Effective Distributed Representations for Academic Expert Search

This repository contains (partial) code that was used in conducting research for the "Effective Distributed Representations for Academic Expert Search" paper that appeared at the SDP 2020 workshop @ EMNLP 20202 by Mark Berger, Jakub Zavrel and Paul Groth.

Link to the paper: Effective distributed representations for academic expert search

Disclaimer

The SDP 2020 paper was written and submitted as a result of the MSc thesis project by Mark Berger. Due to their specificity to the scope of the conducted thesis project, some parts of the processes referenced in the paper are not verbosely presented in this repository, such as the explicit code for the creation and population of the FAISS indexes, and the experiments we performed before arriving at the baseline BERT and GloVe embeddings. However, with the provided datasets and the populated and ready-to-go FAISS indexes, the presented thought process in the notebooks, and the final expert search API, we believe future visitors will be able to re-use our data and ideas to conduct new research.

Structure

The repository is structured in the following way:

The Notebooks directory contains Jupyter notebooks concerning our work process around preparing and embedding the papers and obtaining author metadata. Moreover, the Jupyter notebook at Notebooks/Embedding_and_retrieval/Ranking_complete.ipynb can be considered as the main logic of the produced expert search system and contains most of the processes explained in the paper. This file, however, can better be accessed and read through the ExpertSearchAPI/API/Functions/re_ranking.py file, where each function is documented and explained.

The Utility directory contains fully documented Python scripts containing the data preprocessing, data alignment, custom embedding, and evaluation functions. Users wanting to have a good starting point to re-use our ideas should look at this directory for clearly documented methodologies we used. Note however that the file ExpertSearchAPI/API/Functions/re_ranking.py contains all of the voting model, data fusion and actual retrieval logic that we used in our research and our final system, and is also a very good overview of our work and a good starting point.

Data

The accompanying datasets are published and can be accessed at https://doi.org/10.5281/zenodo.4075181.

Citing this code & using this data

Please cite the following paper when using this code:

@inproceedings{berger-etal-2020-effective,
    title = "Effective distributed representations for academic expert search",
    author = "Berger, Mark  and
      Zavrel, Jakub  and
      Groth, Paul",
    booktitle = "Proceedings of the First Workshop on Scholarly Document Processing",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.sdp-1.7",
    pages = "56--71",
    abstract = "Expert search aims to find and rank experts based on a user{'}s query. In academia, retrieving experts is an efficient way to navigate through a large amount of academic knowledge. Here, we study how different distributed representations of academic papers (i.e. embeddings) impact academic expert retrieval. We use the Microsoft Academic Graph dataset and experiment with different configurations of a document-centric voting model for retrieval. In particular, we explore the impact of the use of contextualized embeddings on search performance. We also present results for paper embeddings that incorporate citation information through retrofitting. Additionally, experiments are conducted using different techniques for assigning author weights based on author order. We observe that using contextual embeddings produced by a transformer model trained for sentence similarity tasks produces the most effective paper representations for document-centric expert retrieval. However, retrofitting the paper embeddings and using elaborate author contribution weighting strategies did not improve retrieval performance.",
}

Please cite the following when using the data:

@misc{berger2020data,
  doi = {10.5281/ZENODO.4075166},
  url = {https://zenodo.org/record/4075166},
  author = {Berger,  Mark},
  title = {Datasets for Effective Distributed Representations for Academic Expert Search},
  publisher = {Zenodo},
  year = {2020},
  copyright = {Creative Commons Attribution 4.0 International}
}