This repository contains work for our Final Project of TTIC 31190: NLP. It is a collaborative project with Benjamin Rothschild(@b-nroths).

Argumentation mining aims to automatically identify structured argument data from unstructured text. One important part of argumentation mining is claim and stance identification. Most of the current approaches are engineered to address specific domains, for example a specific model might be built just to analyze claims in court documents using attributes specific to court documents and legal vocabulary. However, argumentative sentences are often characterized by common rhetorical structures, independently of the domain. The goal is to explore a method that exploits structured parsing information to detect claims without resorting to topic specific information.

Claim stance classification has become an active field of research in the past years. We show that claim identification could be further improved with features includ- ing mean embedding representation of a sentence, the similarity between sentence and topic target, the location of the sen- tence, length of sentence and PageRank score of the sentence. We also achieve a better than average accuracy instance classification using sentiment analysis be- tween claim and topic. This paper also shows how the combination of many fea- tures types and optimization methods are necessary to achieve accuracy in this NLP task.

The dataset we are using is the Claim Stance Dataset from IBM Debater project which can be accessed here http: //www.research.ibm.com/haifa/ dept/vst/debating_data.shtml. It contains 2,394 labeled claims for 55 topics that are pulled from 1,065 Wikipedia articles. For each article we are given the following data points: • Full text from Wikipedia (text) • Topic Target (text) • Claim Text (text) • Claim Start Index (integer) • Claim End Index (integer) • Stance (Pro or Con)

We divided this research problem into two parts, claim identification and stance identification.

Claim identification is to classify whether a sentence is a claim or not.

Stance classification focuses on the problem of de- termining if a claim is PRO or CON towards its topic.

In this paper, we’ve shown that claim identifica- tion, and claim stance classification could bene- fit from a combination of many features types and optimization methods. A remaining challenge is to build models that are less dependent on train- ing data type and achieve similar results on a vari- ety of data sets including online forum discussions and newspaper articles.

For further description, please check our report here