NegNN (Neural Network for Automatic Negation Detection)

The code implements a feed-forward NN (/feed_forward) and a BiLSTM(/bilstm_class or /bilstm) to perform automatic negation scope detection.

Data

Training, test and development data can be found in the /data folder. For training, development and initial testing, we used the data released for the *SEM2012 shared task; please refer to the shared task and related papers for information regarding the annotation style. Additional test data extracted from Simple Wikipedia is available in /data/test/simple_wiki. The data was manually annotated following the guidelines released during the *SEM2012 shared task. Please refer to the .pdf file for ulterior information. The python code in /reader used to read in the data is part of the code made available by Packard et al. (2014) ("Simple Negation Scope Resolution through Deep Parsing: A Semantic Solution to a Semantic Problem").

Dependencies

• Tensorflow (tested on v. 0.7.1)
• scikit-learn (tested on v. 0.17.1) - for score report purposes only, feel free to use any other library instead -
• numpy (tested on v. 1.11.0)

Train

To train the model, first go to the parent directory of the repository and run python NegNN/(bilstm_class|bilstm|feed_forward)/train.py. /bilstm_class is a more elegant implementation that wraps the BiLSTM code inside a separate class so to avoid any repetition. There seems to be however problems with this implementation when run on MacOsX El capitan 10.11.5; if so, please run the less elegant implementation in the /bilstm folder train.py accepts the following flags

-h, --help            show this help message and exit
  --embedding_dim EMBEDDING_DIM
                        Dimensionality of character embedding (default: 50)
  --max_sent_length MAX_SENT_LENGTH
                        Maximum sentence length for padding (default:100)
  --num_hidden NUM_HIDDEN
                        Number of hidden units per layer (default:200)
  --num_classes NUM_CLASSES
                        Number of y classes (default:2)
  --num_epochs NUM_EPOCHS
                        Number of training epochs (default: 50)
  --learning_rate LEARNING_RATE
                        Learning rate(default: 1e-4)
  --scope_detection [SCOPE_DETECTION]
                        True if the task is scope detection or joined
                        scope/event detection
  --noscope_detection
  --event_detection [EVENT_DETECTION]
                        True is the task is event detection or joined
                        scope/event detection
  --noevent_detection
  --POS_emb POS_EMB     0: no POS embeddings; 1: normal POS; 2: universal POS
  --emb_update [EMB_UPDATE]
                        True if input embeddings should be updated (default:
                        False)
  --noemb_update
  --normalize_emb [NORMALIZE_EMB]
                        True to apply L2 regularization on input embeddings
                        (default: False)
  --nonormalize_emb
  --test_set TEST_SET   Path to the test filename (to use only in test mode
  --pre_training [PRE_TRAINING]
                        True to use pretrained embeddings
  --nopre_training
  --training_lang TRAINING_LANG
                        Language of the tranining data (default: en)

Please disregard the event detection flag for the moment; this might be part of future work.