From 815bfe68235d740c8d152330b7f05409921d7f9d Mon Sep 17 00:00:00 2001
From: Language Team <no-reply@google.com>
Date: Thu, 19 Oct 2023 21:37:01 +0000
Subject: [PATCH] Dataset release for EMNLP 2023 paper on reasoning with
 spatial prepositions

PiperOrigin-RevId: 575005500
---
 language/capwap/utils/checkpoint_utils.py     |  2 +-
 language/gscan/data/world.py                  |  8 ++-
 language/quest/xattn/gen_training_examples.py |  2 +-
 language/spatial_prep/README.md               | 52 +++++++++++++++++++
 4 files changed, 60 insertions(+), 4 deletions(-)
 create mode 100644 language/spatial_prep/README.md

diff --git a/language/capwap/utils/checkpoint_utils.py b/language/capwap/utils/checkpoint_utils.py
index 60ce7779..0700c8c6 100644
--- a/language/capwap/utils/checkpoint_utils.py
+++ b/language/capwap/utils/checkpoint_utils.py
@@ -24,7 +24,7 @@
 
 def log_variables(name, var_names):
   tf.logging.info("%s (%d total): %s", name, len(var_names),
-                  random.sample(var_names, min(len(var_names), 5)))
+                  random.sample(list(var_names), min(len(var_names), 5)))
 
 
 def init_from_checkpoint(checkpoint_path,
diff --git a/language/gscan/data/world.py b/language/gscan/data/world.py
index 4fd2564b..71380611 100644
--- a/language/gscan/data/world.py
+++ b/language/gscan/data/world.py
@@ -62,7 +62,9 @@ def sample_nearby_position(self, position, exclude_locations=None):
     actual_available_positions = self.get_nearby_positions(
         position, exclude_locations=exclude_locations)
     if actual_available_positions:
-      sampled_position = random.sample(actual_available_positions, 1).pop()
+      sampled_position = random.sample(
+          list(actual_available_positions), 1
+      ).pop()
       return world.Position(column=sampled_position[0], row=sampled_position[1])
     return None
 
@@ -76,7 +78,9 @@ def sample_non_nearby_position(self, position):
                             for dir in DIR_TO_DIR_VEC.values()])
     actual_available_positions = available_positions - nearby_positions
     if actual_available_positions:
-      sampled_position = random.sample(actual_available_positions, 1).pop()
+      sampled_position = random.sample(
+          list(actual_available_positions), 1
+      ).pop()
       return world.Position(column=sampled_position[0], row=sampled_position[1])
     return None
 
diff --git a/language/quest/xattn/gen_training_examples.py b/language/quest/xattn/gen_training_examples.py
index 549de439..31726554 100644
--- a/language/quest/xattn/gen_training_examples.py
+++ b/language/quest/xattn/gen_training_examples.py
@@ -122,7 +122,7 @@ def main(unused_argv):
 
     # Add additional random negatives.
     if FLAGS.random_negatives > 0:
-      random_titles = random.sample(doc_titles, FLAGS.random_negatives)
+      random_titles = random.sample(list(doc_titles), FLAGS.random_negatives)
       for doc_title in random_titles:
         if doc_title not in relevant_titles:
           new_example = xattn_utils.get_example(
diff --git a/language/spatial_prep/README.md b/language/spatial_prep/README.md
new file mode 100644
index 00000000..ca4e38f3
--- /dev/null
+++ b/language/spatial_prep/README.md
@@ -0,0 +1,52 @@
+# A Benchmark for Reasoning with Spatial Prepositions
+
+This folder contains the dataset described in the paper
+ "A Benchmark for Reasoning with Spatial Prepositions" (EMNLP 2023).
+
+## Download the data
+
+* [English dataset](https://storage.googleapis.com/spatial-prepositions-dataset/spatial_prepositions_benchmark_en.tsv)
+* [Romanian dataset](https://storage.googleapis.com/spatial-prepositions-dataset/spatial_prepositions_benchmark_ro.tsv)
+
+## Data format
+
+The datasets are provided in TSV (tab-separated) format.
+ Each row contains an incongruent and a congruent example, both formed using
+ the same prepositions, with the following tab-separated elements:
+ `premise1_a`, `premise2_a`, `conclusion_a`,
+ `no` (language-specific, indicating that `conclusion_a` is invalid),
+ `premise1_b`, `premise2_b`, `conclusion_b`,
+ `yes` (language-specific, indicating that `conclusion_b` is valid).
+
+## Building examples
+
+From each example, a question can be created by combining the two premises
+ and the conclusion, as follows:
+ "If `premise_1` and `premise_2`, does that imply that `conclusion`?"
+
+The examples with invalid conclusions are designed such that a wrong
+ interpretation of the spatial prepositions in the premises can make the
+ conclusion appear valid. For example:
+
+* If `John is in the crib` and `the crib is in the living room`,
+ does that imply that `John is in the living room`? -> `yes`
+ (congruent example with valid conclusion)
+
+* If `John is in the newspaper` and `the newspaper is in the kitchen`,
+ does that imply that `John is in the kitchen`? -> `no`
+ (incongruent example with invalid conclusion)
+
+
+## Citation
+
+If you use this data, please cite:
+
+```
+@inproceedings{comsa2023,
+    author = {Comșa, Iulia-Maria and Narayanan, Srini},
+    title = "{A Benchmark for Reasoning with Spatial Prepositions}",
+    publisher = "Association for Computational Linguistics",
+    booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
+    year = {2023},
+}
+```