Author: Janna Lipenkova
Market research surveys typically consist of two types of questions, namely “closed” and “open” questions. Closed questions limit the possible range of responses and result in structured data. By contrast, “open” questions allow the respondent to reply with free text, expressing their full-fledged, authentic opinion. This flexibility makes open questions attractive in market research: given the right wording, an open question can trigger responses that are broader, deeper and provide more authentic insight than the rigid, black-and-white multiple-choice question.
Challenges with the analysis of open-ended questions
Why, then, are open-ended questions not widely adopted in market research? One of the reasons is that they are difficult to analyze due to their unstructured character. Most researchers use manual coding, where open questions are manually structured into a classification scheme (the “coding frame”) according to the topics they represent. The following table shows some examples:
Manual coding comes with several issues:
- High cost: manual coding is labor-intensive and thus expensive.
- Errors: depending on their mental and physical shape, human coders make mistakes or provide inconsistent judgments at different points in time.
- Subjectivity: due to the inherent ambiguity and the subtleties involved in human language, different people might code the same response in different ways.
Last but not least, coding hundreds or even thousands of open-end questions can be a frustrating endeavor. In today’s world, where AI is used to automate and optimize virtually every repetitive task, it seems natural to turn to automated processing to eliminate the monotone parts of coding. Beyond the several benefits of automation, this also creates time for more involved challenges that require the creativity, experience and intellectual versatility of the human brain.
Using Natural Language Processing as a solution
Natural Language Processing (NLP) automates the manual work researchers do when they code open-ended questions. It structures a text according to the discussed topics and concepts as well as other relevant metrics, such as the frequency, relevance and sentiment. Beyond speeding up the coding process, NLP can be used to discover additional insights in the data and enrich the end result. The capacity of a machine to look at a large dataset as a whole and discover associations, regularities and outliers is larger than that of the human brain.
Three algorithms – topic modeling and classification, concept extraction and sentiment analysis – are particularly useful in the coding process.
Topic modeling and classification
Topic modeling detects abstract topics in the text. Topic modeling is an unsupervised learning method similar to clustering, and learns lexical similarities between texts without a predefined set of classes. Thus, it is particularly useful in the initial stage of the construction of a coding frame. The following word cloud shows words that are frequently mentioned in texts about comfort:
Topic classification is similar to topic modeling. However, it works with a given coding frame and classifies each text into one of the predefined classes. This means, that it can be used for coding after the coding frame has been constructed.
Concept extraction matches concrete strings in the text. Whereas topic modeling and classification work with – often implicit – lexical information distributed everywhere in the text, concept extraction matches the exact words and phrases that occur in the text. On a more advanced level, concept extraction also uses the structure of the lexicon and can deal with lexical relations, such as:
- Synonymy: EQUALS-relationship, e. g. VW EQUALS Volkswagen
- Hypernymy: IS-A-relationship, e. g. Sedan IS-A Vehicle
- Meronymy: PART-OF relationship, e. g. Engine PART-OF Car
Concept extraction usually focuses on nouns and noun phrases (engine, Volkswagen). In the context of evaluations (open-ended questions), it is also useful to extract concepts that are “hidden” in adjectives (fast ➤ Speed, cozy ➤ Comfort) and verbs (overpay ➤ Price, fail ➤ Reliability).
In terms of implementation, there are two main approaches to concept extraction: the dictionary-based approach and the machine-learning approach. The dictionary-based approach works with predefined lists of terms for each category (also called “gazeteers”). The machine-learning approach, on the other hand, learns concepts of specific types from large quantities of annotated data. As a rule of thumb, the smaller and more specific the available dataset, the more efficient the use of pre-defined lists of concepts and linguistic expressions.
Sentiment analysis detects whether a given text has a positive or a negative connotation. Sentiment analysis can be further detailed to the level of individual aspects mentioned in a text, thus allowing to detect mixed sentiments on the phrase level:
“Classy and reliable, but expensive.”
Sentiment analysis operates on an emotional, subjective and often implicit linguistic level. This subtlety raises several challenges for automation. For example, sentiment analysis is highly context dependent: a vacuum cleaner that sucks would probably get a neutral-to-positive sentiment; by contrast, the internet connection in a car normally shouldn’t “suck”. Another complication is irony and sarcasm: on the lexical level, ironic statements often use vocabulary with a clear polarity orientation. However, when put into the surrounding context, this polarity is inversed:
“Really great engineering… the engine broke after only three weeks!”
Irony is mostly detected from anomalies in the polarity contrasts between neighboring text segments. For instance, in the example above, “really great engineering” gets a strong positive sentiment which radically clashes with the negative feel of “the engine broke after only three weeks”. Since the two phrases are directly juxtaposed without a conjunction such as “but” or “although”, the machine is able to recognize the overall negative connotation.
Combining Human and Artificial Intelligence
Summing up, using NLP for the coding of open-ended questions leverages several benefits of automation: it speeds up the process and saves human labor on the “frustrating” side of things. It achieves better consistency and objectivity, mitigating the effects of human factors such as fatigue and inconsistent judgment. Finally, the ability of the machine to process large data quantities at a high level of detail allows a level of granularity that might be inaccessible to the human brain.
While it is out of question that NLP automation increases the efficiency of verbatim coding, keep in mind that current AI technology is not perfect and should always have a human in the driving seat. Methods such as NLP can process large quantities of data in no time, but they do not yet capture the full complexity of language. A combination of high-quality NLP with a carefully engineered process for continuous optimization will ensure a rewarding journey towards in-depth understanding of the opinions, needs and wants of end consumers.