Open-domain question-answering has emerged as a benchmark for measuring a
system’s capability to read, represent, and retrieve general knowledge.
Retrieval-based question-answering systems require connecting various systems
and services, such as BM25 text search, vector similarity search, NLP model
serving, tokenizers, and middleware to glue all this together. Most of these
are core features of Vespa.ai. In this post, we reproduce the state-of-the-art
baseline for retrieval-based question-answering systems within a single,
scalable production ready application on Vespa.ai.

Introduction

Some of the most effective drivers of scientific progress are benchmarks.
Benchmarks provide a common goal, a purpose, for improving the state-of-the-art
on datasets that are available to everyone. Leaderboards additionally add a
competitive motivation, offering the opportunity to excel among peers. And
rather than just endlessly tinkering to improve relevant metrics, competitions
add deadlines that spur researchers to actually get things done.

Within the field of machine learning, benchmarks have been particularly
important to stimulate innovation and progress. A new competition, the
Efficient Open-Domain Question Answering challenge for NeurIPS 2020, seeks to
advance the state-of-the-art in question answering systems. The goal here is to
develop a system capable of answering questions without any topic restriction.
With all the recent progress in natural language processing, this area has
emerged as a benchmark for measuring a system’s capability to read, represent,
and retrieve general knowledge.

The current retrieval-based state-of-the-art is the Dense Passage Retrieval
system, as described in the Dense
Passage Retrieval for Open-Domain Question Answering
paper. It consists of a set of python
scripts, tools, and models developed primarily for research. There are a lot
of parts in such a system. These include two BERT-based models for encoding
text to embedding vectors, another BERT-based model for extracting answers,
approximate nearest-neighbor similarity search and text-based BM25 methods for
retrieving candidates, tokenizers, and so on. It’s not trivial to bring such a
system to production. We thought it would be interesting to consolidate these
different parts and demonstrate how to build an open-domain question-answering
serving system with Vespa.ai that achieves state-of-the-art accuracy.

Most of these components are core features in Vespa. A while ago, we improved
Vespa’s text search support for term-based retrieval and ranking. We recently
added efficient approximate nearest neighbors for semantic, dense vector
recall. For hybrid retrieval, Vespa supports many types of machine-learned
models, for instance neural networks and decision forests. We have also
improved our support for TensorFlow and PyTorch models to run larger NLP and
Transformer models.

This is interesting because while this has obvious benefits in a research
setting, such systems’ real value lies in their end-use in applications. Vespa
is designed as a highly performant and scalable production-ready system. Thus,
it offers a simplified path to deployment in production without coping with the
complexity of maintaining many different subsystems. That makes Vespa an
attractive package.

During this blog post, we’ll touch upon

• Fast approximate-nearest neighbors for semantic, dense vector retrieval.
• Term-based (BM25) retrieval for sparse vector retrieval.
• Importing of multiple pre-trained BERT-based models in Vespa for encoding embedding vectors and extracting answers.
• Custom logic for tokenization and other things.

For more details we refer to the companion sample
application.

This post’s goal is to recreate the Dense Passage Retrieval (DPR) paper results
for the Natural Questions
benchmark.
We’ll first go through a high-level overview of how a retrieval-based
question-answering system works in the context of this paper. Then we’ll show
how this can all be implemented on Vespa without any external services or
plugins, and recreate the paper’s state-of-the-art results as measured by the
exact match of answers given a set of questions. We’ll wrap up with a look to
the future and the next post in this series.

Background: the anatomy of a retrieval-based QA system

The Natural Questions benchmark consists of natural language questions and
answers. How to retrieve and represent the knowledge required to answer the
questions is up to each system. There are two main approaches to this:
retrieval and parametric. A retrieval-based system uses search terms or
semantic representation vectors to recall a set of sentences or paragraphs
subsequently evaluated with a machine-learned model to extract the exact
answer. A parametric system stores the answers more or less directly in the
weights of a large neural network model. There has also been research into
hybrid retrieval and parametric systems such as the Retrieval-Augmented
Generation system, which recently improved
the state-of-the-art for the natural question benchmark as a whole. In this
blog post, we’ll focus on a retrieval-based system, but will explore parametric
and hybrid approaches in later blog posts.

A retrieval-based question answering system typically stores its “knowledge” in
an information retrieval system. This can be sentences, paragraphs, or entire
documents. Here we’ll use the Wikipedia dataset where each page is split into
passages of 100 words each. The dataset contains 21 million such passages. When
answering a question, we first retrieve the passages that most likely include
the answer. They are then analyzed with a machine-learned model to extract the
spans that most likely results in the correct response. These stages are called
the “retriever” and “reader”, respectively.

The retriever

The retriever is responsible for generating a set of candidate passages. Since
the subsequent reader component is expensive to evaluate, it is crucial to have
an effective retrieval mechanism. There are two main approaches to passage
retrieval: term-based (sparse) such as for BM25, and embedding (dense) vectors,
which each have their strengths and weaknesses.

Term-based (sparse) retrieval

Term-based retrieval is the classic information retrieval method and covers
algorithms such as TF-IDF and BM25. Conceptually, the text is represented by a
vector where each dimension represents a term in a vocabulary. A non-zero value
signifies its presence. As each text only contains a subset of possible terms
in the vocabulary, these vectors are large and sparse. The similarity between
two texts, for instance, a document and a query, can be computed by a dot
product between the sparse vectors with slight modifications (e.g., term
importance) for TF-IDF or BM25. Term-based methods rely on inverted index
structures for efficient retrieval. This can in some cases be further
accelerated by algorithms such as WAND.

Except for any pre-processing such as lemmatization, stemming, and possibly
stop-word removal, terms are matched exactly as found in the text. This can be
a strength as well as a weakness. For salient terms, e.g. names and
places, this cuts down the search space significantly. However, potentially
relevant documents that don’t contain the exact term will not be retrieved
unless one uses query expansion or related techniques. The Dense Passage
Retrieval (DPR) paper uses ElasticSearch as the providing system for BM25.

Embedding-based (dense) retrieval

The number of potential terms in a vocabulary can be vast indeed. The basic
idea behind embedding vectors is to compress this high dimensional sparse
vector to a much smaller dense vector where most dimensions contain a non-zero
value. This has the effect of projecting a query or document vector into a
lower-dimensional space. This can be done so that vectors that are close
geometrically are also close semantically. The DPR paper uses two BERT models
to encode text: one for encoding queries and one for encoding documents. The
two models are trained simultaneously in a two-tower configuration to maximize
the dot product for passages likely to answer the question.

In contrast to the sparse representation, there are no exact methods for
finding the nearest neighbors efficiently. So we trade accuracy for efficiency
in what is called approximate nearest neighbors (ANN). Many different methods
for ANN search have been proposed. Some are compatible with inverted index
structures so they can be readily implemented in existing information retrieval
systems. Examples are k-means clustering, product quantization (and it’s
relatives), and locality sensitive hashing, where the centroids or buckets can
be indexed. A method that is not compatible with inverted indexes is
HNSW (hierarchical navigable small world).
HNSW is based on graph structures, is efficient, and has an attractive
property where the graph can be incrementally built at runtime. This is in
contrast to most other methods that require offline, batch oriented index
building.

Retrieval based on semantic embeddings complements term-based retrieval
well. Semantically similar documents can be recalled even though they don’t
contain the exact same terms. Unlike the bag-of-words approach for term-based
retrieval, word order can provide additional context. Historically, however,
term-based retrieval has outperformed semantic embeddings on question answering
problems, but the DPR paper shows that dense retrieval can be vastly improved
if the encoding has specifically been trained to the task. The DPR paper uses
FAISS with an HNSW index for
similarity search.

The reader

While the retriever component’s job is to produce a set of candidate passages
that hopefully contain the answer to the question, the reader extracts the
passages’ actual answer. This requires some form of natural language
understanding model, and BERT (or other Transformer) models are used. These
models are typically huge and expensive to evaluate, so only a small number of
candidate passages are run through them.

Transformer models take sequences of tokens as input. The tokenization of text
can be done in quite a few different ways to balance vocabulary size and
sequence length. Due to BERT models’ full attention mechanism, evaluation time
increases quadratically with sequence length. So a reasonable balance must be
struck, and BERT-based models use a WordPiece or similar algorithm to split
less common words into subwords.

The reader model’s input is the concatenation of the tokens representing the
question, the document’s title, and the passage itself. The model looks up the
embedding representation for each token, and through a series of layers,
produces a new representation for each token. These representations can then be
used for different tasks. For question-answering, an additional layer is added
to compute the three necessary outputs: the relevance of the passage to the
question, and the start and end indexes of the answer.

To extract the final answer, the passage that produced the largest relevance
score is used. The two other outputs of the model are probabilities for each
token of being a start token and an end token. The final answer is chosen by
finding the span with the largest sum of start probability and end probability.
This results in a sequence of tokens, which must be converted to words by the
tokenizer before returning. The DPR paper uses a BERT-based
model to
output span predictions.

Putting all this together

The retrieval-based question-answering system as described above, capable of
both term- and embedding-based retrieval, requires at least the following
components:

• A BM25 information retrieval system storing the 21 million Wikipedia text passages.
• An efficient vector similarity search system storing the passage embedding vectors.
• A model serving system for the three different BERT-based models: query encoder, document encoder, and reader model.
• A BERT-based tokenizer.
• Middleware to glue this all together.

The tokenizer generates the token sequence for the text. These tokens are
stored for usage in the reader model. They are also used in a document
BERT-base encoder model to create the embedding vector for dense retrieval. The
text and embedding vectors need to be indexed for fast retrieval.

A similar process is followed for each query. The tokenizer produces a token
sequence used to generate an embedding vector in the query BERT-based encoder.
The first stage, retrieval, is done using either term-based retrieval or
embedding based retrieval. The top-N passages are passed to the Reader model
and ranked accordingly. The best passage is analyzed to extract the best span
containing the answer.

This is a non-trivial list of services that need to be set up to implement a
question-answering system. In the next section we show how to implement all
this as a single Vespa application.

Reproducing the baseline on Vespa

Most of the components mentioned above have become core features in Vespa. In
the following we’ll present an overview of setting this up in Vespa. The
details can be seen in the companion sample
application.

Schema

When creating an application with Vespa, one typically starts with a document
schema. The schema contains, among other things, the definition of which data
should be stored with each document. In our case, each document is a passage
from the Wikipedia dataset. So we set up a schema that allows for the different
retrieval methods we have discussed:

• Sparse retrieval using traditional BM25 term-based retrieval.
• Dense retrieval using vector representations encoded by a trained model.
• Hybrid retrieval using a combination of the above.

We set up all of this in a single document schema:

schema wiki {
  document wiki {

    field title type string {
      indexing: summary | index
    }

    field text type string {
      indexing: summary | index
    }

    field title_token_ids type tensor(d0[256]) {
        indexing: summary | attribute
    }

    field text_token_ids type tensor(d0[256]) {
      indexing: summary | attribute
    }

    field text_embedding type tensor(x[769]) {
      indexing: attribute | index
      attribute {
        distance-metric: euclidean
      }
    }
  }
}

Here, each passage’s title and text content is represented both by a string and
a token sequence field. The string fields are indexed to support BM25 and
support WAND for accelerated retrieval. The token sequences are represented as
tensors and are used as inputs to the reader model.

The embedding vector of the title and text is precomputed and stored as a
tensor. This vector is used for dense retrieval, so we enable the HNSW
index on this
field for approximate nearest neighbor matching. A nice feature of Vespa
is that the HNSW index is not pre-built offline; it is constructed online as
data is indexed. This allows for applications that are much more responsive to
new data being fed into the system.

Retrieval and ranking

A query in Vespa defines, among other things, how Vespa should recall documents
(called the matching phase) and how Vespa should score the documents (called
the ranking phase). Vespa provides a rich query
API, where queries are
specified with the Vespa YQL
language.
As the different retrieval strategies (term-based and embedding-based) have
different query syntax, we have built a custom searcher component that allows
us to build a unified search interface and only pass the actual question and
retrieval strategy as parameters. This simplifies comparisons between the
methods.

The retrieval strategies differ both on what is recalled and how they are
scored. In Vespa, scoring is expressed using ranking expressions that are
configured in the document schema. Vespa supports multi-phased
ranking, and we
exploit that here so that the first phase represents the retriever, and the
second phase the reader. We set up the first phase ranking profile like this:

rank-profile sparse inherits openqa {
  first-phase {
    expression: bm25(text) + bm25(title)
  }
}
rank-profile dense inherits openqa {
  first-phase {
    expression: closeness(field, text_embedding)
  }
}
rank-profile hybrid inherits openqa {
  first-phase {
    expression: 1000*closeness(field, text_embedding) + bm25(text) + bm25(title)
  }
}

Here, we set up three ranking profiles, one for sparse retrieval, one for dense
retrieval, and one for hybrid retrieval. The sparse ranking profile uses the
BM25 score of the title and text fields against the query as the scoring
function. The dense profile uses the closeness, e.g., the euclidean distance
between the query and document embedding field. The hybrid profile is an
example that combines both for hybrid retrieval. This first phase represents
the retriever.

For the reader model we set up the base profile openqa which introduces a
second phase, common between the retrieval strategies:

onnx-model reader {
  file: files/reader.onnx
  input  input_ids: input_ids
  input  attention_mask: attention_mask
  output output_0: start_logits
  output output_1: end_logits
  output output_2: relevance_logits
}

rank-profile openqa {
  second-phase {
    rerank-count: 10
    expression: onnxModel(reader).relevance_logits
  }
  summary-features {
    onnxModel(reader).start_logits
    onnxModel(reader).end_logits
  }
}

Here, the top 10 documents are re-ranked by the reader model, defined be the
onnxModel rank feature. The actual model we use is a pre-trained
model the DPR team
has published on HuggingFace’s model repository. HuggingFace has released an
excellent transformer model
export, which makes it
easy to export Transformer models (from either PyTorch or TensorFlow) to ONNX
format. After the model is exported to ONNX, we can put the
model file in the application package and configure its use in the document
schema as seen above. To make this scale, Vespa distributes this model to all
content nodes in the cluster. Thus the model is evaluated on the content node
during ranking, avoiding transferring data to an external model service.

The reader model has three outputs, where the relevance_logits output is used
for scoring. The other two represent the probabilities of each token being a
start or end token, respectively, for the ultimate answer to the question.
These are picked up by a custom searcher, and the actual span these fields
represent is extracted there.

The application contains one additional model, the question encoder
model
used to generate the embedding vector for the query at run time. The
pre-computed document embeddings from the Wikipedia dataset are published by
Facebook Research. We use these embeddings directly.

Vespa container middleware – putting it all together

The application has the following custom plugins implemented as Java in Vespa containers:

• A BERT tokenizer component that is responsible for generating the sequence of
  BERT tokens from a text.
• A custom document processor which uses the BERT tokenization during indexing.
• A custom searcher that retrieves the embedding representation for the query
  invoking the BERT question encoder model.
• A custom searcher that controls the retrieval logic (e.g., sparse, dense, or
  hybrid).
• A custom searcher that reads the output of the reader model and extracts the
  best matching answer span and converts that token sequence to the actual text
  which is returned to the user.

The result is a service which takes a textual question and returns the
predicted answer, all within one Vespa application.

Results

The benchmark we use for this application is the Natural Questions data set.
All experiments in this section are run by querying the Vespa instance and
checking the predicted answer to the golden reference answer.

Retriever Accuracy Summary

We use Recall@position as the main evaluation metric for the retriever. The
obvious goal of the retriever is to have the highest recall possible at the
lowest possible position. Since the final top position passages are re-ranked
using the BERT-based reader, the fewer passages we need to evaluate the better
the run time complexity and performance.

The following table summarizes the retriever accuracy using the original 3,610
dev questions in the Natural Questions for Open Domain Question Answering tasks
(NQ-open.dev.jsonl).

The DPR paper reports recall@20 of 79.4, so our results are inline with their
reported results for the dense retrieval method. We attribute the slight
difference in our favor here for a different set of settings for the HNSW
index.

Reader Accuracy Summary

We evaluate the reader accuracy using the Exact Match (EM) metric. The Exact
Match metric measures the percentage of predictions that match any one of the
ground truth answers exactly. To get an EM score of 1 for a query the answer
prediction must match the golden answer exactly as given in the dataset. This
is challenging. For instance, the question “When was the last moon landing?”
has golden answers “14 December 1972 UTC” or “December 1972”, and the predicted
answer “14 December 1972” will be scored 0.

The results are for the same data set above, with a re-rank of the top 10
results from the retriever:

The above results reproduce the results of the DPR paper which at the writing
of this post is the current state of the art for retrieval-based systems.

Further work

In this blog post we’ve been focused on reproducing the results of the DPR
paper within a single instance of Vespa. In the next part of this series we
will introduce a more sophisticated hybrid model and show how we can use other
model types to hopefully improve accuracy. We will follow up with another post
on lowering total system latency down to more reasonable levels while measuring
any impact this has on accuracy.

Stay tuned!