The Retriever is a lightweight filter that can quickly go through the full document store and pass a set of candidate documents to the Reader. It is an tool for sifting out the obvious negative cases, saving the Reader from doing more work than it needs to and speeding up the querying process.
Dense Passage Retrieval (dense)
Note that not all Retrievers can be paired with every DocumentStore. Here are the combinations which are supported:
See Optimization for suggestions on how to choose top-k values.
TF-IDF is a commonly used baseline for information retrieval that exploits two key intuitions:
documents that have more lexical overlap with the query are more likely to be relevant
words that occur in fewer documents are more significant than words that occur in many documents
Given a query, a tf-idf score is computed for each document as follows:
score = tf * idf
tfis how many times words in the query occur in that document.
idfis the inverse of the fraction of documents containing the word.
In practice, both terms are usually log normalised. If you’d like to learn more about the exact details of the algorithm, have a look at this video.
document_store = InMemoryDocumentStore()...retriever = TfidfRetriever(document_store)...finder = Finder(reader, retriever)
BM25 is a variant of TF-IDF that we recommend you use if you are looking for a retrieval method that does not need a neural network for indexing. It improves upon its predecessor in two main aspects:
tfafter a set number of occurrences of the given term in the document
It normalises by document length so that short documents are favoured over long documents if they have the same amount of word overlap with the query
document_store = ElasticsearchDocumentStore()...retriever = ElasticsearchRetriever(document_store)...finder = Finder(reader, retriever)
See this blog post for more details about the algorithm.
Dense Passage Retrieval (Recommended)
Dense Passage Retrieval is a highly performant retrieval method that calculates relevance using dense representations. Key features:
One BERT base model to encode documents
One Bert base model to encode queries
Ranking of documents done by dot product similarity between query and document embeddings
Indexing using DPR is comparatively expensive in terms of required computation since all documents in the database need to be processed through the transformer. The embeddings that are created in this step can be stored in FAISS, a database optimized for vector similarity. DPR can also work with the ElasticsearchDocumentStore or the InMemoryDocumentStore.
When using DPR, it is recommended that you use the dot product similarity function since that is how it is trained.
To do so, simply provide
similarity="dot_product" when initializing the DocumentStore
as is done in the code example below.
There are two design decisions that have made DPR particularly performant.
Separate encoders for document and query helps since queries are much shorter than documents
Training with ‘In-batch negatives’ (gold labels are treated as negative examples for other samples in same batch) is highly efficient
In Haystack, you can simply download the pretrained encoders needed to start using DPR. If you’d like to learn how to set up a DPR based system, have a look at our tutorials.
document_store = FAISSDocumentStore(similarity="dot_product")...retriever = DensePassageRetriever(document_store=document_store,query_embedding_model="facebook/dpr-question_encoder-single-nq-base",passage_embedding_model="facebook/dpr-ctx_encoder-single-nq-base")...finder = Finder(reader, retriever)
Tip: Haystack supports training of your own DPR model! Tutorial coming soon!
In Haystack, you also have the option of using a single transformer model to encode document and query. One style of model that is suited to this kind of retrieval is that of Sentence Transformers. These models are trained in Siamese Networks and use triplet loss such that they learn to embed similar sentences near to each other in a shared embedding space.
They are particular suited to cases where your query input is similar in style to that of the documents in your database i.e. when you are searching for most similar documents. This is not inherently suited to query based search where the length, language and format of the query usually significantly differs from the searched for text.
When using Sentence Transformer models, we recommend that you use a cosine similarity function.
To do so, simply provide
similarity="cosine" when initializing the DocumentStore
as is done in the code example below.
document_store = ElasticsearchDocumentStore(similarity="cosine")...retriever = EmbeddingRetriever(document_store=document_store,embedding_model="deepset/sentence_bert")...finder = Finder(reader, retriever)
Deeper Dive: Dense vs Sparse
Broadly speaking, retrieval methods can be split into two categories: dense and sparse.
Sparse methods, like TF-IDF and BM25, operate by looking for shared keywords between the document and query. They are:
simple but effective
don’t need to be trained
work on any language
More recently, dense approaches such as Dense Passage Retrieval (DPR) have shown even better performance than their sparse counter parts. These methods embed both document and query into a shared embedding space using deep neural networks and the top candidates are the nearest neighbour documents to the query. They are:
powerful but computationally more expensive especially during indexing
trained using labelled datasets
Between these two types there are also some qualitative differences too. For example, sparse methods treat text as a bag-of-words meaning that they do not take word order and syntax into account, while the latest generation of dense methods use transformer based encoders which are designed to be sensitive to these factors.
Also dense methods are very capable of building strong semantic representations of text, but they struggle when encountering out-of-vocabulary words such as new names. By contrast, sparse methods don’t need to learn representations of words, they only care about whether they are present or absent in the text. As such, they handle out-of-vocabulary words with no problem.
Dense methods perform indexing by processing all the documents through a neural network and storing the resulting vectors. This is a much more expensive operation than the creation of the inverted-index in sparse methods and will require significant computational power and time.
The terms dense and sparse refer to the representations that the algorithms build for each document and query. Sparse methods characterise texts using vectors with one dimension corresponding to each word in the vocabulary. Dimensions will be zero if the word is absent and non-zero if it is present. Since most documents contain only a small subset of the full vocabulary, these vectors are considered sparse since non-zero values are few and far between.
Dense methods, by contrast, pass text as input into neural network encoders and represent text in a vector of a manually defined size (usually 768). Though individual dimensions are not mapped to any corresponding vocabulary or linguistic feature, each dimension encodes some information about the text. There are rarely 0s in these vectors hence their relative density.