IRDM Course Project Part II
IRDM 2020
March 15, 2021
1 Task Definition
An information retrieval model is an essential component for many applications (e.g. search,
question answering, recommendation etc.). Similar to the first part of the project, your task
in this assignment is to develop an information retrieval model that solves the problem of
passage retrieval, i.e., a model that can effectively and efficiently return a ranked list of short
texts (i.e. passages) relevant to a given query. In this part of the assignment, your goal is to
improve the basic models that you implemented in the first part.
This is an individual project, therefore, everyone is expected to submit their own project
report.
2 Data
The dataset from previous task is available through this url and the dataset for training and
validation is available through this url. Our dataset consists of 5 files:
• test-queries.tsv is a tab separated file, where each row contains a query ID (qid) and
the query (i.e., query text).
• candidate_passages_top1000.tsv is a tab separated file, containing initial rankings that
contain 1000 passages for each of the given queries (as it was in the first part of the
assignment) in file test-queries.tsv. The format of this file is <qid pid query passage>,
where qid is the query ID, pid is the ID of the passage retrieved, query is the query
text and passage is the passage text, all tab separated. Figure 1 shows some sample
rows from the file.
• train_data.tsv and validation_data.tsv. These are the datasets you will be using for
training and validation. You are expected to train your model on the training set and
evaluate your models’performance on the validation set. In these datasets, you are
given additional relevance column indicating the relevance of the passage to the query
which you will need during training and validation.
3 Subtasks
The course project involves several subtasks that are required to be solved. The four subtasks
of this project are described below.
1
IRDM Course Project Part II
IRDM 2020
March 15, 2021
Figure 1: Sample rows from candidate_passages_top1000.tsv file
- Evaluating Retrieval Quality. (20 marks) Implement methods to compute the average
precision and NDCG metrics. Compute the performance of using BM25 as the retrieval
model using these metrics. Your marks for this part will mainly depend on the implementation
of metrics (as opposed to your implementation of BM25, since we already
focused on that as part of the first assignment). - Logistic Regression (LR). (25 marks) Represent passages and query based on a word
embedding method, (such as Word2Vec, GloVe, FastText, or ELMo). Compute query
(/passage) embeddings by averaging embeddings of all the words in that query (/passage).
With these query and passage embeddings as input, implement a logistic regression
model to assess relevance of a passage to a given query. Describe how you
perform input processing & representation or features used. Using the metrics you
have implemented in the previous part, report the performance of your model based on
the validation data. Analyze the effect of the learning rate on the model training loss.
All implementations for logistic regression algorithm must be your own for this part.
Important Notes:
• The training data size you are given is quite small, so it should not cause you much
difficulty in training but in case you have any issues with the data size, please feel
free to use a sample of the training data.
• If you think it is necessary, you are allowed to use negative sampling for generating
a subset of training data (possibly together with other sampling methods if
needed). - LambdaMART Model (LM). (25 marks) Use the LambdaMART [1] learning to rank algorithm
(a variant of LambdaRank we have learned in the class) from XGBoost gradient
boosting library 1
to learn a model that can re-rank passages. You can command XGBoost
to use LambdaMART algorithm for ranking by setting the appropriate value to
the objective parameter as described in the documentation 2
. You are expected to
carry out hyper-parameter tuning in this task and describe the methodology used in
1https://xgboost.readthedocs.i…
2https://xgboost.readthedocs.i…
2
IRDM Course Project Part II
IRDM 2020
March 15, 2021
deriving the best performing model. Using the metrics you have implemented in the
first part, report the performance of your model on the validation data. Describe how
you perform input processing & representation or features used. - Neural Network Model (NN). (30 marks) Using the same training data representation
from the previous question, build a neural network based model that can re-rank passages.
You may use existing packages, namely Tensorflow or PyTorch in this subtask.
You are expected to justify your neural network architecture by providing the motivation
and how it fits to our problem. You are allowed to use different types of (deep)
neural network architectures (e.g. feed forward, convolutional, recurrent and/or transformer
based neural networks) for this part. Using the metrics you have implemented
in the first part, report the performance of your model on the validation data. Describe
how you perform input processing & representation or features used.
3.1 Submission of Test Results.
You should have one file per model (named LR.txt, LM.txt, and NN.txt, respectively),
where the format of the file is:
<qid1 A1 pid1 rank1 score1 algoname2>
<qid1 A1 pid2 rank2 score2 algoname2>
<qid1 A1 pid3 rank3 score3 algoname2>
<qid1 A1 pid4 rank4 score4 algoname2>
…
The width of columns in the format is not important, but it is important to have exactly
six columns per line with at least one space between the columns. In this format: - The first column is the query number.
- The second column is currently unused and should always be“A1”, to refer to the
fact that this is your submission for Assignment 1. - The third column is the passage identifier.
- The fourth column is the rank the passage/document is retrieved (starting from 1,
down to 100). - The fifth column shows the score (integer or floating point) of the model that generated
the ranking. - The sixth column refers to the algorithm you used for retrieval (would either be LR,
LM or NN, depending on which model you used) .
3
IRDM Course Project Part II
IRDM 2020
March 15, 2021 - Submission
You are expected to submit all the codes you have implemented for all the parts of the
assignment (e.g. evaluation metrics, data representation, logistic regression, LambdaMART
training, neural network implementation, etc.) All the code should be your own and you are
not allowed to reuse any code that is available online. You are allowed to use both Python
and Java as the programming language.
You are also expected to submit a written report whose size should not exceed 6 pages,
including references. Your report should describe the work you have done for each of the
aforementioned steps. Your report should explicitly describe the performance of the models
you have implemented, the input representations (or features) you have used, how you have
used the training and validation sets (any sub-sampling done, etc.), how you have done
hyper-parameter tuning, the neural architecture you have used and why, etc.
You are required to use the SIGIR 2020 style template for your report. You can either use
LaTeX or Word available from the ACM Website 3
(use the“sigconf”proceedings template).
Please do not change the template (e.g. reducing or increasing the font size, margins, etc.). - Deadline
The deadline will be announced later.
References
[1] C. J. Burges. From ranknet to lambdarank to lambdamart: An overview. Technical
Report MSR-TR-2010-82, June 2010.