In [1], the work presented has a very strong argument for the need for language models in social computing platforms. This can be deconstructed using the following two sections:
- SUMMARY: The paper first gives a theoretical base to the concepts that are used, along with a survey of related work. Here modality, subjectivity and the other linguistic measures used have been defined to capture the different perceived dimensions of a language model. The claims of all of them are warranted with the help of previous work. The statistical framework considered the problem as an ordered logistic regression one resulting in phrase collinearity (A common property in natural language expressions). The performance of the model is well documented with a sound defense for the validity. The overall accuracy of the model is a clear indicator of its use against the considered baseline classifiers. Implications are drawn on each of the defined measures based on the inferential statistical results of the model.
- REFLECTION: As a proponent for credibility level assessments of social media content, I favor the establishment of well-founded metrics to filter content. The paper is a strong step in the direction with a detailed account of the design process for a good linguistic model. The few immediate design opportunities that are regurgitated from the paper are:
- The creation of a deployable automated system for content analysis adopting such a model. This can be a very interesting project where a Multi-agent Machine learning model using the CREDBANK system as its Supervised Learner, can help classify tweets in real time assigning credits to the source of the content. This will be monitored by another agent which reinforces the Supervised Learner, essentially creating a meta-learner.[2]-[5]
- Adaptation of an ensemble of such models to form a global system which cross-verifies and credits information not just from a single platform but across multiple ones to give the metaphorical “global mean” as against the “local mean” in information. [6]
- The model should account for the linguistic chaos even with newly created “Purrrrr” or “covfefe”. These lexiconic outliers could be captured with the use of Chaos Theory in Reinforcement Learning, which could be an entirely new avenue of research.
The paper also helped me understand the importance of capturing the different dimensions of a language model and corroborating it with evidence with tools of statistical inference.
[1] T. Mitra, G. P. Wright, and E. Gilbert, “A Parsimonious Language Model of Social Media Credibility Across Disparate Events,” Proc. 2017 ACM Conf. Comput. Support. Coop. Work Soc. Comput. – CSCW ’17, pp. 126–145, 2017.
[2] D. Li, Y. Yang, Y.-Z. Song, and T. M. Hospedales, “Learning to Generalize: Meta-Learning for Domain Generalization,” Oct. 2017.
[3] F. Sung, L. Zhang, T. Xiang, T. Hospedales, and Y. Yang, “Learning to Learn: Meta-Critic Networks for Sample Efficient Learning,” Jun. 2017.
[4] R. Houthooft et al., “Evolved Policy Gradients,” Feb. 2018.
[5] Z. Xu, H. van Hasselt, and D. Silver, “Meta-Gradient Reinforcement Learning,” May 2018.
[6] Marios Michailidis (2017), StackNet, StackNet Meta-Modelling Framework, URL https://github.com/kaz-Anova/StackNet
