| Title | Neural Networks For Negation Scope Detection |
| Authors | Federico Fancellu, Adam Lopez, Bonnie Webber |
| Year | 2016 |
| URL | http://www.aclweb.org/anthology/P16-1047 |
In negation scope detection, the goal is to identify the words in a sentence that fall under the scope of a negation, e.g., He was not driving the car and she left to go home. Fancellu et al. demonstrate that a bidirectional LSTM without hand-crafted features is competitive with more complex systems that rely on heuristics or rules. It even outperforms them when the test data is from the same domain as the training data.
The LSTM takes as input the tokens in the sentence together with a sentence-length vector that identifies the negation cue. Both inputs are embedded into 50-dimensional embeddings, with the embedding matrix for the cue input having just two embeddings: cue and notcue. Pre-trained word embeddings improve the performance of the system, as does adding universal part-of-speech information (which is embedded in the same way as the token and cue input).
It is worth noting that the system is evaluated on the shared task of *SEM2012, which has a small training set of only 848 sentences. An error analysis shows that incorrect decisions are mainly related to the syntactic structure of the sentence, which is not directly modelled.