Words Of Persuasion: Best Adjectives For Persuasive Ted Talks

Data Collection

The dataset is retrieved from Data World database by OwenTemple’s Project, containing a complete listing of all TED Talks from official events posted in TED.com. The dataset includes the URL where video can be viewed, full English transcript URL, speaker name, headline, description, month and year filmed, event, duration, date published, topic tags and extended with abstract variables that describe the ratings of the talks, which is ‘persuasive’, ‘inspiring’ and ‘unconvincing’. The rating variables are gathered by TED.com from their users who had voted for a particular talk using TED.com’s ratings tool.

Data Pre-Processing and Data Cleaning

As the gathered dataset consist of too many irrelevant attributes, the first step of this section is to redefine a new dataframe with only relevant attributes: Transcript, Tag, Persuasive and ViewCounts. A complete data cleaning approaches (Fill NaN values, fix attribute types) is performed to make sure the data is cleaned.

Considering that the number of views over time since the talk posted might create a misleading data in ‘Persuasive’ attribute, a normalized attribute should be derived to better representing the persuasiveness of each talks. To normalize the rating to account for the number of views for each session, a new variable ‘norm_persuasive’ is extracted through dividing the count of persuasive votes by the number of times the talks had been viewed, which can be also defined as Persuasive votes per View of Talk (Figure 2 ).

See Full Size >

Meanwhile, each transcript is originally tagged with multiple topic category that best describe what the talks about. However, each title has multiple tags that used to describe the talks. To better segregate the speech type according to each tag allocated to each talk, each list of tags is first split into multiple rows attached with only single tag.

As the main objective of current work is to extract on the frequent words used in both persuasive and less persuasive talks, the instances are separated into persuasive (Norm_Persuasive >= 250) and less persuasive (Norm_Persuasive > 250) and categorized into top 4 tag category (Technology, Global Issue, Science and Business). The final dataset used is reduced to only the English transcript, tag, Normalized Persuasive Rating attributes.

Text Processing

After the cleaning process, each instance is tokenized using NLTK tokenizer (Bird, Klein, & Loper, 2009) into unigram bag of words. The reason of only applying unigram to the transcript is because we believed that a persuasive transcript might contain some ‘Convincing Adjectives’ that might better bring resonance across the listeners to the speaker. In such a way, unigram could be better in identifying the ‘Adjectives’ in POS tagging section at later.

During the tokenization of the transcript, a stop word filter is first applied to remove the commonly used word that does not contributes to later process. Each word is applied to retain only lower case before going into lemmatization process to further reduce the inflectional forms and derivationally related forms of a word to a common base form.

After applying lemmatization to the tokens, each word is classified into their parts of speech and labelling them accordingly to their lexical category using NLTK part-of-speech tagger. The motivations behind such steps is to identify the adjectives and retain only them for later uses. The tokens extracted to be only consisted of tags: ‘JJ – Adjective’, ‘JJR – Comparative Adjective’ and ‘JJS – Superlative Adjective’. The frequency of each words is calculated using NLTK FreqDist to further validate the processed bag of words. Afterall, the same techniques and approaches are applied to both persuasive talks and less persuasive talks and ready to be further extraction of insights from the context.

Word Cloud and Frequency Histogram

After each word are processed and engineered to be fit into the wordcloud analysis, the cleaned data is loaded into Word Cloud and plotted with the frequency distribution histogram. The key steps in current section is to prepare a better and presentable visualization of data from the cleaned dataset, in another word is to prepare the bag of words illustrated informatic wordcloud and bar plot. Each set of instances (persuasive and less persuasive) is implemented with same methods to finally output their Word Cloud board and Frequent Word histogram. The most frequent words that contributes to a persuasive speech is then identified.

Evaluation Criteria

To further test and validate the suitability of the Word Cloud and Frequency Histogram, a set of formal evaluation and informal evaluation approach is conducted to evaluate the current model.

First session is to validate again the output of model (WordCloud and Frequency Histogram) through internal testing (Alpha test) on the retrieved bag of words. The challenge here is to recap the actual suitable words that appeared in the top 50 adjectives and fine tune the word tokens structure. The initial output is tending to have a lot of irrelevant nouns and adverb that particularly not making any sense. A process of choosing different part of speech tag in Penn Treebank tag set is iterated until the best word cloud that contains only adjective is generated.

Second session is to gather feedback from public speaking enthusiast and regular listeners using google doc survey form. The involved participants are explained with the motivations behind the project and required to comment and rate for the suitability of the WordCloud generated. Each participant is showed with different WordCloud according to different tag category (Technology, Global Issue, Science and Business) and let them rates on how helpful the output could be. The results show positive feedback experience as most participants give positive ratings to the output. Figure 03 shows the overview of expert feedback’s results depicted in pie chart.

See Full Size >

Technology Talk

The generated Word Cloud of Persuasive Technology is surprisingly containing more information than expected. As depicted in Figure 4 , the top words used by the persuasive transcript is ‘Antiangiogenic’, ‘Intergalactic’, ‘Cooperative’, ‘Testable’, ‘Recyclable’. The relevant topics that might be related to the words extracted is impressed, where adjective Anti angiogenic is used to describe the drugs that used in treatment to stop tumours from growing own blood vessels, Intergalactic is used to describe the travel between galaxies in science fiction and speculation, Recyclable is used to describe the waste that could be reuse or recycle. It is clearly describe that people are more convincing if the technology related stuff is related to these keywords’ main purpose. While words like ‘headtailhead’, ‘readwrite’ could be categorized as irrelevant words as they do not provide much information.

For Less Persuasive transcript (Figure 5 ), there appeared to have some commonly used adjectives, such as Litte, many, first, different, much, good, great, able, big and so on. The insights here showcase that these adjectives are tended to be less persuasive when used in a technology talk.

See Full Size >

For Less Persuasive transcript (Figure 5 ), there appeared to have some commonly used adjectives, such as Litte, many, first, different, much, good, great, able, big and so on. The insights here showcase that these adjectives are tended to be less persuasive when used in a technology talk.

See Full Size >

Global Issues Talk

Meanwhile, the global issue talks transcript contained lesser information than expected and only certain relevant keywords is managed to be extracted. However, it is still visible that if adjective that related to ethical behaviour is being used, it has better chance to persuade the listener, for example Moral, Rational, Serious, Nonviolent and Liberal. Word Cloud of global issue talks transcript containing more irrelevant words (such as adverb) compared to technology talks, it is assumed that global issues topic might require more adverb style of adjective in supporting their emotional persuading techniques (Figure 6 ). While the less persuasive speech again showcases the commonly used adjectives just like technology, which can be assumed that the speech will be less persuasive if such keywords is being used during the talk (Figure 7 ).

See Full Size >

See Full Size >

Science Talk

For the science talk, the retrieved top adjective used in persuasive talk is surprisingly have some common words compared with Technology talk. Some adjective like Intergalactic, Antiangiogenic is again raised in the transcript, which again proof that these keywords could actually help in persuading their listeners and could be comprehended that these issues is more concerned by the listener. Key adjective that could raise resonance is Financial, Agnostic and Irrational (Figure 8 ) while the less persuasive transcript is again containing the commonly used adjectives (Figure 9 ).

See Full Size >

See Full Size >

Business Talk

For business talks, it is surprisingly that the keywords extracted is more descriptive than being expected. The worth pinpoint extracted top words in persuasive transcript is varied across different categories, such as Moral, Copyright, Nuclear, Intangible, Tangible, Risky, Civil, Intrinsic and many more (Figure 10 ). Across these words, we could make assumption on a persuasive talk is tended to be rich in descriptive adjectives and more intense adjective. The less persuasive talks again contained the commonly used adjective that is less intense and does not makes them be convincing enough to persuade the listener (Figure 11 ).

See Full Size >

See Full Size >

Discussions

From the work conducted, multiple text processing processes are performed to final generate a suitable model to better visualize on the frequent word used in persuasive speech and less persuasive speech. In the design stage, each process is compared with several similar techniques. Only suitable technique to was chosen in the design of the framework. In this section, we provide the discussions of each process and justify the selection of the selected techniques.

One challenge revealed during stop word filtering is the removal of punctuation of each word tokens. Although a list of stop words is used to filter the token list, but there still exist unwanted punctuation such as ‘100-year-old’, ‘low-power’ and more. Such a case is less efficient when plug into Word Cloud, where the word cloud module will separate the tokens into ‘100’, ‘year’, ‘old’ and provide an inaccurate output. To further fix this, the punctuation is manually removed with panda library function is left only word without ‘100yearold’, so each token is correctly used in word counting in Word Cloud module.

Besides, the tokens that will be filtered to contain only adjective might have different type adjective, such as superlative adjective and comparative adjective. To further reduce the inflectional forms and derivationally related forms of a word to a common base form, Lemmatization is applied. From there, the bag of words is then tagged with part of speech category and filtered to leave only the adjective for evaluation.

Lastly, as there appeared to have overlapped words set appeared in both persuasive and less persuasive transcript, we discovered that there is a need to segregate only the unique words that makes that particular speech to be persuasive. Considering the fact that some commonly used adjective would appear in both persuasive and less persuasive transcript, we compare both set bag of words and remove the duplicated word set, leaving only the unique word tokens to be our final data model for output generation.