Background Linking

Code to replicate my experiments for background linking in TREC News

Docker

After cloning the repository build the docker image using the dockerfile:

docker build . -t blimg

Test the setup with sample resources:

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.sample --db sample.db --topics topics.sample.txt \
           --qrels qrels.sample.txt --candidates candidates.sample.txt \
           --nr-terms 100 --output sample.txt --run-tag sample

The ndcg_cut_5 should be 0.7505.

Resources

In order to reproduce the experiments, you need to specify the exact same resources as described below.

• index: Index of the Washington Post Corpus (v2 or v3)
• db: Database with terms and named entities for all topic/candidate docs
• embeddings: Word embedding file
• topics: File with topics (TREC format)
• qrels: Query relevance file for the specified topics
• candidates: Candidate documents

Index

TREC's Washington Post index was build using Anserini, see Regressions for TREC 2019 Background Linking. In order to obtain the corpus, an individual agreement form has to be completed first. The exact command we used is shown bellow (note that we used version 2 of the corpus for the 2019 topics, and version 3 for the 2020 topics):

./target/appassembler/bin/IndexCollection -collection WashingtonPostCollection \
 -input /WashingtonPost.v2/data -generator WashingtonPostGenerator \
 -index lucene-index.core18.pos+docvectors+rawdocs_all \
 -threads 1 -storePositions -storeDocvectors -storeRaw -optimize -storeContents

The obtained index should be stored in bglinking/resources/Index.

Database

A database was created to speed up the graph generation. Named entities and tf-idf terms were stored per candidate document in a database. REL was used for the extraction of named entities, see build_db.py.

The database should be stored in bglinking/resources/db.

Candidates

Candidates were obtained using BM25 + RM3 via Anserini, see Regressions for TREC 2019 Background Linking.

The candidates file should be stored in bglinking/resources/candidates

Embeddings

We made use of embeddings from GEEER, they can be downloaded from this link.

The embeddings should be extracted and stored in bglinking/resources/embeddings.

Topics and Qrels

Topics and query relevance files can be downloaded from the News Track page.

Store in bglinking/resources/topics-and-qrels

graph configuration

• nr-terms: Number of terms used in the graph (default = 100)

• term-tfidf: Scaler for tf-idf weight of node (default = 1.0)

• term-postition: Scaler for position weight of node (default = 0.0)

• term-embedding: Scaler for embedding weight of edges, i.e. cosine similarity term vectors (default = 0.0)

• text-distance: Scaler for distance weight for edges (default = 0.0)

• use-entities: Append named entities to graph nodes

• textrank: Apply textrank to current graph

output

• output: Name of output file
• run-tag: Run-tag in output file

other

• diversity: Apply diversity filter
• stats: Show index stats
• year: Year of TREC edition

Experiments Graph Configurations

Results are stored in bglinking/resources/output

Graph [100 terms, no edges]

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --nr-terms 100 --output simple_graph_19.txt --run-tag simple_graph

Graph [100 terms, edges based on text distance]

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --nr-terms 100 --text-distance 1 --output simple_graph_text_distance_19.txt \
           --run-tag simple_graph_text_distance

Graph [100 terms, edges based on word embeddings]

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-embedding 1 --output simple_graph_term_embedding_19.txt \
           --run-tag simple_graph_term_embedding

Graph [100 terms - weights based on term position]

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --nr-terms 100 --term-position 1 --output simple_graph_term_position_19.txt \
           --run-tag simple_graph_term_position

Graph configurations combining: term position, text distance & word embedding.

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --nr-terms 100 --term-position 1 --text-distance 1 \
           --output simple_graph_term_position_text_rank_19.txt \
           --run-tag simple_graph_term_position_text_rank

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 \
           --output simple_graph_term_position_text_embedding_19.txt \
           --run-tag simple_graph_term_position_text_embedding

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --text-distance 1 --term-embedding 1 \
           --output simple_graph_term_position_text_distance_19.txt \
           --run-tag simple_graph_term_position_text_distance

Add named entities to graph nodes (simplest configuration)

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --use-entities --output only_entities_19.txt --run-tag only_entities

Add named entities to graph nodes (best performing run)

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 --use-entities \
           --output best_graph_entities_19.txt \
           --run-tag best_graph_entities

Test effect of novelty algorithm (without named entities)

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 --novelty 0.05 \
           --output best_graph_novelty_19.txt \
           --run-tag best_graph_novelty

Test effect of novelty algorithm (with named entities)

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 --use-entities --novelty 0.05 \
           --output best_graph_novelty_entities_19.txt --run-tag best_graph_novelty_entities

Test best run with TextRank algorithm

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 --textrank \
           --output best_graph_textrank_19.txt \
           --run-tag best_graph_textrank

Test diversification

docker run --rm -v $PWD/bglinking/resources:/opt/background-linking/bglinking/resources blimg \
           --index lucene-index.core18.pos+docvectors+rawdocs_all --db entity_database_19.db \
           --topics topics.backgroundlinking19.txt --qrels qrels.backgroundlinking19.txt \
           --candidates run.backgroundlinking19.bm25+rm3.topics.backgroundlinking19.txt \
           --embedding WKN-vectors/WKN-vectors.bin \
           --nr-terms 100 --term-position 1 --term-embedding 1 --diversify --use-entities \
           --output best_graph_diversified_19.txt --run-tag best_graph_diversified