neulab / InterpretEval

Interpretable Evaluation for (Almost) All NLP Tasks

by Pengfei Liu, Jinlan Fu, Yang Xiao, Graham Neubig and other contributors.

This project is supported by two following works:

• Interpretable Multi-dataset Evaluation for Named Entity Recognition (EMNLP-2020)

• RethinkCWS: Is Chinese Word Segmentation a Solved Task? (EMNLP-2020)

Final Product: ExplainaBoard (Updating)

Updates:

• Added one SOTA NER system and two top-scoring systems on CoNLL03, check out. Thanks Ikuya Yamada and Stefan Schweter for their sharing results.

1. Motivated Questions

• Performance of many NLP tasks has reached a plateau. What works, and what's next?

• Is XX a solved task? What's left?

• A good evaluation metric can not only rank different systems but also tell their relative advantages (strengths and weaknesses) of them.

• Next-generation of Leaderboard: equip it with powerful analysis ability?

2. Interpretable Evaluation Methodology

The evaluation methodology generally consists of following steps.

2.1 Attribute Definition

Taking NER and CWS tasks for example, we have defined 8 attributes for the NER task, and 7 attributes for the CWS task.

Id	NER	CWS
1	Entity Length	Word Length
2	Sentence Length	Sentence Length
3	OOV Density	OOV Density
4	Token Frequency	Character Frequency
5	Entity Frequency	Word Frequency
6	Label Consistency of Token	Label Consistency of Character
7	Label Consistency of Entity	Label Consistency of Word
8	Entity Density

2.2 Bucketing

Bucketing is an operation that breaks down the holistic performance into different categories. This can be achieved by dividing the set of test entities into different subsets of test entities (regarding spanand sentence-level attributes) or test tokens (regarding token-level attributes).

2.3 Breakdown

Calculate the performance of each bucket.

Summary Measures

Summarize quantifiable results using statistical measures

3. Application

3.1 System Diagnosis

• Self-diagnosis
• Aided-diagnosis

3.2 Dataset Bias Analysis

3.3 Structural Bias Analysis

4. Interpreting Your Results?

4.1 Method 1: Upload your files to the ExplainaBoard website

4.2 Method 2: Run it Locally

Give the Named Entity Recognition task as an example. Run the shell: ./run_task_ner.sh.

The shell scripts include the following three aspects:

• tensorEvaluation-ner.py -> Calculate the dependent results of the fine-grained analysis.

• genFig.py -> Drawing figures to show the results of the fine-grained analysis.

• genHtml.py -> Put the figures drawing in the previous step into the web page.

After running the above command, a web page named tEval-ner.html will be generated for displaying the analysis and diagnosis results of the models.

The running process of the Chinese Word Segmentation task is similar.

4.2.1 Requirements:

• python3
• texlive
• poppler
• pip3 install -r requirements.txt

4.2.2 Analysis and diagnosis your own model.

Take CoNLL-2003 datasets as an example.

• Put the result-file of your model on this path: data/ner/conll03/results/ (It contains three columns separated by space: token, true-tag, and predicted-tag). In order to carry out model diagnosis, two or more model result files must be included. You can also choose one of the result files provided by us as the reference model.

• Name the train- and test-set (the dataset related to your result-file) as 'train.txt' and 'test.txt', and then put them on the path: data/ner/conll03/data/.

• Set the path_data (path of training set), datasets[-] (dataset name), model1 (the first model's name), model2 (the second model's name), resfiles[-] (the paths of the results) in run_task_ner.sh according to your data.

• Run: ./run_task_ner.sh. The analysis results will be generated on the path output_tensorEval/ner/your_model_name/.

   Notably, so far, our system only supports limited tasks and datasets, 
   we're extending them currently!

4.2.3 Generate the HTML code

As introduced in section 4.2.2, we have generated the analysis results on the path output_tensorEval/ner/your_model_name/. Next, we will generate the HTML code base on the analysis results. In the ./run_task_ner.sh, the codes after #run pdflatex .tex are used to generate the HTML code. Before running ./run_task_ner.sh, you need to make sure that you have installed the texlive and poppler.

Other illustrations of the ./run_task_ner.sh code are as follows:

• genFig.py generates the latex codes about the analysis charts (e.g. bar-chart, heatmap).

• pdflatex $file.tex generates a figure with .pdf format based on the latex code.

• pdftoppm -png $file.pdf converts the figure with .pdf into the .png format.

• genHtml.py generates the HTML code that arranges the analysis figures and tables.

4.2.4 Note:

• More than two result files are required. Because comparative-diagnosis is to compare the strengths and weaknesses of the model architectures and pre-trained knowledge between two or more models, it is necessary to input as least two model results.

• The result file must include three columns of words, true-tags, and predicted-tags, separated by space. If your result file is not in the required format, you can modify the function read_data() in file tensorEvaluation-ner.py to adapt to your format.

Here are some generated results of preliminary evaluation systems: Named Entity Recognition (NER), Chinese Word Segmentation (CWS), Part-of-Speech (POS), and Chunking.

• NER
• CWS
• POS
• Chunk