Table Question Answering

👍

Tutorial

Checkout our Open-Domain QA on Tables tutorial for a hands on guide on how to build your own system.

Indexing Data

To index data, cast your tables into pandas DataFrame format and use them to initialize Document objects. Then write them into your document store using the write_documents() method.

from haystack import Document
import pandas as pd

table = pd.DataFrame()
docs = [
    Document(content=table,
             content_type="table",
             meta={}),
    ...
]
document_store.write_documents(docs)

If you have your tables embedded in raw files (e.g. within a PDF), you can use the AzureConverter to parse them into the required format (see File Converters).

Retrieval

The TableTextRetriever is designed to perform document retrieval on both text and tabular Documents. It is a tri-encoder model with a separate encoder for the query, text passage and table. To learn more about the design of this component and also the training of the default models, have a look at Multi-modal Retrieval of Tables and Texts Using Tri-encoder Models which was accepted at EMNLP 2021.

retriever = TableTextRetriever(
    document_store=document_store,
    query_embedding_model="deepset/bert-small-mm_retrieval-question_encoder",
    passage_embedding_model="deepset/bert-small-mm_retrieval-passage_encoder",
    table_embedding_model="deepset/bert-small-mm_retrieval-table_encoder",
    embed_meta_fields=["title", "section_title"]
)
document_store.update_embeddings(retriever=retriever)

TableReader

With the TableReader, you can get Answers to your questions even if the answer is buried in a table. It is designed to use the TAPAS model created by Google. These models are able to pick out single cells, or a set of cells and perform an aggregation operating to form a final answer.

reader = TableReader(model_name_or_path="google/tapas-base-finetuned-wtq", max_seq_len=512)

🚧

Deprecation Note

Starting with 1.18, linearized offsets will be deprecated. Offsets that use the row and column indices of the table cell will be used instead.

👍

Note

• When the Answers are returned by the TableReader, the offsets indicate the cells that answer the question. These offsets are start and end indices when counting through the table in a linear fashion which means the first cell is top left and the last cell is bottom right.
  Starting with 1.16, you can enable offsets that use the row and column indices of the table that answer the question, instead of the linearized offsets. To do this, pass return_table_cell=True in the TableReader init() method.
• In the Answer's meta field, you can find the aggregation operator used to combine a set of cells into a final answer.

Pipeline

All the table QA components can be combined together in a pipeline.

table_qa_pipeline = Pipeline()
table_qa_pipeline.add_node(component=retriever, name="TableTextRetriever", inputs=["Query"])
table_qa_pipeline.add_node(component=reader, name="TableReader", inputs=["TableTextRetriever"])

You can run queries on this pipeline as follows.

prediction = table_qa_pipeline.run("How many twin buildings are under construction?")

The output of the pipeline is compatible with the print_answers() utility function

from haystack.utils import print_answers

print_answers(prediction, details="minimal")