softgitron/LeviRank

Relevant paper and code repository links

Closed this issue · 6 comments

softgitron commented
Relevant paper and code repository links
ashishrana160796 commented

Hi All,

We can post all the paper links here for helping each other and may be also express why that paper/code is important in the later stages.


Comparative argument retrieval paper links:

* Combined Approaches: {Mixed Methods}
    - Mixed Approaches, identify bottlenecks for improvements. {Paper: 3, Code: 4}
    - paper: http://ceur-ws.org/Vol-2936/paper-215.pdf, code: https://github.com/JanNiklasWeder/Touche-21-Task-2

* Pre-trained Language Models and NLU: {Important paper**}
    - Seq2Seq, Transformer, and BERT embedding. {Paper: 3, Code: 4.5**}
    - paper: http://ceur-ws.org/Vol-2696/paper_210.pdf, code: https://github.com/skoltech-nlp/touche

* DistilBERT-based Argumentation Retrieval: {Our Baseline Paper**}
    - query expansion, argument extraction, scoring, and sorting + BERT {Paper: 4, Code: 1}
    - paper: http://ceur-ws.org/Vol-2936/paper-209.pdf, code: https://github.com/jmriebold/BoilerPy3, https://github.com/adbar/trafilatura, https://github.com/UKPLab/sentence-transformers

* ensemble based retrieval methods + BERT re-ranking methodology:
    - comparative argument touche {Paper: 3, Code: 2}
    - paper: http://ceur-ws.org/Vol-2936/paper-211.pdf, code: https://github.com/Georgetown-IR-Lab/OpenNIR

* Search Engine Paper: Usage of whoosh and targer library {can be used to add intrepretability component at last stage}
    - whoosh library usage for index creation: https://whoosh.readthedocs.io/en/latest/quickstart.html {Paper: 2, Code: 1}
    - paper: http://www.dei.unipd.it/~ferro/CLEF-WN-Drafts/CLEF2020/paper_178.pdf, code: https://github.com/uhh-lt/targer
    
* Baseline paper approaches: http://ceur-ws.org/Vol-2696/paper_207.pdf
* Interesting Approach: http://ceur-ws.org/Vol-2696/paper_174.pdf
* [X] Not good code: https://github.com/touche-webis-de/gienapp21

Concepts Used:

https://github.com/touche-webis-de/weder21multi-stage
BERT ranker: https://arxiv.org/pdf/1910.14424.pdf
docT5query: https://github.com/castorini/docTTTTTquery
Document Query Expansion paper: https://arxiv.org/pdf/1904.08375.pdf
ashishrana160796 commented

Query Expansion Methods discussion:

* Synonym and antonym generation method: {paper: http://ceur-ws.org/Vol-2936/paper-209.pdf, code: https://github.com/touche-webis-de/weder21/blob/main/src/preprocessing/QueryExpansion.py}
  - We should use synonyms and antonyms of comparative quantities + noun alternative replacements for constructing alternative queries in CNF/DNF form.
  
* More relevant approach, Doc2Query expando-mono-duo design-pattern (H-1): {paper: https://arxiv.org/pdf/2101.05667.pdf, code: https://www.tira.io/t/expanded-passages-for-the-touche-22-task-2-argument-retrieval-for-comparative-questions/578, touche blog link: https://www.tira.io/t/expanded-passages-for-the-touche-22-task-2-argument-retrieval-for-comparative-questions/578}
  - MS MACRO trained docT5queries are available for touche 22 documents, fine-tuning may be required for better performance (future work).
  - Ideal SOTA design needs to be followed for the implementation.
ashishrana160796 commented

Papers that can be relevant for evaluation metrics:

* QMUL-SDS's "step-by-step" binary classification approach: https://github.com/XiaZeng0223/sciverbinary
* SCIFACT task and its evaluation metrics : https://arxiv.org/pdf/2004.14974.pdf
* VERT5ERINI System Approach: https://aclanthology.org/2021.louhi-1.11.pdf
ashishrana160796 commented

Approaches are finalized, closing the issue!