Tweet Search Engine

Information Retrieval (IR) system based for a collection of documents (Twitter messages)

Setup

Install anaconda and activate with conda activate.

Install the requirements

pip install -r requirements.txt 

Fetch the Pre-trained Vector Model

sh init.sh

Preprocessing

python src/preprocessing.py

Building an Inverted Index

python src/inverted_index.py

Usage

General Usage

usage: main.py [-h] [-Q QUERY] [-F] [-M {default,BERT}] [-QE {level-1,level-2,none}]

optional arguments:
  -h, --help            show this help message and exit
  -Q QUERY, --query QUERY
                        string to query
  -F, --file            Run queries defined in ../data/topics_MB1-49.txt
  -M {default,BERT}, --method {default,BERT}
                        Neural Method for Retrieval
  -QE {level-1,level-2,none}, --queryexpansion {level-1,level-2,none}
                        Level of query expansion

Running a Query

Note:default corresponds to tf-idf weighted cosine similarity retrieval method

Individual Query

To run specified query in command and output results in console:

python src/main.py -M {BERT|default} -Q "<query>" 

Batch Query

To run default batch defined in data/topics_MB1-49.txt and produce Results.txt:

python src/main.py -M {BERT|default} -F 

TREC Results

Using TF-IDF with Cosine Similarity

map                     all     0.2075
P_10                    all     0.2408

Using stsb-roberta-base BERT model with Cosine Similarity

map                     all     0.0274
P_10                    all     0.0327

Using stsb-roberta-large BERT model with Cosine Similarity

map                     all     0.0356
P_10                    all     0.0388

Using TF-IDF with Cosine Similarity and Query Expansion Level 1

map                   	all	0.2070
P_10                  	all	0.2408

Using TF-IDF with Cosine Similarity and Query Expansion Level 2

map                   	all	0.2070
P_10                  	all	0.2408