Fabon Dzogang is the author of 'Early Fusion of Low Level Features for Emotion Mining', recently published in the Biomedical Informatics Insights Supplement.  We asked Mr Dzogang to tell us about the background of this paper.

To start please tell us about the challenge this year.  Why did you decide to become involved, and what goals did you and/or your team expect to accomplish?

I2B2 challenge (track2) consists in classifying emotions from suicide notes. Participants are provided with sentences labelled from a set of 15 emotions (e.g. “love”, “anger”, “joy” …) and the goal is to predict the relevant emotions expressed in unknown sentences.

Opinion/Emotion mining from texts is a new area of research. The very ambiguous and idiosyncratic nature of emotions as well as their subtle expressions in texts makes it a difficult learning task. As part of my PhD thesis, the I2B2 challenge is an opportunity to study a consistent corpus and to compare our work with other systems as well as to learn from other participant’s experience.

In writing this paper what were the particular challenges you faced?  How did you overcome these challenges?

This task involved many challenges:

Firstly, emotion labels are highly unbalanced in the corpus. Even though our system employs asymmetric classification costs penalizing the most prevalent emotion labels, rare emotions are highly underrepresented in the corpus (every systems could not learn much for the most rare emotions).

Secondly, sentences are multi-labelled, but when implementing our system, we chose to ignore this facet as we observed that only 7% of the sentences are associated with more than one emotion. The task is a hard one and we wanted to focus our whole attention on the detection of emotions. We learned from other participants that in a one versus all strategy, keeping every positive prediction proves efficient for this corpus.

Finally, emotion mining in texts is known to be a difficult task, subjectivity of emotions, complex linguistic constructs, a lack of descriptive features makes it hard to learn a good system.

What has been the major benefit for you in the work discussed in your article?  How has it contributed to our knowledge of the area?  

We implemented a 2-step method consisting of detecting neutral sentences first, then discriminating among the 15 emotions labels. Our method is based on the fusion of n-grams at the vector level. We also compute dictionaries specialized for each emotion by filtering n-grams according to Shannon's entropy measures.

Making use of low level features, our work shows the benefits and limitations of methods purely based on data. The fusion strategy we employ allows coarse to fine representation of descriptive features, for instance considering the unigram “bad”, the bigram “not bad” models negation while the trigram “not really bad” takes account of linguistic intensifiers. While our system performs well, emotions for which no discriminative vocabulary exists in the corpus are not handled well which was particularly true for emotions occurring rarely in the corpus. Our work shows that some emotions do possess a naturally discriminative vocabulary: even though “Love” and “Thankfulness” are not the most frequent emotions in the corpus, they are the easiest to detect.

Other participants who made use of external and semantic resources on emotions could alleviate the problem we discussed above, also known as the semantic gap. One perspective of our work is to hybrid low level features with semantic features.

View Fabon Dzogang's LinkedIn page and personal website

share on