numeric nlg - template changes (#1041)

prepare/cards/numeric_nlg.py

@@ -12,13 +12,20 @@
 card = TaskCard(
     loader=LoadHF(path="kasnerz/numericnlg"),  # TODO: load from github repo
     preprocess_steps=[
-        Set(fields={"type_of_input": "table", "type_of_output": "description"}),
+        Set(
+            fields={
+                "type_of_input_a": "table",
+                "type_of_input_b": "caption",
+                "type_of_output": "description",
+            }
+        ),
         MapHTMLTableToJSON(field="table_html_clean", to_field="table_out"),
-        SerializeTableAsMarkdown(field="table_out", to_field="input"),
+        SerializeTableAsMarkdown(field="table_out", to_field="input_a"),
         RenameFields(field="description", to_field="output"),
+        RenameFields(field="caption", to_field="input_b"),
     ],
-    task="tasks.generation[metrics=[metrics.bleu,metrics.rouge,metrics.bert_score.bert_base_uncased,metrics.meteor]]",
-    templates="templates.generation.all",
+    task="tasks.generation.from_pair",
+    templates="templates.generation.from_pair.all",
     __description__="NumericNLG is a dataset for numerical table-to-text generation using pairs of a table and a paragraph of a table description with richer inference from scientific papers.",
     __tags__={
         "modality": "table",

prepare/metrics/normalized_sacrebleu.py

@@ -28,6 +28,7 @@
 
 metric = MetricPipeline(
     main_score="sacrebleu",
+    prediction_type="str",
     preprocess_steps=[
         Copy(
             field="task_data/target_language",

prepare/tasks/generation.py

@@ -13,3 +13,27 @@
     "tasks.generation",
     overwrite=True,
 )
+
+add_to_catalog(
+    Task(
+        input_fields={
+            "input_a": "str",
+            "type_of_input_a": "str",
+            "input_b": "str",
+            "type_of_input_b": "str",
+            "type_of_output": "str",
+        },
+        reference_fields={"output": "str"},
+        prediction_type="str",
+        metrics=[
+            "metrics.bleu",
+            "metrics.rouge",
+            "metrics.bert_score.bert_base_uncased",
+            "metrics.meteor",
+        ],
+        augmentable_inputs=["input_a", "input_b"],
+        defaults={"type_of_output": "Text"},
+    ),
+    "tasks.generation.from_pair",
+    overwrite=True,
+)

prepare/templates/generation/generation.py

@@ -30,3 +30,23 @@
     "templates.generation.all",
     overwrite=True,
 )
+
+add_to_catalog(
+    InputOutputTemplate(
+        input_format="Given the following {type_of_input_a} and {type_of_input_b}, generate the corresponding {type_of_output}."
+        "\n{type_of_input_a}: \n{input_a} \n{type_of_input_b}: \n{input_b} \n{type_of_output}:",
+        output_format="{output}",
+        postprocessors=[
+            "processors.take_first_non_empty_line",
+            "processors.lower_case_till_punc",
+        ],
+    ),
+    "templates.generation.from_pair.default",
+    overwrite=True,
+)
+
+add_to_catalog(
+    TemplatesList(["templates.generation.from_pair.default"]),
+    "templates.generation.from_pair.all",
+    overwrite=True,
+)

src/unitxt/catalog/cards/numeric_nlg.json

@@ -8,7 +8,8 @@
         {
             "__type__": "set",
             "fields": {
-                "type_of_input": "table",
+                "type_of_input_a": "table",
+                "type_of_input_b": "caption",
                 "type_of_output": "description"
             }
         },
@@ -20,16 +21,21 @@
         {
             "__type__": "serialize_table_as_markdown",
             "field": "table_out",
-            "to_field": "input"
+            "to_field": "input_a"
         },
         {
             "__type__": "rename_fields",
             "field": "description",
             "to_field": "output"
+        },
+        {
+            "__type__": "rename_fields",
+            "field": "caption",
+            "to_field": "input_b"
         }
     ],
-    "task": "tasks.generation[metrics=[metrics.bleu,metrics.rouge,metrics.bert_score.bert_base_uncased,metrics.meteor]]",
-    "templates": "templates.generation.all",
+    "task": "tasks.generation.from_pair",
+    "templates": "templates.generation.from_pair.all",
     "__description__": "NumericNLG is a dataset for numerical table-to-text generation using pairs of a table and a paragraph of a table description with richer inference from scientific papers.",
     "__tags__": {
         "modality": "table",

src/unitxt/catalog/metrics/normalized_sacrebleu.json

@@ -1,6 +1,7 @@
 {
     "__type__": "metric_pipeline",
     "main_score": "sacrebleu",
+    "prediction_type": "str",
     "preprocess_steps": [
         {
             "__type__": "copy",

src/unitxt/catalog/tasks/generation/from_pair.json

@@ -0,0 +1,27 @@
+{
+    "__type__": "task",
+    "input_fields": {
+        "input_a": "str",
+        "type_of_input_a": "str",
+        "input_b": "str",
+        "type_of_input_b": "str",
+        "type_of_output": "str"
+    },
+    "reference_fields": {
+        "output": "str"
+    },
+    "prediction_type": "str",
+    "metrics": [
+        "metrics.bleu",
+        "metrics.rouge",
+        "metrics.bert_score.bert_base_uncased",
+        "metrics.meteor"
+    ],
+    "augmentable_inputs": [
+        "input_a",
+        "input_b"
+    ],
+    "defaults": {
+        "type_of_output": "Text"
+    }
+}

src/unitxt/catalog/templates/generation/from_pair/all.json

@@ -0,0 +1,6 @@
+{
+    "__type__": "templates_list",
+    "items": [
+        "templates.generation.from_pair.default"
+    ]
+}

src/unitxt/catalog/templates/generation/from_pair/default.json

@@ -0,0 +1,9 @@
+{
+    "__type__": "input_output_template",
+    "input_format": "Given the following {type_of_input_a} and {type_of_input_b}, generate the corresponding {type_of_output}.\n{type_of_input_a}: \n{input_a} \n{type_of_input_b}: \n{input_b} \n{type_of_output}:",
+    "output_format": "{output}",
+    "postprocessors": [
+        "processors.take_first_non_empty_line",
+        "processors.lower_case_till_punc"
+    ]
+}