculated as a post-processing step. Then, at prediction time, the text is analyzed by each ensemble member. The scores from all methods are considered, and the predicted sentiment is assigned the label that occurred most often in the post-processing data set with the particular combination of predicted scores. Lastly, the sentiment is binned into a five-point scale according to the ratio of occurrences of the ensemble output between positively- and negatively-labeled text in the post-processing data, which allows the detection of neutral sentiment.
This ensemble is implemented as a component of SDL's Customer Commitment Dashboard (CCD), which is part of the suite of Customer Experience Management software solutions offered by SDL. CCD provides unique insight into the end-to-end customer experience by tracking consumer conversations through social media monitoring at all stages of the customer journey.
Mark Gingrich, SDL
- Contextual Sensing and Sentiment Classification
(mouseover here for description)Traditionally, sentiment classification uses models trained on example phrases that are coded for desired sentiment. When constrained to a corpus of well-constrained utterances, such as product reviews on a website, this approach works well. We argue sentiment classification for less constrained corpora can be improved by considering context. Context can simply be the current location of a person, or as complex as knowing the person is at work after a longer than usual day. Context and sentiment classification can be combined in two ways. First, contextual information can improve sentiment classification of text. Information such as where a person was when they created the text could help interpret the content or sentiment behind the text, particularly with content that might be sarcastic or ironic. Second, we can apply sentiment classification techniques to contextual data streams to identify the sentiment of a person at a point in time.For example, knowing that a person had a fairly busy day after not sleeping well could identify that person as tired or grumpy. This can be derived from information such as a wearable sleep sensor and calendar information, activity sensors, or location information, all of which can be derived from sensors on a mobile phone.
Adrienne Andrew, ARO, Inc.
- Analyzing user behavior to add context to sentiments: Community structure mining for user role identification
(mouseover here for description)Analyzing user behavior to add context to sentiments: Community structure mining for user role identification. Description= This talk will focus on how community structure mining of social media to identify and classify user roles based on interaction behaviors can be a useful preprocessing step for sentiment analysis.Understanding the roles users play in community interactions has a two fold benefit.First, once a user is cast into a role their sentiments can be weighted based on the importance of the role.The comments of a power user may be treated differently than those of a routine user.Second, role identification can be used for dataset reduction. Reductions in the data set size can lead to a reduction in processing cost associated with full textual analysis.Data from roles that are identified as not essential for sentiment analysis can be dropped or dimensional reduction can be implemented based on the identified roles.
This presentation will demonstrate the application of structure mining techniques on a social media data set.The methods presented will include social network analysis, graph theory and the development of weighting schemas. The results of an experimental study carried out to empirically evaluate the approach with real- world data is presented highlighting the identified roles.
Robert Nolker, Analyze
- Beyond Sentiment: Context-aware emotion modeling for customer insights
(mouseover here for description)When we reduce a tweet, a post, or any social message to a single positive/negative data point, we do violence to the complexity of our customers' expressions. Using simple sentiment analysis to understand the customer experience is no better than using black-and-white photographs to understand color.
This session introduces methods for modeling the emotional and conceptual resonances in social media data by exploiting the expression-centric structure of corpora like the Experience Project. Through speaker-sensitive word association modeling, we can embrace the complexity of language without compromising the perspicuity of our results.
Moritz Sudhof, Kanjoya