Call for Participation in Linked Data Models for Emotion and Sentiment Analysis Community Group
Posted on:The Linked Data Models for Emotion and Sentiment Analysis Community Group has been launched:
The Sentiment Analysis Community Group is a forum to promote sentiment analysis research. Topics addressed are:
- Definition of a Linked Data based vocabulary for emotion and sentiment analysis.
- Requirements beyond text-based analysis, i.e. emotion/sentiment analysis from images, video, social network analysis, etc.
- Clarifying requirements and the need for consensus as e.g. systems currently use widely varying features for describing polarity values (1-5, -2/-1/0/1/2, positive/neutral/negative, good/very good etc.).
- Marl and Onyx are vocabularies for emotion and sentiment analysis that can be taken as a starting point for discussion in the CG.
This group will not publish specifications.
In order to join the group, you will need a W3C account.
This is a community initiative. This group was originally proposed on 2013-12-20 by Daniel Molina. The following people supported its creation: Daniel Molina, Vincenzo Masucci, Michele Mostarda, Francesco Adolfo Danza, Roberto Maestre, Paul Buitelaar, Carlos A. Iglesias, Björn Schuller and J. Fernando Sánchez. W3C’s hosting of this group does not imply endorsement of its activities.
If you believe that there is an issue with this group that requires the attention of the W3C staff, please send us email on site-comments@w3.org
Thank you,
W3C Community Development Team
I would like to raise the issue of creating gold-standard and standardising on evaluation metrics for Aspect-Based Sentiment Analysis. In my understanding the current gold standard is SEMEVAL’s, but its based on neg/neut/pos scores in annotations and accuracy scores in evaluation. We strongly believe this is not the way to approach the problem. The fundamental problem is that the neg/pos is not sufficient. For example stating: “I almost liked this hotel” and “This hotel is a terrible place to stay” convey very different sentiment values, but could both be in the neg category. When this is carried over to classification, even a high accuracy rate, like 95% could mean two very different things. :
1. There are no big errors in the classification, just a big number of small errors at the threshold of pos/neg. So this is pretty much perfect classification
2. There are few errors, but they are all extremely significant, e.g. strongly-positive judgement is mistaken for strongly-negative.
We would like to see a gold-standard and results reported with much more fine-grained metrics, so we can really compare approaches based on the results. Currently, such comparisons are not possible and accuracy numbers that are reported in literature can be misleading.