# How do you map phrase IDs to sentence IDs in the Stanford Sentiment Analysis dataset?

I am trying to reproduce results obtained on the Stanford Sentiment Analysis dataset, which contains sentiment annotations for syntactic components of sentences taken from Rotten Tomatoes reviews.

The data is partitioned into training, development, and test sets by sentence in the datasetSentences.txt and datasetSplit.txt files. Components within these sentences are mapped to sentiment annotations in the dictionary.txt and sentiment_labels.txt files. The README states that the sentence and phrase IDs are different, but I can't find any mapping between them, so I don't know how to split the sentiment annotations up into the partitions used in the experiment I am trying to reproduce.

Of course I would expect the mapping from phrases to sentences to be many-to-one (because the same phrase might appear in multiple sentences) but I'd still expect to be given an explicit mapping instead of having to compare substrings.

Has anybody worked with this data set? Is there something I am overlooking?

Only full sentences are used for testing and validation, though sentences and phrases are used for training. See Yoon Kim 2014, Convolutional Neural Networks for Sentence Classification for an overview of various text classification experiments.

Here is code that creates training, dev, and test .CSV files from the various text files in the dataset download.

"""
Put all the Stanford Sentiment Treebank phrase data into test, training, and dev CSVs.

Socher, R., Perelygin, A., Wu, J. Y., Chuang, J., Manning, C. D., Ng, A. Y., & Potts, C. (2013). Recursive Deep Models
for Semantic Compositionality Over a Sentiment Treebank. Presented at the Conference on Empirical Methods in Natural
Language Processing EMNLP.

https://nlp.stanford.edu/sentiment/
"""

import os
import sys

import pandas

def get_phrase_sentiments(base_directory):
def group_labels(label):
if label in ["very negative", "negative"]:
return "negative"
elif label in ["positive", "very positive"]:
return "positive"
else:
return "neutral"

dictionary.columns = ["phrase", "id"]
dictionary = dictionary.set_index("id")

sentiment_labels.columns = ["id", "sentiment"]
sentiment_labels = sentiment_labels.set_index("id")

phrase_sentiments = dictionary.join(sentiment_labels)

phrase_sentiments["fine"] = pandas.cut(phrase_sentiments.sentiment, [0, 0.2, 0.4, 0.6, 0.8, 1.0],
include_lowest=True,
labels=["very negative", "negative", "neutral", "positive", "very positive"])
phrase_sentiments["coarse"] = phrase_sentiments.fine.apply(group_labels)
return phrase_sentiments

def get_sentence_partitions(base_directory):
sep="\t")
return sentences.join(splits).set_index("sentence")

def partition(base_directory):
phrase_sentiments = get_phrase_sentiments(base_directory)
sentence_partitions = get_sentence_partitions(base_directory)
# noinspection PyUnresolvedReferences
data = phrase_sentiments.join(sentence_partitions, on="phrase")
data["splitset_label"] = data["splitset_label"].fillna(1).astype(int)
data["phrase"] = data["phrase"].str.replace(r"\s('s|'d|'re|'ll|'m|'ve|n't)\b", lambda m: m.group(1))
return data.groupby("splitset_label")

base_directory, output_directory = sys.argv[1:3]
os.makedirs(output_directory, exist_ok=True)
for splitset, partition in partition(base_directory):
split_name = {1: "train", 2: "test", 3: "dev"}[splitset]
filename = os.path.join(output_directory, "stanford-sentiment-treebank.%s.csv" % split_name)
del partition["splitset_label"]
partition.to_csv(filename)


There is a gist for the code here.

• the code is in completed. At data["phrase"] = return data.groupby("splitset_label") Aug 15, 2017 at 11:33