Anaphora resolution

​IA161 NLP in Practice Course, Course Guarantee: Aleš Horák

Prepared by: Marek Medveď

State of the Art

Anaphora resolution (or pronoun resolution) is the problem of resolving references to earlier or later items in the discourse.
Main approaches:

1. Knowledge-rich approaches:
   1. Syntax-based approaches
   2. Discourse-Based Approaches
   3. Hybrid Approaches
   4. Corpus based Approaches
2. Knowledge-poor Approaches:
   1. Machine learning techniques

References

1. Anaphora Resolution, Studies in Language and Linguistics by Mitkov, R., 2014, Taylor & Francis
2. Wikicrem: A large unsupervised corpus for coreference resolution, Kocijan, V., Camburu, O.-M., Cretu, A.-M., Yordanov, Y., Blunsom, P., and Lukasiewicz, T. (2019).
3. A comprehensive review on feature set used for anaphora resolution, Lata, K., Singh, P., and Dutta, K. (2020).
4. The referential reader: A recurrent entity network for anaphora resolution, Liu, F., Zettlemoyer, L., and Eisenstein, J. (2019).
5. Petra: A sparsely supervised memory model for people tracking, Toshniwal, S., Ettinger, A., Gimpel, K., and Livescu, K. (2020).

Practical Session

Student has to understand Hobbs' definition of anaphora resolution and according to it implement the main function of Hobbs' algorithm in the proposed python script that contains all necessary functions. According to real data (syntactic trees) students test their adjusted program and evaluate it. At the end of the session students hand in the results to prove completing the task. An additional task is to find sentence structures that are not covered by Hobbs' algorithm.

The task:

1. download the script with data from here
2. NLTK package is required for hobbs.py. When running at your computer, paste

   pip3 install nltk --user
   

   in the terminal to install NLTK package. Faculty machines should have nltk already installed.
3. understand Hobbs' definition of anaphora resolution and replace XXX function calls with correct ones
4. find 20 nontrivial sentences wit anaphora: 10 that Hobbs algorithm can recognize and 10 sentences it dos not. You can use ​the Stanford parser to test new sentences - copy the tree to one line and remove the ROOT tag.
5. submit your hobbs.py script with 10 examples that are correctly recognized with hobbs.py and 10 examples that are not correctly recognized by hobbs.py in the homework vault. For each unrecognized example write an explanation into one separate file unrecognized_notes.txt (first column: example id, second column: explanation).

Commands:

1. execute Hobbs script:

   python3 ./hobbs.py demosents.txt He