updated/enhanced modeling readme

modeling/README.md

@@ -6,4 +6,66 @@ The files in this directory are responsible for training and optimizing the mach
 
 * `1_featurize_training_data.py` : responsible for reading in the manually-tagged training data and featurizing it for model training. 
 
-* `2_grid_search_CV.py` : runs a grid search to optimize hyperparameters for both the opinion and sentiment models. In practice, this was always run using [domino](http://www.dominoup.com/). Note that the final models tuned by the grid search are ultimately pickled, for later use in YUMM's main summary-generation pipeline (see `../classes`). 
+* `2_grid_search_CV.py` : runs a grid search to optimize hyperparameters for both the opinion and sentiment models. In practice, this was always run using [domino](http://www.dominoup.com/). Note that the final models tuned by the grid search are ultimately pickled, for later use in YUMM's main summary-generation pipeline (see `../classes`). The results of the grid search are stored in `./results`
+
+## Model Results
+
+YUMM takes a two-stage modeling approach, first attempting to classify a given sentence as Opinionated vs. Not Opinionated, and only subsequently attempting to classify Opinionated sentences as being Positive or Negative. 
+
+Currently, the Opinion Model is an SVM, and the Sentiment Model is a Logisitc Regression. Both models are currently trained using an extremely lightweight featureset (only 16 total features) that includes primaryily lexion- and syntactic-features (e.g. number/fraction of positive/negative words, number/fraction of various parts of speech). 
+
+Given the minimal featuresets, both models currently perform quite respectably on cross validation:
+
+##### Opinion Model:
+
+Here is the performace report for the tuned SVM opinion model (0 is objective, 1 is subjective).
+
+```
+Opinion_Model Classification Report:
+             precision    recall  f1-score   support
+
+          0       0.40      0.62      0.49       152
+          1       0.85      0.69      0.76       456
+
+avg / total       0.74      0.67      0.69       608
+```
+
+![](./results/Opinion_Model.png =500x)
+
+##### Sentiment Model: 
+
+Here is the performance report for the tuned Logisitc Regression sentiment model (1 is positive, -1 is negative): 
+
+```
+Sentiment_Model Classification Report:
+             precision    recall  f1-score   support
+
+         -1       0.46      0.75      0.57        84
+          1       0.93      0.79      0.86       361
+
+avg / total       0.84      0.78      0.80       445
+```
+
+![](./results/Sentiment_Model.png =500x)
+
+Note that precision is currently quite low on classifying Negative sentences; I expect that this is primarily due to the small amount of training data that was available, the majority of which was *not* negative: 
+
+```
+Size of complete development set: 3040
+Target class breakdowns:
+ 1    1759 (positive)
+ 0     816 (objective)
+-1     465 (negative)
+```
+
+Although the model does its best to account for this imabalance (e.g. by heavily weighting negative examples in the loss function), sometimes more data is simply needed. 
+
+Future iterations of YUMM intend to build out the modeling featureset further, hopefully incorporating more lexicon-based features derived from a WordNet propogation scheme such as that of [Blair-Goldensohn et al. (2008)](http://www.ryanmcd.com/papers/local_service_summ.pdf). However, YUMM aims to avoid the complexity of a full-blown Bag-of-Words/TFIDF approach.  
+
+
+
+
+
+
+
+