Language as a way of seeing

A conversation in my Intercultural Communication class on Tuesday inspired me to think about language in a different way.

There is a fantastic paper by Charles Goodwin called Professional Vision that shows how people train each other into ways of seeing. The example he uses is an anthropological investigation of soil, and he shows how the new anthropologist is trained by the more experienced anthropologist to notice and describe gradations in the color or quality of the soil. These gradations have a specific meaning within their investigation, and being able to describe them is an important tool of the study.

Language provides two important functions that I want to focus on today. The one we think of most often is the tool function. Language gives us a way to communicate what we see. But it is much less often that we think about language as a way of seeing.

I’ve talked about our constant stimulation before on this blog. We are constantly exposed to more sensory data than we could ever process. Language provides a lens through which to see the world in front of us. Languages have inherent sets of coding structures to apply to the world.

There is another function of language that corpus and computational linguistics really highlights, and that’s habituation. We may have the tools within a language to describe something in any number of ways, but very few of these ways are actually used with any frequency. We learn in corpus linguistics that language is never random. In many ways, in fact, language follows a zipfian distribution, with a high concentration of high frequency words or expressions, and each after the most frequent exponentially less frequent.The language itself may be large, but the language we use is much smaller.

This is an important element to consider in the conceptualization of translation. One person who reviewed the original Russian translation of our Global Survey of Physicists complained that if they did not know English, they would not have understood the Russian. The more I learn about language, the more I understand how that could be possible. Language is not just grammatical, it is habitual, and it holds a way of seeing the world.

I don’t believe this is an idea whose time has come

I ran across this presentation of a Living Social style performance review system:

http://i.bnet.com/video/itunes/12z0315_GeorgeHu_Salesforce_sd_DL.m4v

I’m pretty sure that it is not yet April 1st, so I’m going to assume this program is for real. If it catches on, it represents a giant step toward the gamification of performance reviews.

We are currently in the process of adapting to a new performance review system at my workplace, and the process is immensely complicated. There is definitely some crowd sourcing potential in these systems and room for improvement. But I think that what’s really missing from these systems is a basis for internal motivation. I’m a firm believer that people will naturally try to excel. If people are not trying to excel, it should be taken as a sign that something is disrupting the system. Given that basic assumption, the best kind of performance review system is one that naturally allows, rewards and encourages people to strive to do their best, one that respects and recognizes peoples’ natural talents and interests. Gamification of performance just makes a disrespectful joke out of this. I have to wonder if this is a consequence of an increasingly pervasive artificial system of punishments and rewards. Instead of people finding and experiencing the natural consequences of their actions, are people instead looking for external rewards, such as badges on social media interfaces?

While I truly believe in embracing technology, I don’t believe that this is a step forward. In my ideal work situation, we would all be encouraged to do our best, credited for the work that we do, and given work that we’re most enthusiastic about.

If you think that is naive, please prove me wrong!

Amazing Presentation on Infographics

I had the privilege this week of attending a webinar by Matthew Erickson of the New York Times about innovative graphic presentations of data. There were some truly amazing interactive displays included in this presentation, and the presenter had a lot of very insightful suggestions for rethinking data presentation:

 

http://www.ericson.net/files/aapor-shared.pdf

 

He spoke about the role of good interactive presentation in situating data, providing context, developing layers, and telling a story. A lot of times, the distribution of the data, and it’s relationship with data from other sources, is its most interesting layer. In an innovative presentation of data, we must balance the expectations of the audience, who become interactants with the data and must be able to manipulate it easily, with a complementary layer of expertise or context.

For example, data about Manny Rivera’s pitching style could best be understood by the placement of the ball at the hitter’s decision making point. In a graphic about Rivera’s success, the reporters were able to show how radically different pitches were virtually indistinguishable at the crucial decision making point for the hitter.

He referred to infographics as the “gamification of news.”

 

To connect his presentation to this ongoing discussion of text analytics, check out the way he displayed word frequencies:

http://www.ericson.net/presentations/aiga-for-web.pdf

Interestingly, it is still problematic, but it is super cool looking…

 

And, speaking of infographics, check out this awesome one that Pew debuted today:

http://features.pewforum.org/religious-migration/map.php#/Destination/UnitedStates/

Calling Respondents Stupid

From a Politico article titled “How much do voters know?” by Alexander
Burns:

“The first lesson you learn as a pollster is that people are stupid,”
said Tom Jensen of Public Policy Polling, a Democratic polling firm. “I
tell a client trying to make sense of numbers on a poll that are
inherently contradictory that at least once a week.”

Full article at:

http://www.politico.com/news/stories/0312/73947.html

Message to Tom Jenson:
It’s not so much that people are stupid as it is that Tom Jenson is ignorant of the cognitive underpinnings of response methodology and filling in his gaps with unnecessary condescension. Truly, all survey researchers deal with these contradictions. We might be able to garner more consistency in our response sets if we could instead survey computers or robots, but, alas, that’s not why we conduct surveys.

Qualitative and Quantitative methods revisited

At the GURT 2012 conference last week, another graduate student whose work was primarily quantitative asked me about the issue of data quality in qualitative research. She asked how, if your results are not repeated over a large group of people, do you know that your results are reliable or representative? In the process of answering her, I realized that in some ways quantitative and qualitative research are ideologically opposed. Whereas quantitative research is validated in part by the reliability of a result across a variety of people and contexts, the people and contexts for which the result is being quantified are not the focus but the rationale. However, in qualitative research, the context and the people are the focus. Qualitative research is more about putting a microscope up to an element in the data and closely observing how it works in the context of its surroundings. The research questions are starkly different, but they do clearly complement each other.

In the world of text analytics, the analytic focus is largely quantitative, which in some ways counteracts the very nature of the data. The unbalanced nature of the data collection then calls the quality of the quantitative analysis into question, and the analysis is not used in a traditional way to recreate the microfocus, so the advantages of the qualitative microfocus are also diminished.

We spoke a lot at GURT about the paucity of theory in the field of text analytics. The technical ability applied to the problems is quite strong. There are quite a few very talented programmers who are hard at work at conquering some of the technical issues inherent in text analysis, some on the computer science end of things, some at the computational linguistic end of things and some fortunate enough to work from both ends with knowledge from both fields. But it is not enough to be able to solve problems and answer questions, we have to know which questions to ask.

It has been relatively easy to blame the fast growing set of consumers for their immediate hunger for the data analysis. In order to satisfy this hunger, companies like Open Amplify double their output monthly and are working as hard and fast as they can to keep up with the demand. But the demand generally comes with the same level of linguistic knowledge that most laypeople have. We, as language users, are constantly inundated with language, and we only consciously process a very small proportion of it. So we don’t instinctively ask the questions that our data is really best suited to answer. The text analytic world is responding to questions of “what are people talking about?” with word frequencies, comparative word frequencies and sentiment analyses that are tied to those word frequencies. But we don’t use language that way! If I ask you about your phone, you’re likely to respond about its features of its usefulness of its price, or how well you’ve adapted to it. If I ask 100 people about their phones, how much good will it do me to aggregate across responses? There is a good deal of work that needs to be done in terms of finding intertextual references to phones (e.g. “springy keypad” or “data plan”) and assigning a negative value to “limited calling plan” and a positive to “limited call interference.” When I asked a coworker how our advisory committee meeting was going while I was at the conference, she answered “delicious.” We communicate by keying on shared knowledge, and as we communicate we build senses of particular topics that are related specifically to our conversation. If my coworker had answered with a comment about the potato salad, and I had played off of that, would we be talking about potato salad in any equivalent way to the way we might at a summer picnic? I would argue that as we joke on, we be talking in fact about something quite different than the potato salad itself. In fact, we would likely use the potato salad as a stand-in for the meeting that we were really discussing. Should that conversation be used by market researchers in a potato salad corporation?

In fact, the topics that we discuss are quite variable. The specific meanings that elements take on within a conversation is best understood in the context of that conversation as a part of a qualitative analysis than an aggregated quantitative analysis.

The big essential questions that we need to grapple with, as a field, at this point are questions like:

‘what kind of questions is this kind of data best suited to answer?’

‘how can our knowledge of linguistics and discourse be transferred into quantifiable questions that could feed the field of text analysis’

‘what kinds of questions can we ask of textual data that will reframe the way that people think about the usefulness of textual data?’

‘how can we best harness this fast growing mass of textual data in the most useful, reliable ways?’

I would argue that these are questions that discourse analysts are best suited to answer, but in order to ask them, they/we must be able to leave our qualitative bunkers and open our minds to the complementary potential for quantitative analysis. I would also argue that a popularized appreciation for the value of discourse analysis would also lend some legitimacy to a field that is largely unknown.

On the way to work this morning, I listened to an interview with Naomi Wolf. She spoke in part of the chutzpah of presenting academic knowledge in a widely accessible format. Academic perspective, she argued, is too often maintained in academic circles, far away from the general population who could really use and appreciate it. Georgetown professor Deborah Tannen made some important steps in the popularization of sociolinguistics. I believe that what I am suggesting is a quantitative extension of the popularization. People could not have imagined that a book about something as obscure as conversation analysis could be interesting or so widely applicable to their own lives. There was no rushing the doors from the general population of people desperate for these books. Hers was not a case of giving people what they wanted. It was a case of giving people something that would be widely useful. And people embraced Dr Tannen’s books as such.

Let us now use the luxury of time that the academic sector has but the commercial sector certainly does not to do what we do best: theorize! A new, great plane awaits. Let us head west!

Keeping Research Responsible

Reliability and validity are the two most important principals in social science research. They are the measures that maintain the integrity, quality and ultimate usefulness of our work.

Reliability refers to the replicability of findings, which is a crucial element in any scientific process. Repeatability under varying conditions helps to establish the consistency and boundaries of a phenomena. Oftentimes, we like to compare ourselves to scientists and emphasize our take on the scientific method. I would argue that some of the most important lessons we can learn from science include the basic doubt (inherent in statistics as well, thanks to the null hypothesis) that assures that no one study can make knowledge so much as suggest knowledge that can lead to further testing and potential verification. Research is inherently tied to its underlying questions, and a well conducted study based on one question could easily lead to very different findings than another, even slightly different, research question. Reliability cannot be valued highly enough in social science research.

Validity refers to the value of our findings. What do these findings mean? What do they refer to? In survey research, understanding the validity of a finding often entails retracing what question was asked to whom under what circumstances and anchoring any conclusions to those basic truths instead of extrapolating to a wider principle that we would like to have observed.

In text analytics, their are two more anchors that guide research; precision and recall. Recall is the measure of how many of the correct instances of a phenomena you were able to isolate in your programming, and precision refers to the percent of the instances that you collected that were indeed instances of your target phenomena. Text analysis is a dance of queries, toggling between collecting correct matches and dropping incorrect matches. It is within the context of this dance that language seems most staggering. Here we see how little people say what they mean or mean what they say, how much context matters, how often people refer to subjects by proxy, how dependent on ongoing conversation new elements are, …

As users of language, we are constantly inundated with words. We cope with this by only focusing on selective elements. We focus on metamessages, not mechanics. It is easy to assume, from our perspective, that language is straightforward. Indeed, if it was straightforward, text analytics would be easy and misunderstandings would not be so rampant! All of the linguistic data that we are flooded with would be harnessed regularly, and society would reflect that data better through ubiquitously increased customization.

But the reality of language stands in stark contrast to what we assume it to be, and that reality is the lifeblood of linguistics.

Rethinking the Future of Survey Methodology; Finding a Place for Linguistics

Where is the future of survey research?

The technical context in which survey methodology lives is evolving quickly. Where will surveys fit into this context in the future?

In the past, surveys were a valuable and unique source of data. As society became more focused on customization and understanding certain populations, surveys became in invaluable tool for data collection. But at this point, we are inundated with data. The amount of content generatd every minute on the net is staggering. In an environment where content is so omnipresent, what role can surveys play? How can we justify our particular brand of data?

Survey methodology has become structured around a set of ethics and practices, including representativeness and respect for the respondents. Without that structure, the most vocal win out, and the resulting picture is not representative.

I recently had the pleasure of reading a bit of Don Dillman’s rewrite of ‘The Tailored Design Method,’ which is the defining classic reference in survey research. The book includes research based strategies for designing and targeting a survey population with the highest possible degree of success. It is often referred to as a bible of sorts to survey practitioners. This time around, I began to think about why the suggestions in the book are so successful. I believe that the success of Dillman’s suggestions has to do with his working title- it is a tailored method, designed around the respondents. And, indeed, the book borrows some principles of respondent or user centered design.

So where does text analysis fit into that? In a context where content is increasingly targeted, and people expect content to be increasingly targeted, surveys as well need to be targeted and tailored for respondents. In an era where the cost benefit equation of survey response is increasingly weighted against us, where potential responses are inundated with web content and web surveys, any misstep can be enough to drive respondents away and even to cause a potential viral backlash. It has never been more important to get it right.

And yet we are pressured not to get it right but to get it fast. So the traditional methods of focus groups and cognitive interviews are increasingly too costly and too timely to use. But their role is an important one. They act to add a layer of quality control to the surveys we produce. They keep us from thinking that because we are the survey experts we are also the respondent experts and the response experts.

A good example of this is Shaeffer’s key idea of native terms. I have a brief story to illustrate this. Our building daycare is about to close, and I have been involved in many discussions about the impact of its closure as well as the planning and musing about the upcoming final farewell reunion celebration. The other day I ran into one of the kids’ grandparents, someone who I have frequentky discussed the daycare with. She asked me if I was planning to go to Audrey’s party. I told her I didn’t know anything about it and wasn’t planning to go. I said this, because I associate the terms she used with retirement celebrations. I assumed that she was talking about a party specifically in honor of the director, not the reunion for all of the kids.

It’s easy as a survey developer to assume that if you ask something that is near enough to what you want to know, the respondent can extrapolate the rest. But that belies the actual way in which we communicate. When it comes to communication, we are inundated with verbal information, and we only really consciously take the gloss of it. That’s what linguistics is all about; unpacking the aspects of communication that communicators don’t naturally focus on, don’t notice, and even can’t notice in the process of communication.

So where am I going with all of this?

One of the most frequent aspects of text analysis is a word frequency count. This is often used as a psuedo content analysis, but that is a very problematic extrapolation, for reasons that I’ve mentioned before in this blog and in my paper on ths topic. However, word frequency counts are a good way of extrapolating native terms from which to do targeting.

Text analytics aren’t representative, but they have the ability of being more representative than many of the other predevelopment methods that we employ. Their best use may not be so much as a supplement to our data analysis as a precursor to our data collection.

However, that data has more uses than this.

It CAN be used as a supplement to data analysis as well, but not by going broad. By going DEEP. Taking segments and applying discourse analytic methodology can be a way of supplementing the numbers and figures collected with surveys with a deeper understanding of the dynamics of the respondent population.

Using this perspective, linguistics has a role both in the development of tailored questionnaires and in the in depth analysis of the respondenses and respondents.

More work on Twitter, Google Searches and Text Analytics in Survey research

I am so excited to see this blog post and read the paper that it was based on!

 

blog post:

https://blogs.rti.org/surveypost/2012/01/04/can-surveillance-of-tweets-and-google-searches-substitute-survey-research-2

paper:

http://www.rti.org/pubs/twitter_google_search_surveillance.pdf

 

Kudos to RTI for continuing to carve out a place for text analytics in the future of survey research!

Word Clouds

Here is an interesting application of word clouds. It is a word cloud analysis of Public Opinion Quarterly, the leading journal in Public Opinion Research:

https://blogs.rti.org/surveypost/2012/02/26/a-visual-history-of-poq-1937-to-present/

Word clouds are a fast and easy tool that produce a visual picture of the most frequently used words in a body of text or ‘bag of words.’ They are frequently used as a tool for content analysis.

On my ‘my research’ page above, there is a link to a paper I wrote about text analytic strategies. In the paper, I addressed word clouds in great detail. I did that because word clouds are fast gaining popularity and recognition in the survey research community and in the wider society at large. However, the clouds have a lot of limitations that are rarely considered by people who use them.

One of the complications of a word cloud is that word frequency alone doesn’t speak to the particular ways in which a word was used. So when you see ‘public,’ you think of the public/private dichotomy that is such a big debate in the current public sphere. However, in the context of a survey ‘public’ could also easily be used as a noun, to refer to potential respondents. While word clouds appear to give a lot of information in a quick visual, the picture underlying that information can be clouded by the complexities of language use.

I don’t think that these pictures can map directly onto the underling topical landscape, but they can provide a quick window into the specific words that we have used over the years and the changes in our lexicon over time.

Another CLIP

I missed today’s CLIP. Too much work and too much rain. But the description of it made it sound especially interesting, because the speaker is obviously really grappling with the concept of context. It would have been interesting to have heard what he did with it and how he used linguistics (he specifically mentioned the field, albeit probably not in a discourse analytic type of way). I will have to follow up with him or with his papers. Thankfully, he’s local!

Here’s the sum:

February 29: Vlad Eidelman, Unsupervised Textual Analysis with Rich Features

Learning how to properly partition a set of documents into categories in an unsupervised manner is quite challenging, since documents are inherently multidimensional, and a given set of documents can be correctly partitioned along a number of dimensions, depending on the criterion. Since the partition criterion for a supervised model is encoded in the data via the class labels, even the standard information retrieval representation of a document as a vector of term frequencies is sufficient for many state-of-the-art classification models. This representation is especially well suited for the most common application: topic (or thematic) analysis, where term presence is highly indicative of class. Furthermore, for tasks where term presence may not be adequate, such as sentiment or perspective analysis, discriminative models have the ability to incorporate complex features, allowing them to generalize and adapt to the specific domain. In the case where we do not have access to resources for supervised training, we must turn to unsupervised clustering models. Clustering models rely almost exclusively on a simple bag-of-words vector representation, which performs well for topic analysis, but unfortunately, is not guaranteed to perform well for a different task.

In this talk, I will present a feature-enhanced unsupervised model for categorizing textual data. The presented model allows for the integration of arbitrary features of the observations within a document. While in generative models the observed context is usually a single unigram, or bigram, our model can robustly expand the context to extract features from a block of text of larger size. After presenting the model derivation, I will describe the use of complex automatically derived linguistic and statistical features across three practical tasks with different criterion: perspective, sentiment, and topic analysis. I show that by introducing domain relevant features, we can guide the model towards the task-specific partition we want to learn. For each task, our feature enhanced model outperforms strong baselines and state-of-the-art models.

Bio: Vladimir Eidelman is a fourth-year Ph.D. student in the Department of Computer Science at the University of Maryland, working primarily with Philip Resnik. He received his B.S. in Computer Science and Philosophy from Columbia University in 2008 and a M.S in Computer Science from UMD in 2010. His research interests are in machine learning and natural language processing problems, such as machine translation, structured prediction, and unsupervised learning. He is the recipient of the National Science Foundation Graduate Research and National Defense Science and Engineering Graduate Fellowships.