Kristian Aune

Head of Customer Success, Vespa

05 Nov 2019

In the September Vespa product update, we mentioned Tensor Float Support, Reduced Memory Use for Text Attributes, Prometheus Monitoring Support, and Query Dispatch Integrated in Container.

This month, we’re excited to share the following updates:

Nearest Neighbor and Tensor Ranking

Tensors are native to Vespa. We compared elastic.co to vespa.ai testing nearest neighbor ranking using dense tensor dot product. The result of an out-of-the-box configuration demonstrated that Vespa performed 5 times faster than Elastic. View the test results.

Optimized JSON Tensor Feed Format

A tensor is a data type used for advanced ranking and recommendation use cases in Vespa. This month, we released an optimized tensor format, enabling a more than 10x improvement in feed rate. Read more.

Matched Elements in Complex Multi-value Fields 

Vespa is used in many use cases with structured data – documents can have arrays of structs or maps. Such arrays and maps can grow large, and often only the entries matching the query are relevant. You can now use the recently released matched-elements-only setting to return matches only. This increases performance and simplifies front-end code.

Large Weighted Set Update Performance

Weighted sets in documents are used to store a large number of elements used in ranking. Such sets are often updated at high volume, in real-time, enabling online big data serving. Vespa-7.129 includes a performance optimization for updating large sets. E.g. a set with 10K elements, without fast-search, is 86.5% faster to update.

Datadog Monitoring Support

Vespa is often used in large scale mission-critical applications. For easy integration into dashboards,
Vespa is now in Datadog’s integrations-extras GitHub repository.
Existing Datadog users will now find it easy to monitor Vespa.
Read more.

About Vespa: Largely developed by Yahoo engineers, Vespa is an open source big data processing and serving engine. It’s in use by many products, such as Yahoo News, Yahoo Sports, Yahoo Finance, and the Verizon Media Ad Platform. Thanks to feedback and contributions from the community, Vespa continues to grow.

We welcome your contributions and feedback (tweet or email) about any of these new features or future improvements you’d like to request.

Photo by Norbert
Braun on Unsplash

If you are planning to implement search functionality but have not
yet collected data from user
interactions to
train ranking models, where should you begin? In this series of
blog posts, we will examine the concept of zero-shot text ranking.
We implement a hybrid ranking method using Vespa and evaluate it on a
large set of text relevancy datasets in a zero-shot setting.

In the first post, we will discuss the distinction between in-domain
and out-of-domain (zero-shot) ranking and present the BEIR benchmark.
Furthermore, we highlight situations where in-domain embedding
ranking effectiveness does not carry over to a different domain in
a zero-shot setting.

Introduction

Pre-trained neural language models, such as BERT, fine-tuned for
text ranking, have demonstrated remarkable effectiveness compared
to baseline text ranking methods when evaluating the models in-domain.
For example, in the Pretrained Transformer Language Models for
Search
blog post series, we described three methods for using pre-trained
language models for text ranking, which all outperformed the
traditional lexical matching baseline
(BM25).

In the transformer ranking blog series,
we used in-domain data for training and production (test), and the
documents and the queries were drawn from the same in-domain data
distribution.

In-domain trained and evaluated ranking methods. All models are
end-to-end represented using Vespa, open-sourced in the msmarco-ranking
sample
app.
The Mean Reciprocal Rank
(MRR@10) is
reported for the dev query split of the MS MARCO passage ranking
dataset.

This blog post looks at zero-shot text ranking, taking a ranking
model and applying it to new domains without adapting or fine-tuning
the model.

In-domain training and inference (ranking) versus zero-shot inference
(ranking) with a model trained in a different domain.

Information Retrieval Evaluation

Information retrieval (IR) evaluation is the process of measuring
the effectiveness of an information retrieval system. Measuring
effectiveness is important because it allows us to compare different
ranking strategies and identify the most effective at retrieving
relevant information.

We need a corpus of documents and a set of queries with relevance
judgments to perform an IR evaluation. We can start experimenting
and evaluating different ranking methods using standard IR metrics
such as nDCG@10, Precision@10, and Recall@100. These IR metrics
allow us to reason about the strengths and weaknesses of proposed
ranking models, especially if we can evaluate the model in multiple
domains or relevance datasets. Evaluation like this contrasts the
industry’s most commonly used IR evaluation metric, LGTM (Looks
Good To Me)@10, for a small number of queries.

Evaluating ranking models in a zero-shot setting

In BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of
Information Retrieval Models,
Thakur et al. introduce a benchmark for assessing text ranking
models in a zero-shot setting.

The benchmark includes 18 diverse datasets sampled from different
domains and task definitions. All BEIR datasets have relevance
judgments with varying relevance grading resolutions. For example,
TREC-COVID, a dataset of the BEIR benchmark, consists of 50 test
queries, with many graded relevance labels, on average 493,5 document
judgments per query. On the other hand, the BEIR Natural Questions
(NQ) dataset uses binary relevance labels, with 4352 test queries,
with, on average, 1.2 judgments per query.

The datasets included in BEIR also have a varying number of documents,
queries, and document lengths, but all the datasets are monolingual
(English).

Statistics of the BEIR datasets. Table from BEIR: A Heterogeneous
Benchmark for Zero-shot Evaluation of Information Retrieval
Models; see also
BEIR Benchmark
datasets.

The BEIR benchmark uses the Normalised Cumulative Discount
Gain
(nDCG@10) ranking metric. The nDCG@10 metric handles both datasets
with binary (relevant/not-relevant) and graded relevance judgments.
Since not all the datasets are available in the public domain (E.g.,
Robust04), it’s common to report nDCG@10 on the 13 datasets that
can be downloaded from the BEIR GitHub repository
or
using the wonderful ir_datasets
library. It’s also possible to aggregate the reported nDCG@10 metrics
per dataset to obtain an overall nDCG@10 score, for example, using
the average across the selected BEIR datasets. It’s important to
note which datasets are included in the average overall score, as
they differ significantly in retrieval difficulty.

Zero-Shot evaluation of models trained on Natural Questions

The most common BEIR experimental setup uses the MS MARCO labels
to train models and apply the models in a zero-shot setting on the
BEIR datasets. The simple reason for this setup is that MS MARCO
is the largest relevance dataset in the public domain, with more
than ½ million training queries. As with NQ, there are few, an
average of 1.1, relevant passages per query. Another setup variant
we highlight in detail in this section is to use a ranking model
trained on Natural Questions(NQ) labels with about 100K training
queries and evaluate it in an out-of-domain setting on MS MARCO
labels.

MS MARCO and Natural Questions datasets have fixed document corpora,
and queries are split into train and test. We can train a ranking
model using the train queries and evaluate the ranking method on
the test set. Both datasets are monolingual (English) and have user
queries formulated as natural questions.

MS MARCO sample queries

how many years did william bradford serve as governor of plymouth colony?

define preventive

color overlay photoshop

Natural Questions (NQ) sample queries

what is non controlling interest on balance sheet

how many episodes are in chicago fire season 4

who sings love will keep us alive by the eagles

On the surface, these two datasets are similar. Still, NQ has longer
queries and documents compared to MS MARCO. There are also subtle
differences in how these datasets were created. For example, NQ
uses passages from Wikipedia only, while MS MARCO is sampled from
web search results.

The above table summarizes basic statistics of
the two datasets. Words are counted after simple space tokenization.
Query lengths are calculated using the dev/test splits. Both
datasets have train splits with many queries to train the ranking
model.

In open-domain question answering with
Vespa,
we described the Dense Passage Retriever (DPR) model, which uses
the Natural Questions dataset to train a dense 768-dimensional
vector representation of both queries and Wikipedia paragraphs. The
Wikipedia passages are encoded using the DPR model, representing
each passage as a dense vector. The Wikipedia passage vector
representations can be indexed and efficiently searched using an
approximate nearest neighbor
search.

At query time, the text query is encoded with the DPR model into a dense
vector representation used to search the vector index. This ranking
model is an example of dense retrieval over vector text
representationns
from BERT. DPR was one of the first dense retrieval methods that
outperformed the BM25 baseline significantly on NQ. Since then,
much water has flown down the river, and dense vector models are
closing in on more computationally expensive cross-encoders in an
in-domain setting on MS MARCO.

In-domain effectiveness versus out-of-domain in a zero-shot setting

The DPR model trained on NQ labels outperforms the BM25 baseline
when evaluated on NQ. This is an example where the in-domain
application of the trained model improves the ranking accuracy over
baseline BM25.

In-domain evaluation of the Dense Passage Retriever (DPR). DPR is
an example of Embedding Based Retrieval (EMB).

Suppose we use the DPR model trained on NQ (and other question-answering
datasets) and apply the model on MS MARCO. Then we can say something
about generalization in a zero-shot setting on MS MARCO.

Out-of-domain evaluation of the Dense Passage Retriever (DPR). DPR
is an example of Embedding Based Retrieval (EMB). In this zero-shot
setting, the DPR model underperforms the BM25 baseline.

This case illustrates that in-domain effectiveness does not necessarily
transfer to an out-of-domain zero-shot application of the model.
Generally, as observed on the BEIR dense
leaderboard,
dense embeddings models trained on NQ labels underperform the BM25
baseline across almost all BEIR datasets.

Summary

In this blog post, we introduced zero-shot and out-of-domain IR
evaluation. We also introduced the important BEIR benchmark.
Furthermore, we highlighted a case study of the DPR model and its
generalization when applied out-of-domain in a zero-shot setting.

We summarize this blog post with the following quote from the
BEIR paper:

In-domain performance is not a good indicator for out-of-domain
generalization. We observe that BM25 heavily underperforms neural
approaches by 7-18 points on in-domain MS MARCO. However, BEIR
reveals it to be a strong baseline for generalization and generally
outperforming many other, more complex approaches. This stresses
the point that retrieval methods must be evaluated on a broad range
of datasets.

Next blog post in this series

In the next post in this series on zero-shot ranking, we introduce a
hybrid ranking model, a model which combines multi-vector representations with BM25.
This hybrid model overcomes the limitations of single-vector embedding models,
and we prove its effectiveness in a zero-shot setting on the BEIR benchmark.

Photo by Tamarcus Brown
on Unsplash

Where should you begin if you plan to implement search functionality
but have not yet collected data from user
interactions to
train ranking models?

In the first
post in
the series, we introduced the difference between in-domain and
out-of-domain (zero-shot) ranking. We also presented the BEIR
benchmark and highlighted cases where in-domain effectiveness does
not transfer to another domain in a zero-shot setting.

In this second post in this series, we introduce and evaluate three
different Vespa ranking methods on the
BEIR benchmark in a zero-shot
setting. We establish a new and strong BM25 baseline for the BEIR
dataset, which outperforms previously reported BM25 results. We
then show how a unique hybrid approach, combining a neural ranking
method with BM25, outperforms other evaluated methods on 12 out of
13 datasets on the BEIR benchmark. We also compare the effectiveness
of the hybrid ranking method with emerging few-shot methods that
generate in-domain synthetic training data via prompting large
language models (LLMs).

Establishing a strong baseline

In the BEIR paper,
the authors find that BM25
is a strong generalizable baseline text ranking model. Many, if not
most, of the dense single vector embedding models trained on MS
MARCO labels are outperformed by BM25 in an out-of-domain setting.
Quote from BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation
of Information Retrieval
Models:

In-domain performance is not a good indicator for out-of-domain
generalization. We observe that BM25 heavily underperforms neural
approaches by 7-18 points on in-domain MS MARCO. However, BEIR
reveals it to be a strong baseline for generalization and generally
outperforming many other, more complex approaches. This stresses
the point, that retrieval methods must be evaluated on a broad range
of datasets.

What is interesting about reporting BM25 baselines is that there
are multiple implementations, variants, and performance tweaks, as
demonstrated in Which BM25 Do You Mean? A Large-Scale Reproducibility
Study of Scoring
Variants.
Unfortunately, various papers have reported conflicting results for BM25 on
the same BEIR benchmark datasets. The BM25 effectiveness can vary due to different
hyperparameters and different linguistic processing methods used in different system implementations,
such as removing stop words, stemming, and tokenization. Furthermore, researchers want to contrast their proposed ranking
approach with a baseline ranking method. It could be tempting to
report a weak BM25 baseline, which makes the proposed ranking method
stand out better.

Several serving systems implement BM25 scoring, including Vespa.
Vespa’s lexical or sparse retrieval is also accelerated using the
weakAnd Vespa query
operator.
This is important because implementing a BM25 scoring function in
a system is trivial, but scoring all documents that contains at
least one of the query terms approaches linear complexity. Dynamic
pruning algorithms like weakAnd improve the
retrieval efficiency significantly compared to naive
brute-force implementations that scores all documents matching any of the query terms.

BM25 has two
hyperparameters, k1 and b, which impact ranking effectiveness.
Additionally, most (14 out of 18) of the BEIR datasets have both
title and text document fields, which in a real-production environment
would be the first thing that a seasoned search practitioner would
tune the relative importance of. In our BM25 baseline,
we configure Vespa to independently calculate the
BM25 score of both
title and text, and we combine the two BM25 scores
linearly. The complete Vespa rank
profile is given below.

rank-profile bm25 inherits default {
   first-phase {
      expression: bm25(title) + bm25(text)
   }
   rank-properties {
      bm25(title).k1: 0.9
      bm25(title).b: 0.4
      bm25(text).k1: 0.9
      bm25(text).b: 0.4
   }
}

We modify the BM25 k1 and b parameters but use the same parameters for
both fields. The values align with Anserini
defaults
(k1=0.9, b=0.4).

The following table reports nDCG@10 scores on a subset (13) of the
BEIR benchmark datasets. We exclude the four datasets that are not
publicly available. We also exclude the BEIR CQADupStack dataset
because it consists of 12 sub-datasets where the overall nDCG@10
score is found by averaging each
sub-dataset’s
nDCG@10 score. Adding these sub-datasets would significantly increase
the evaluation effort.

The table summarizes the BM25 nDCG@10 results. Vespa BM25 versus
BM25 from BEIR paper.

The table above demonstrates that the Vespa implementation has set a new high
standard, outperforming reported BM25 baselines on the BEIR benchmark.

Evaluating Vespa ranking models in a zero-shot setting

With the new strong BM25 baseline established in the above section, we
will now introduce two neural ranking models and compare their performance with the baseline.

Vespa ColBERT

We have previously described the Vespa ColBERT implementation in
this blog
post,
and we use the same model
weights in this
work. The Vespa ColBERT model is based on a distilled 6-layer MiniLM
model with 22M parameters, using quantized int8 weights (post-training
quantization). The model uses only 32 vector dimensions per query
and document wordpiece),
in contrast to the original ColBERT model, which
uses 128 dimensions. Furthermore, we use Vespa’s support for
bfloat16
to reduce the per-dimension storage usage from 4 bytes per dimension
with float to 2 bytes with bfloat16. We configure the maximum query
length to 32 wordpieces, and maximum document length to 180
wordpieces. Both maximum length parameters align with the training and experiments
on MS MARCO.

The ColBERT MaxSim scoring is implemented as a re-ranking model
using Vespa phased ranking,
re-ranking the top 2K hits ranked by BM25. We also compute and store
the title term embeddings for datasets with titles, meaning we have
two MaxSim scores for datasets with titles. We use a linear combination
to combine the title and text MaxSim scores.

The complete Vespa rank profile
is given below.

rank-profile colbert inherits bm25 {
   inputs {
      query(qt) tensor<float>(qt{}, x[32])
      query(title_weight): 0.5
   }
   second-phase {
      rerank-count: 2000
	   expression {
	   (1 - query(title_weight))* sum(
	    reduce(
	      sum(
		      query(qt) * cell_cast(attribute(dt), float), x
	      ),
	      max, dt
	    ),
	    qt
	   ) +
	   query(title_weight) * sum(
	    reduce(
	      sum(
		      query(qt) * cell_cast(attribute(title_dt), float), x
	      ),
	      max, dt
	    ),
	    qt
	  )
	}
}

The per wordpiece ColBERT vectors are stored in Vespa using Vespa’s
support for storing and computing over
tensors.

Note: Users
can also trade efficiency versus cost by storing the tensors on
disk, or in-memory using
paging
options. Paging is highly efficient in a re-ranking pipeline, as
just a few K tensors values are potentially paged on-demand.

Vespa Hybrid ColBERT + BM25

There are several ways to combine the ColBERT MaxSim with BM25,
including reciprocal rank
fusion(RRF)
which does not consider model scores, just the ordering (ranking) the
scores produce. Quote from Reciprocal Rank Fusion outperforms Condorcet and
individual Rank Learning Methods:

RRF is simpler and more effective than Condorcet Fuse, while sharing
the valuable property that it combines ranks without regard to the
arbitrary scores returned by particular ranking methods

Another approach is to combine the model scores into a new score
to produce a new ranking. We use a linear combination in this work
to compute the hybrid score. Like the ColBERT-only model, we use
BM25 as the first-phase ranking model and only calculate the hybrid
score for the global top-ranking K documents from the BM25 model.

Before combining the scores, we want to normalize both the unbound
BM25 and the bound ColBERT score. Normalization is accomplished by
simple max-min
scaling of the scores. With max-min scaling, scores from any ranking
model are scaled from 0 to 1. This makes it easier to combine the
two using relative weighting.

Since scoring in a production serving system might be spread across
multiple nodes, each node involved in the query will not know the
global max scores. We solve this problem by letting Vespa content
nodes involved in the query return both scores using Vespa
match-features.

A custom searcher
is injected in the query dispatching stateless Vespa service. This
searcher calculates the max and min for both model scores using
match features for hits within the window of global top-k hits
ranked by BM25. As with the ColBERT rank profile, we use a re-ranking
window of 2000 hits, but we perform feature-score scaling and
re-ranking in a stateless custom searcher instead of on the content
nodes.

The complete Vespa rank profile is given below. Notice the
match-features, which are returned with each hit to the stateless
searcher
(implementation),
which performs the normalization and re-scoring. The first-phase
scoring function is inherited from the previously described bm25
rank profile.

rank-profile hybrid-colbert inherits bm25 {
   function bm25() {
	   expression: bm25(title) + bm25(text)
   }

   function colbert_maxsim() {
	   expression {
	      2*sum(
	         reduce(
	            sum(
		            query(qt) * cell_cast(attribute(dt), float) , x
	            ),
	         max, dt
	         ),
	         qt
	      ) +
	      sum(
	         reduce(
	            sum(
		            query(qt) * cell_cast(attribute(title_dt), float), x
	            ),
	         max, dt
	         ),
	         qt
	      )
      }
   }
   match-features {
	   bm25
	   colbert_maxsim
   }
}

​​Results and analysis

As with the BM25 baseline model, we index one of the BEIR datasets
at a time on a Vespa instance and evaluate the models. The following
table summarizes the results. All numbers are nDCG@10. The
best-performing model score per dataset is in bold.

The table summarizes the nDCG@10 results per dataset. Note that MS MARCO is in-domain for ColBERT and Hybrid.
Average nDCG@10 is only computed for zero-shot and out-of-domain datasets.

As shown in the table above, in a in-domain setting on MS MARCO,
the Vespa ColBERT model outperforms the BM25
baseline significantly. The resulting nDCG@10 score aligns with reported MRR@10
results from previous work using
ColBERT
in-domain on MS MARCO. However, mixing the baseline BM25 using the hybrid model on MS MARCO evaluation
hurts the nDCG@10 score, as we combine two models where the unsupervised BM25
model is significantly weaker than the ColBERT model.

The Vespa ColBERT model underperforms BM25 on out-of-domain datasets,
especially CLIMATE-FEVER. The CLIMATE-FEVER dataset has very long
queries (avg 20.2 words). The long questions challenge the ColBERT
model, configured with a max query length of 32 wordpieces in the
experimental setup. Additionally, the Vespa ColBERT model underperforms
reported results for the full-sized ColBERT
V2 model using 110M parameters
and 128 dimensions. This result could indicate that the compressed
(in the number of dimensions) and model distillation have a more
significant negative impact when applied in a zero-shot setting
compared to in-domain.

These exceptions aside, the data shows that the unique hybrid Vespa ColBERT and BM25 combination is highly
effective, performing the best on 12 of 13 datasets. Its average
nDCG@10 score improves from 0.453 to 0.481 compared to the strong
Vespa BM25 baseline.

To reproduce the results of this benchmark,
follow the open-sourced
instructions.

Comparing hybrid zero-shot with few-shot methods

To compare the hybrid Vespa ranking performance with other models,
we include the results reported in Promptagator: Few-shot Dense
Retrieval From 8 Examples from
Google Research.

Generating synthetic training data in-domain via prompting LLMs is a recent
emerging Information Retrieval(IR) trend also described in
InPars: Data Augmentation for Information Retrieval using Large Language
Models.

The basic idea is to “prompt” a large language model (LLM) to generate synthetic queries
for use in training of in-domain ranking models. A typical prompt include a few
examples of queries and relevant documents, then the LLM is “asked”
to generate syntetic queries for many of the documents in the corpus.
The generated syntetic query, document pairs can be used to train neural ranking models.
We include a quote describing the approach from the Promptagator paper:

Photo by Maxime VALCARCE on Unsplash

This blog post explores using large language models (LLMs) to
generate labeled data for training ranking models. Distilling the
knowledge and power of generative models with billions of parameters
to ranking models with a few million parameters. The approach uses
a handful of human-annotated labeled examples (few-shot) and prompts
the LLM to generate synthetic queries for documents in the corpus.

The ability to create high-quality synthetic training data might
be a turning point with the potential to revolutionize information
retrieval. With a handful of human annotations, the LLMs can generate
infinite amounts of high-quality labeled data at a low cost. Training data which
is used to train much smaller and compute efficient ranking models.

Introduction

Language models built on the
Transformer architecture have
revolutionized text ranking overnight,
advancing the state-of-the-art by more than 30% on the MS MARCO
relevance dataset. However, Transformer-based ranking models need
significant amounts of labeled data and supervision to realize their
potential. Obtaining high-quality annotated data for training deep
ranking models is labor-intensive and costly. Therefore, many
organizations try to overcome the labeling cost problem by using
pseudo-labels derived from click models. Click models use query-document
user interaction from previously seen queries to label documents.
Unfortunately, ranking models trained on click-data suffer from
multiple bias issues, such as presentation bias and survivorship
bias towards the existing ranking model.

In addition, what if you want to build a great search experience
in a new domain without any interaction data (cold-start) or resources
to obtain sufficient amounts of labeled data to train neural ranking
models? The answer might be generative large language models (LLMs),
which can generate training data to train a ranking model adapted
to the domain and use case.

Generative large language models (LLMs)

The public interest in generative language models (LLMs) has
skyrocketed since OpenAI released ChatGPT in November
2022. The
GPT-3 model is trained on a
massive amount of text data using unsupervised learning and can
generate human-like text given a text prompt input.

Google is another leader in the language model space, except they
have not exposed any of them in a public chat-like interface like
OpenAI. In Scaling Instruction-Finetuned Language
Models, researchers from Google
describe a few of their generative language models and instruction
fine-tuning. In contrast to OpenAI and many other language model
providers, Google has open-sourced their generative FLAN-T5 models
in various model
sizes,
up to 11B parameters, using a permissive Apache 2.0 License.

Figure from Scaling Instruction-Finetuned Language
Models

A critical difference between large generative language models
and a vanilla BERT ranking model is that they necessarily don’t
require task-specific fine-tuning of the model weights. Massive
self-supervised training on piles of text, coupled with later
fine-tuning on a broad, diverse set of tasks, is one of the reasons
they are called foundation models
(FM).

The foundation model weights are frozen, but we can adapt the model
to our task by mixing natural language instructions with data in
the prompt input. The art of mixing instructions and data in an LLM
instruction prompt has created a new artistic engineering field;
prompt engineering.
We can improve the model’s generated output using prompt engineering
by changing the instructions written in natural language.

Generating labeled data via instruction-prompting Large Language Models

Instruction-prompting large language models (LLMs) have also entered
the information retrieval research (IR). A recent trend in IR
research is to use generative LLMs, such as GPT-3, to generate
synthetic data to train ranking models .

The general idea is to design an instruction prompt with a few
labeled relevance examples fed to the LLM (large language model)
to generate synthetic queries or documents. The instruction prompts
that generate artificial questions are the most promising direction
since all you need is a few labeled queries, document examples, and
samples from the document corpus. In addition, with synthetic query
generators, you avoid running inference with computationally expensive
LLMs at user time, which can be unrealistic for most
organizations. Instead, the
synthetic generation process is performed offline with LLM-powered
query generators. Offline inference with LLMs is considerably less
engineering-intensive than online usage, as inference latency is
not a concern.

LLMs will hallucinate and sometimes produce fake queries that are
too generic or irrelevant. To overcome this, researchers
use a ranking model (RM) to test the query quality. One can, for example, rank
documents from the corpus for each generated synthetic query using
the RM. The synthetic query is only retained for model training if
the source document ranks highly. This query consistency check
grounds the LLM output and improves the training data. Once the
synthetic queries and positive (relevant) document pairs are filtered,
one can sample negative (potentially irrelevant) examples and train
a ranking model adapted to the domain, the characteristics of the
document corpus, and the few-show query examples.

Illustration of distilling the knowledge and power of generative Large Language Models (LLMs) with billions of parameters to ranking models with a few million parameters.

Experiments

In the previous posts on zero-shot ranking
experiments,
we evaluated zero-shot ranking models on 13 BEIR
benchmark datasets. We reported
large gains in ranking effectiveness by a hybrid ranking model that
combines unsupervised BM25 ranking with a neural ranking model
trained on a large relevance dataset. This hybrid model is an example
of a zero-shot ranking model without any fine-tuning or adaption
to the new domain.

In this work, we want to see if we can improve the effectiveness
by using a large language model to create synthetic training data
for the new domain. We focus on one of the BEIR datasets, the
bio-medical trec-covid IR dataset.
We chose this dataset because we have previously built a simple
demo search UI and Vespa
app, indexing the
final CORD-19 dataset. With this demo, we can deploy the ranking
models to a production app in Vespa cloud. The following subsections
describe the experimental setup. We also publish three
notebooks
demonstrating the 3-stage process.

Synthetic query generation

We chose the open-source 3 Billion
flan-t5-xl generative
model as our artificial query generator. The model is genuine
open-source, licensed with a permissive Apache 2.0 license. We
devise the following few-shot instruction prompt template.

These are examples of queries with sample relevant documents for
each query. The query must be specific and detailed.

Example 1:
document: $document_example_1
query: $query_example_1

Example 2:
document: #document_example_2
query: $query_example_2

Example 3:
document: $document_example_3
query: $query_example

Example 4:
document: $input_document
query:

Our first prompt attempt did not include “the query must be specific
and detailed” phrase. Without it, many eyeballed queries were too
generic. Changing the prompt made the model produce more specific
queries. The change in output quality is an example of the magic of prompt engineering.

We use the three first trec-covid test queries (originally from
https://ir.nist.gov/trec-covid/data/topics-rnd5.xml, which is not available anymore)
as our in-domain examples for few-shot instruction examples.
We pick the first document annotated
as highly relevant to form the complete query-document example.

Finally, we iterate over the document collection, replace the
$input_document variable with a concatenation of the title and
abstract, then run an inference with the flan-t5-xl model and store the generated
query. We use a single A100 40GB GPU, which costs about 1$/hour and can
generate about 3600 synthetic queries per hour (depending on prompt size).

At completion, we ended up with synthetic queries for 33,099 documents
out of 171K docs. Notice that the three query-document examples are
the same for all document-to-query creations and that the model’s
max input sequence length limits the number of examples we can fit
into the prompt.

Query consistency checking

We use a robust zero-shot hybrid ranking
model
for query consistency checking. The generated query is retained for
training only if the source document is ranked #1 by the zero-shot
model. If the question passes this test, we also sample two negative
query-document pairs from the top 100 documents ranked by the
zero-shot model. After the consistency filter, the number of positive
query, document pairs drops to 14,156 (43% retention). The high retention
percentage demonstrates that the flan-t5-xl and prompt combination is creating
specific and detailed queries.

Dataframe with the generated synthetic queries and the document
metadata. The document abstract is summarized by the query contextual
Vespa dynamic
summary
feature. There are three rows in the data frame for each unique
generated question, one positive (relevant) document and two
irrelevant. This is the input to the next step, which is to train
a ranking model on this purely synthetic data.

Rank model training

After query generation and consistency checking, we use the synthetic
query and document pairs to train a ranking model. As in previous
work,
we use a cross-encoder based on a 6-layer MiniLM model with just
22M trainable parameters. We train the model for two epochs on the
synthetic data and export the model to ONNX for inference in
Vespa.

Finally, we deploy the fine-tuned model as a re-ranking phase
on top of the top 30 results from the hybrid model.

Evaluation & Results

We contrast the model tuned by synthetic queries with ranking models
evaluated in the zero-shot ranking blog
post
on the trec-covid dataset.

We gain four nDCG@10 points over the strong hybrid zero-shot model
and 10 nDCG@10 points over unsupervised BM25. These are significant
gains, especially given the low training and inference costs. We
also note that the Vespa BM25 baseline is strong, beating other
BM25 implementations on the same dataset. The model trained on
synthetic data outperforms the
PROMPTAGTOR model, which uses
a proprietary 137B FLAN checkpoint to generate synthetic queries.
In the paper they report a nDCG@10 of 76.2 on trec-covid. Finally,
we contrast with OpenAI’s GPT embedding
paper, where OpenAI reporst a nDCG@10 score of 64.9 for their GPT
XL embeddings on trec-covid.

Deploying to production

There are two reasons for choosing a cross-encoder model over a
bi-encoder for our synthetic fueled ranking model.

• Cross-encoders are generally more effective than bi-encoders.
• Updating the cross-encoder model weight does not require re-processing
  the document corpus. With bi-encoders using single vector
  representations, developers would have to re-process the document-side
  embeddings every time a new prompt-trained ranking model is available.
  With a cross-encoder, model versioning and A/B testing are easier
  to operationalize.

Cross-encoder’s downside is the computational complexity at query
time, which is quadratic with model sequence input length. We deploy
a trick to reduce the model input sequence; we input the query,
title, and a query contextual Vespa dynamic
summary
of the abstract. The dynamic abstract summarization reduces the
sequence length while retaining segments of the abstract that matches
the query. Furthermore, we limit the complexity by reducing the
re-ranking depth to 30. Finally, we deploy the trained model to our
https://cord19.vespa.ai/ demo site,
where users can choose between the prompt-generated model or the
previously described zero-shot ranking
models.

Conclusion

Synthetic query generation via instruction-prompted LLMs is a
promising approach for overcoming the label shortage problem. With
just three human labeled examples, the query generator, built on
an open-source flan-t5 model, could generate high-quality training
data.

As part of this work, we open-source three notebooks:

In addition to these resources; we open-source the generated
synthetic
queries
and the consistency-checked training data
(trec_covid_train_data_k1.parquet).
The training data includes the zero-shot scores, the consistenty checked query, the document title, and the
query contextual summarization of the abstract. The end-to-end Vespa
application is also
open-sourced.

In future work, we’ll look at how generative models can be used for
re-ranking, summarization, and generative question answering with Vespa.

Regardless, improving retrieval quality is step one in improving the
overall effectiveness of any retrieval-augmented system.

References