Tags

, , , ,

Sentiment is intellectualized emotion; emotion precipitated, as it were, in pretty crystals by the fancy. – James Russel Lowell

What other people think about a particular topic and their opinions regarding the same have always been important to us. Sentiment analysis, which is also called opinion mining, aims to analyse the sentiment of a given document or a given topic by employing methods such as computational linguistics, natural language processing and text mining. Basic sentiment analysis tries to classify the polarity of a given text i.e whether it is positive, negative or neutral. A more advanced analysis tries to capture the author’s emotional state – whether she was “happy“, “angry“, “sad” etc.

Despite its great potential, application of sentiment analysis is emerging only slowly in the field of digital humanities. The current focus is on applying sentiment analysis mainly in areas of social media, reviews, news etc.

In their paper, Marchetti et.al. talk about applying sentiment analysis methods to historical texts. Apart from identifying sentiments of single documents, they propose that it can be applied to a large collection of documents spread over a period of time to understand how the attitude of people towards a specific topic changed over time. Such an understanding would be very helpful for those involved in research on say, history of ideology or evolution of political thought.

The data set they used  consists of about 3000 documents by De Gasperi, an Italian statesman and politician. They conducted two experiments of the data set the first of which used only prior polarity ie, polarity of a word without any context. They extracted 5,224 lemmas on which the method can be applied. Among those, there were 449 lemmas with a high positive polarity (words like ‘giubilo’=rejoicing), 576 with a high negative polarity(words like ‘affranto’=broken-hearted) and the remaining lemmas had an intermediate polarity. The general sentiment of each document was obtained by averaging the (prior) polarity of those lemmas in the document which have a known polarity. The results obtained were accepted by most historians showing that their estimation of polarity was good. The following is a video of a similar, but more general text mining, experiment on the same data set

 But these documents contain several topics and historians are interested in understanding how the sentiment varied for different topics with time. In their second experiment, they got several sentences (each along with its context) regarding a particular topic evaluated for their polarity by two expert annotators and compared results with those obtained from crowd sourcing contribution and through prior polarity. The results are summarized in the following figures

DH2014_183_fleiss

Reliability of agreement between experts and crowd

DH2014_183_accuracy

Accuracy with respect to polarity evaluated by experts

These results indicate the complexity of identifying the basic polarity and shows scope for improvement for contextual polarity.

Another emerging area of research is Digital Happiness. Similar to applying  basic Sentiment Analysis on historic texts, a slightly more advanced analysis can be applied on social media to know the well-being of people. In her paper, Jill Belli talks about Digital Happiness, why it is important and how it is still a utopian project. The following video shows some existing work on this topic and the application is provided in references.

In her paper, she summarizes different methods used to assess as well maximize well-being. These methods mainly include big data, sentiment analysis, crowdsourcing, social networking, biometrics etc. She also analyzes the aims and methods of such projects and critically analyze their utopian aspirations.

Sentiment analysis can be done not only to measure happiness but hatred as well. In their paper, Quinn et.al. presents analysis methods that can detect linguistic patterns that initiate and sometimes result in compounding hate speech in online environments. Unlike the method in the first paper which was trying to evaluate prior polarity, ie sentiment without any context, the method in this paper gives more importance to the context. They have built dictionary of offensive words along with catalogues of clausal structure where these words are used. The data set they used for their research comprises of texts derived from comments on YouTube videos where some drug users are subjected to dehumanizing language.

They use Sentiment Analysis along with Systemic Functional Linguistics methods to visualize patterns of hate speech as the text develops and as these patterns assimilate enough power to cause further hate speech. These visualizations help to detect the less explicit hate speech which could otherwise be undetected. In these visualizations, they use lexical cohesion and clausal structures of sentences to detect patterns corresponding to hate speech. The results of one of their experiments is summarized below.

 DH2014_278_Slide16DH2014_278_Slide18

Every node represents a cluster of chat conversations except  the node at the centre which represents the video. Among these, the purple nodes represent anaphoric conversations which relate to the video directly and blue nodes represent  exophoric conversations which are not directly related to the video. The lines joining the nodes are the result of their sentiment analysis tool. The red lines denote negative sentiment and the green ones (very thin in the picture) positive. The thicker the lines are, the stronger is the sentiment. The result shows strong negative sentiments between nodes of conversations.

All the above are examples of applying Sentiment Analysis to Digital Humanities. It can be seen that it is an emerging trend and will soon be popular.

References

  1. Marchetti, A. , Sprugnoli, R. , and  Tonelli, S. (2014). Sentiment  Analysis For The Humanities: The Case Of Historical Texts
  2. Belli, J. (2014). Unhappy? There’s An App For That: Digital Happiness, Data Mining, And Networks Of Well-Being
  3. Quinn, D. Maycock, K. and Keating, J. (2014). Fractures And Cohesion: Using Systemic Functional Linguistics To Detect And Analyse Hate Speech In An Online Environment
  4. Pang, B. and Lee, L. (2008). Opinion mining and sentiment analysis, Foundations and Trends in Information Retrieval 2 (1-2) , 1-135.
  5. http://www.uvm.edu/storylab/
  6. http://hedonometer.org/
  7. dh.fbk.eu/projects/alcide-analysis-language-and-content-digital-environment
Advertisements