Skip to content

Latest commit

 

History

History
144 lines (91 loc) · 6.62 KB

README.md

File metadata and controls

144 lines (91 loc) · 6.62 KB
Raw

Entity Graph for German

The entity-graph repository provides an implementation for computing, analyzing and visualizing German texts using the entity graph.

Entity graph for first 50 sentences of Die Erbschaft by Arthur Schnitzler.

The Entity Graph

The entity graph, first introduced by Guinaudeau and Strube (2013), represents the relations between sentences and entities in a text and thus computes local coherence. We have extended it by adding a syntactic category for possession modifiers and by reducing the weights for entities found in embedded structures. Furthermore, we extracted a set of graph metrics and show that they are valuable features in text analysis of literary works and author classification.

For more details, please consult the paper by Suter and Strube (2018). If you would like to use this code for academic work, please cite this paper.

@InProceedings{entgraph2018,  
Title                    = {Extending and Exploiting the Entity Graph for Analysis, Classification and Visualization of German Texts},
  Author                   = {Suter, Julia and Strube, Michael},
  Booktitle                = {Proceedings of the 14th Conference on Natural Language Processing (KONVENS 2018)},
  Year                     = {2018}}

Parsing

This work employs ParZu and CorZu as syntactic parser and coreference resolution systems. For full functionality, install ParZu and CorZu and set the right paths in settings.py.

However, the system works without ParZu and CorZu as well. You can pass already parsed texts to the functions as long as they are CoNLL-formatted.

Requirements

The system is implemented in Python 3.7 and tested under Ubuntu Linux, but it should work under other environments which have Python installed (no guarantee).

If you want to use the embedded parser functions, please install ParZu and CorZu.

For running this code, you will need the following libraries installed:

  • numpy
  • matplotlib
  • scipy
  • networkx
  • pygobject3
  • gtk3
  • graph-tool (click here for Anaconda installation help)

Datasets

For evaluating on the sentence reordering tasks, we used the German Tüba/DZ news corpus Tuebadz-10.0-conll2010_utf8. This corpus provides gold annotations for syntactic parsing and coreference resolution. In 0_Datasets/Tueba_DZ, you find a zipped collection of single texts with gold annotations, unparsed and parsed with Parzu/CorZu.

  • gold_docs
  • unparsed_docs
  • parzu_parsed_docs

How to Run

Entity Graph for single document

To compute the entity graph and average out degree for a document, you can pass the filename to the test_driver function in RUN_entity_graph.py.

Alternatively, you can use the command line. The flag -p indicates that the file was pre-parsed and does not need to be passed to ParZu and CorZu.

python RUN_entity_graph.py ./1_Info/single_docs/KAFKA_Tuersteher.txt
python RUN_entity_graph.py ./1_Info/single_docs/KAFKA_Tuersteher_parsed.txt -p

Per default, the test driver function will create the entity graph with networkx and save it in the /2_Results/entity_graph directory. You can disable this by setting draw_single_graph to False.

Evaluation: Sentence Reordering Task

The evaluation script RUN_evaluation.py will evaluate the entity graph on the sentence reordering tasks discrimination and insertion as described in the paper. Simply run the script with python. Do not forget to unzip 0_Datasets/Tueba_DZ.tar.gz.

In the config.py, you can select the one-mode projection. Set version to:

  • pu for P unweighted
  • pw for P weighted
  • pacc or full for P accumulated

You can also activate or deactivate the possession modifiers and weight reduction extensions.

Original version

possession_modifiers = False
reduce_weights = False

Adjusted version

possession_modifiers = True
reduce_weights = True

Without possession modifiers

possession_modifiers = False
reduce_weights = True

Without weight reduction

possession_modifiers = True
reduce_weights = False

Of course, you can change many other settings, for instance the weights for the different syntactic role categories.

Correlations between Entity Graph Evaluation Metrics and Text Features

The Jupyter notebook Entity Graph Correlations.ipynb in 3_Scripts/3_JupyterNotebooks/Entity Graph Correlations visualizes the correlations between the entity graph metrics (discrimination accuracy, insertion score and insertion accuracy) and the features of the respective texts. The feature set includes the number of sentences per text, the number of discourse entities per text, percentage of connected sentence pairs, the coherence measure (= average outdegree), average frequencies of discourse entities, and average number of discourse entities per sentence.

Graph metrics and author/genre classification

From the entity graph, you can extract graph metrics that can be used as features for training a Support Vector Classifier for any text classification. In order to reproduce the classification experiments described in the paper, follow the these instructions.

  1. Unzip /0_Datasets/GUTENBERG_parsed_texts.tar.gz.

  2. Run GET_entity_graphs_for_corpus.py in order to pre-compute and save the entity graph for each text sample. Select the desired projection version pu, pw or pacc in config.py

  3. Run GET_entgraph_features.py in order to compute and save graph metric features for each saved entity graph. Select the desired projection version pu, pw or pacc in config.py .

  4. Run GET_baseline_features.py in order to compute and save the linguistic features used for the baseline system.

  5. Run RUN_classification.py in order to evaluate the performance of the entity graph and baseline features on text classification. Select feature set, projection type, classification task and other settings in the settings part at the beginning of the script.

Note that the results computed with these scripts may vary slightly from the ones described in the paper due to minor changes and updates. However, the relations in performance between the different experiments follow the same trends

Contact

If you have any questions, please do not hesitate to contact me by email.