Extractive Explanations
for Interpretable Text Ranking
Jurek Leonhardt, Koustav Rudra, Avishek Anand
Ranking should be explainable
Neural ranking in a nutshell:
(of course, the following is vastly simplified)
$ \underbrace{\phi(q, d)}_{\textbf{relevance}} = \operatorname{BERT} \underbrace{ (\texttt{[CLS]}\ q\ \texttt{[SEP]}\ d\ \texttt{[SEP]}) }_{\textbf{LLM inputs}} $
LLMs (i.e., Transformers or BERT) are complex and hard to interpret/explain.
What if the trained model is biased? Racist? Sexist?
Extractive explanations
Example document for query: "how long to hold bow in yoga" 🤔
Why should we even show the ranking model the whole document?
Select-and-Rank
Assumption: $k$ sentences of a document are enough to estimate its relevance w.r.t. a query.
The Select-and-Rank paradigm.
A document $d$ is split into sentences $s_i$.
The selector assigns a score to each sentence $s_i$ w.r.t. the query $q$.
The ranker sees only the $k$ highest scoring sentences.
Select-and-Rank
The Select-and-Rank paradigm.
The selector $\Psi$ computes a weight $w_i$ for each sentence:
$\left(w_1, \dots, w_{|d|} \right) = \Psi(q, d)$
End-to-end training: We draw a relaxed $k$-hot sample:1
$\left(\hat{w}_1, \dots, \hat{w}_{|d|} \right) = \operatorname{SubsetSample}(w, k, \tau)$
The query-document relevance is computed by the ranker $\Phi$ using the selected sentences $\hat{d}$:
$\phi(q, d) = \Phi \left(q, \hat{d} \right)$
Each token in $\hat{d}$ is multiplied by its corresponding $\hat{w_i}$ in order to preserve the gradients.
1Sang Michael Xie and Stefano Ermon. Reparameterizable Subset Sampling via Continuous Relaxations. IJCAI 2019.
Selectors
S&R-Lin
S&R-Attn
Ranker
$ \underbrace{\phi(q, d)}_{\textbf{relevance}} = \operatorname{BERT} (\texttt{[CLS]}, \underbrace{q_1, \dots, q_n}_{\textbf{input query}}, \texttt{[SEP]}, \underbrace{\hat{d}_1, \dots, \hat{d}_m}_{\textbf{selected sentences}}, \texttt{[SEP]}) $
Select-and-Rank example
Example selections on Fever. Highlighted sentences contain the answer.
BEIR benchmark
Results of Select-and-Rank models on the BEIR benchmark.
BEIR benchmark
There is a trade-off between the number of sentences $k$ and the effectiveness:
Performance on Fever.
Performance on SciFact.
Comprehensiveness
Comprehensiveness measures the quality of rationales:
How well does the ranking model perform using the document without the selected sentences?
Ranking performance on TREC-DL-Doc’19 using $k = 20$, where $N$ sentences are removed (leaving $k-N$ sentences).
Faithfulness
Faithfulness measures the degree to which the explanations represent the model's reasoning:
How well do the selected sentences represent the document they originate from?
We perform a user study to determine the utility of Select-and-Rank for humans.
The user study interface.
Faithfulness
240 query-document pairs from 30 queries judged by 80 users (4 judgments per instance).
Accuracy of relevance judgments.
Time taken to complete relevance judgments.
Application: Detecting label leakage
Unaltered relevant document
Relevant document with label leakage
Application: Detecting label leakage
Documents where the leakage sentence has been selected.
Distribution of the ranks of the leakage sentence.
Summary
- We proposed Select-and-Rank, a ranking framework that is interpretable by design.
- We showed how Select-and-Rank can be used to explain the decisions for a large number of ranking tasks.
- We performed a user study to highlight the utility of our extractive explanations to humans.
What's next?
Efficiency
Select-and-Rank requires a BERT forward pass for each document.
Can we change the architecture to allow for pre-computations?
Granularity
Select-and-Rank operates strictly on a sentence level.
Can we make the extractive explanations more fine-grained?