University of Washington LING575 Multi-Document Summarization Project

D1

Team Setup

D2

Project 1: Process a docSet

proc_docset.sh input_xml_file1 input_xml_file2 input_xml_file3 training_files_path training_files_output_dir dev_files_path dev_files_output_dir eval_files_path eval_files_output_dir

Structure Overview

src/

Location of all source code

• extract.py
  • Extract topic ids and corresponding doc id from DocSetA (into format such as topic.docSetA.id)
  • Transform doc id into path
• {training|dev|eval}_process.py
  • Use XML.etree.ElementTree library to parse XML
    • If the file is NOT a standard XML (e.g. /corpora/LDC/LDC02T31/apw/1999/19990914_APW_ENG) -> modify the file by adding <DOCSTREAM> tag
  • Find doc id by indexing DOCNO keyword
    • If the file is standard XML, find the doc by looking for id keyword.
  • Extract headline and dateline from HEADLINE and DATELINE keyword.
  • Get individual sentence by doc.TEXT.findall(“P”)

data/

2009_TAC

• Including the 3 guided summary xml files for training, dev and eval set
• The data come from TAC 2010 Guided Summarization task
• devtest and evaltest each has an accompanying categories.txt file that captures the 5 types of topics occurred in the docset

selected_files

outputs/

• Including the results from each of the 3 sets respectively under {training|devtest|evaltest}_output/
  • Create sub directories named by docsetA_id (e.g., D0901A-A) by joining output_directory with docID_docsetID_pairs[docID]
    • Store the outputs in each sub file of the above by the doc_id (e.g., XIN_ENG_20041113.0001)
  • In total, 71 docsetA sub directories under training_output/, 88 under devtest_output and 44 under evaltest_output

results/

Anomalies & Missing Files Handling

• XIN docs only exist after 2000. For these files, we use Fei's convention to find th path.
• XIN docs are actually XIE docs before 2000. XIN's between the period 1996-2000, we use Fei's convention to find th path.
• NYT docs before the year 2000 correspond to files without _ENG in the doc_id.

Other Materials for D2

• Notebook Demo
• Slides
• Report

Misc

• Use XML.etree.ElementTree library to parse XML in extract.py and {training|dev|eval}_process.py
• Use nltk.word_tokenize to tokenize the sentences in {training|dev|eval}_process.py

D3

Initial model

D4

Improved Model

• Clone the pretrained model (with git-lfs so make sure you have it installed https://git-lfs.com)

git lfs install
git clone https://huggingface.co/google/bert_uncased_L-12_H-768_A-12

D5

Final Summarization Systems

D6

Final Report