Update project_proposal.md

[Paul&Frank(2019)](https://www.aclweb.org/anthology/N19-1368/)s approach is based on narrative texts. As their dataset they use the ROCStories dataset ([Mostafazadeh, 2016]()) which contains a collection of narrative texts. We extend this approach by looking at argumentative texts and extending the procedure for our usage.

## Approach

First of all, we look for a suitable [data set](#dataset). It is important here that these are argumentative texts. The selected data set must then be prepared according to the model. To do this, we first manually annotated each of the four hundred essays from our data set with one of the maslow and one of the reiss motives. <br> After having done that we use fleiss kappa to calculate the Inter Annotator Agreement. For gold standard we then annotated 25 of each 100 essays which was annotated by another one. In sum this calculated to 75 more Annotations per person. *E.g. If I annotated essays 1-101 in the first pass, I will now annotate the first 25 of each of 102-202, 203-303, and 304-404.* In sum we then had 100 essays out of the 400 which were annotated by all four Annotators. <br> Since our selected data was already seperated into train and test, we only had to put our train and test files into correct format (see files attached). <br> For our project we only had a look at the last paragraph of each essay (more details under [data set](#dataset)). For this, we selected the last paragraph of each essay of our dataset in *Comparer.py* and compared each word in it with the concepts from ConceptNet. <br> After doing that we start with the steps equivalent to [this](https://github.com/debjitpaul/Multi-Hop-Knowledge-Paths-Human-Needs). Because some of the steps and attached filed from Debjit Pauls github were not working for us (see project_report, problems) we changed a few things, which can be taken from our README.md. After constructing the subgraphs and extracting the relevant knowledge paths we extract the human needs from the created knowledge paths and assign them to the essays (*Human_needs_assigner.py*). <br> The last thing we do is evaluate and assess our obtained results. <br> For textual analysis we wanted to use OpenFraming. But the tool is till today (march 2021) not working due to an intern server error.

First of all, we look for a suitable [data set](#dataset). It is important here that these are argumentative texts. The selected data set must then be prepared according to the format required by the model. To do this, we begin by manually annotating all of the four hundred essays from our data set with one of the maslow and one of the reiss motives. <br> After having done that we use the fleiss kappa score to calculate the Inter Annotator Agreement. To produce a gold standard, each of us annotated an additional 75 essay, resulting in 100 essays total wich have been annotated by all 4 members.  *E.g. If I annotated essays 1-101 in the first pass, I will now annotate the first 25 of each of 102-202, 203-303, and 304-404.* In sum we then had 100 essays out of the 400 which were annotated by all four Annotators. <br> Since our selected data was already seperated into train and test, we only had to put our train and test files into correct format (see files attached). <br> For our project we only had a look at the last paragraph of each essay (more details under [data set](#dataset)). For this, we selected the last paragraph of each essay of our dataset in *Comparer.py* and compared each word in it with the concepts from ConceptNet. <br> After doing that we start with the steps equivalent to [this](https://github.com/debjitpaul/Multi-Hop-Knowledge-Paths-Human-Needs). Because some of the steps and attached filed from Debjit Pauls github were not working for us (see project_report, problems) we changed a few things, which can be taken from our README.md. After constructing the subgraphs and extracting the relevant knowledge paths we extract the human needs from the created knowledge paths and assign them to the essays (*Human_needs_assigner.py*). <br> The last thing we do is evaluate and assess our obtained results. <br> For textual analysis we wanted to use OpenFraming. But the tool is till today (march 2021) not working due to an intern server error.

## Dataset