package language.tapas

Get desktop application:
View/edit binary Protocol Buffers messages

• repeated Annotation annotations = 1

Used in: AnnotatedText

• optional int64 begin_byte_index = 1

  Indices refer to 'original_text' of a question or 'text' of a cell. Inclusive begin byte.

• optional int64 end_byte_index = 2

  Exclusive end byte.

• optional string identifier = 3

  An identifier for example a Wikipedia URL.

• map<string, string> descriptions = 1

  For each entity that appears in the interaction, the map has a textual description of the entity, like the first section of its Wikipedia page.

Used in: Question

• repeated AnswerCoordinate answer_coordinates = 1

  Coordinates of cells that contain the answers.

• optional Answer.AggregationFunction aggregation_function = 2

• repeated string answer_texts = 3

  Answers in text format.

• optional float float_value = 4

  Present if the answer can be represented as a single float value, for example produced by an aggregation ('the average population of all countries').

• optional bool is_valid = 5

  If true, this answer can be used to construct training/test examples. If false some error were triggered during parsing of this answer.

• optional int32 class_index = 6

  Present if the answer can be represented as a single integer value, for example when it's a classification or entailment task.

A function that is applied to the answer cells in order to obtain the final answer.

Used in: Answer

• NONE = 0

• SUM = 1

  Sums all cell values. Numeric cells only.

• AVERAGE = 2

  Averages all cell values. Numeric cells only.

• COUNT = 3

  Counts the number of answers.

Used in: Answer

• optional int32 row_index = 1

• optional int32 column_index = 2

Uses the average cosine similarity to score the tokens.

Used in: TablePruningModel

• optional bool use_positional_embeddings = 2

  Enables the use of positional embeddingins to compute the average cosine similarity.

• optional Loss loss = 3

  The loss used to learn the weights.

Used in: Cells, Table

• optional string text = 1

• optional NumericValue numeric_value = 2

Used in: Table

• repeated Cell cells = 1

Used in: NumericValue

• optional int32 year = 1

• optional int32 month = 2

• optional int32 day = 3

Select the k first tokens up to max_num_tokens. If max_num_tokens = tapas_max_num_tokens no table pruing is used.

Used in: TablePruningModel

(message has no fields)

An interaction represents a sequences of question answerable from a single table.

• optional string id = 1

• optional Table table = 2

• repeated Question questions = 3

The loss used to learn the weights.

Used in: AvgCosSimilarity, TAPAS

• oneof loss

  • Loss.Unsupervised unsupervised = 200

• optional Loss.HardSelection train = 400

• optional Loss.HardSelection eval = 500

• optional bool add_classification_loss = 600

  Enables the pruning model to use a loss similar to the tapas model.

The hard selection strategy used for train and/or for test.

Used in: Loss

• optional HardSelection.SelectionFn selection_fn = 1

Used in: HardSelection

• ALL = 0

  No hard selection is used. Select all the tokens.

• TOP_K = 1

  Selects the best tokens up to max_num_tokens. Returns the TOP_K scores.

• MASK_TOP_K = 2

  Selects the best tokens up to max_num_tokens. Returns the TOP_K mask values in 0,1.

Uses an unsupervised model to learn the required columns. The back probagation is always activated.

Used in: Loss

• optional Unsupervised.Regularization regularization = 5

Used in: Unsupervised

• NONE = 0

  No Regularization is used.

• L1 = 1

  Computes L1 over all the tokens scores.

• L2 = 2

  Computes L2 over all the tokens scores.

• L1_L2 = 3

  Computes l1 on tokens sequence then l2 on the batch.

Used in: NegativeRetrievalExamples

• optional Table table = 1

  One negative table

• optional NegativeRetrievalExample.Type type = 2

• optional int32 rank = 3

  If table was retrieved from Baseline its rank.

• optional float score = 4

  The similarity score between the negative table and the question. A positive score represent a high similarity.

Used in: NegativeRetrievalExample

• BASELINE = 1

• DOCUMENT = 2

• CORRUPTED = 3

• repeated NegativeRetrievalExample examples = 1

  The examples correspend to the negative tables.

Used in: Cell, NumericValueSpan

• oneof value

  • float float_value = 1

  • Date date = 2

Used in: NumericValueSpans

• optional int32 begin_index = 1

• optional int32 end_index = 2

• repeated NumericValue values = 3

Used in: Question

• repeated NumericValueSpan spans = 1

Used in: Interaction

• optional string id = 1

• optional string text = 2

  The question string after normalization.

• optional string original_text = 3

  The original raw question string.

• optional NumericValueSpans annotations = 4

  Numeric value spans in 'text'.

• optional Answer answer = 5

• repeated Answer alternative_answers = 6

• optional int32 rank = 1

• optional double score = 2

Uses a TAPAS model to score the columns or the tokens.

Used in: TablePruningModel

• optional TAPAS.Selection selection = 2

• optional string bert_config_file = 3

• optional string bert_init_checkpoint = 4

• optional bool reset_position_index_per_cell = 6

• optional Loss loss = 5

  The loss used to learn the weights.

Specifies the use of the columns scores or tokens scores.

Used in: TAPAS

• COLUMNS = 0

• TOKENS = 1

Represents a simple table with m rows and n columns.

Used in: Interaction, NegativeRetrievalExample

• repeated Cell columns = 1

  The names of the n columns.

• repeated Cells rows = 2

  m rows containing n cells each.

• optional string table_id = 3

  Some unique identifier of this table.

• optional string document_title = 4

  The title of the document the table appears in.

• optional string caption = 5

  Title or caption of the table.

• optional string document_url = 6

  The URL the table was found on.

• repeated string alternative_document_urls = 7

  Other versions of the same document that the table occurs on.

• repeated string alternative_table_ids = 8

  Other versions of the same table.

• optional string context_heading = 9

  Heading of the table on the document.

Options for table pruning models.

• optional int32 max_num_tokens = 1

• oneof table_pruning_model

  • AvgCosSimilarity avg_cos_similarity = 2

  • TAPAS tapas = 3

  • FirstTokens first_tokens = 4

• repeated TableSelection.TokenCoordinates selected_tokens = 3

• optional TableSelection.ModelPredictionStats model_prediction_stats = 2

• optional TableSelection.DebugInfo debug = 100

Used in: TableSelection

• repeated DebugInfo.Column columns = 1

Used in: DebugInfo

• optional int32 index = 1

  Index of the column.

• optional double score = 2

  The score assigned by some scorer.

• optional bool is_selected = 3

  True if the column was selected as relevant by some column selector.

• optional bool is_required = 4

  True if the column is needed to find the final answer (gold data).

Used in: TableSelection

• repeated ModelPredictionStatsPerColumn column_prediction_stats = 1

• repeated ModelPredictionStatsPerModel model_prediction_stats = 2

  Model predictions of the umodified inputs.

Model 2, 3 and 4 didn't answer the question correctly even when running on the whole table. When column 0 was removed from the input, model 1 answered the question correctly and model 5 incorrectly.

Used in: ModelPredictionStats

• optional int32 column = 3

  Column index.

• repeated ModelPredictionStatsPerModel model_prediction_stats = 2

Used in: ModelPredictionStats, ModelPredictionStatsPerColumn

• optional string model_id = 1

  Identifier of the model that produced this result.

• optional bool is_correct = 2

  Whether the model predictions is correct.

Tokens that should be added to the TF example. Must not be empty!

Used in: TableSelection

• optional int32 row_index = 1

  The header row has index 0, the first data row index 1.

• optional int32 column_index = 2

• optional int32 token_index = 3