Skip to content
sumgram is a tool that summarizes a collection of text documents by generating the most frequent sumgrams (conjoined ngrams)
Python
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
pics
sumgram
tests
.gitignore implemenation of logging Aug 31, 2019
LICENSE
README.md Merge pull request #19 from oduwsdl/dev Dec 20, 2019
README_LONG.md switched off default collocations calculation Sep 21, 2019
setup.py

README.md

sumgram

sumgram (see blogpost) is a tool that summarizes a collection of text documents by generating the most frequent "sumgrams" (conjoined ngrams) in the collection. Sumgrams are higher-order ngrams (e.g., "world health organization") generated by conjoining lower-order ngrams (e.g., "world health" and "health organization"). Unlike convention ngram generators that split multi-word proper nouns, sumgram works hard to avoid this by applying two (pos_glue_split_ngrams and mvg_window_glue_split_ngrams) algorithms. These algorithms enable sumgram to generate conjoined ngrams, or sumgrams of different ngram classes (bigrams, trigrams, k-grams, etc.) as part of the summary, instead of limiting the summary to a single ngram class (e.g., bigrams).

From Fig. 1, the six-gram "centers for disease control and prevention" was split (stopwords removed) into 3 different bigrams ("centers disease," "disease control," and "control prevention") by a conventional algorithm that generates bigrams. But sumgram detected and "glued" such split ngrams.

Fig. 1: Comparison of top 20 (first column) bigrams, top 20 (second column) six-grams, and top 20 (third column) sumgrams (conjoined ngrams) generated by sumgram for a collection of documents about the 2014 Ebola Virus Outbreak. Proper nouns of more than two words (e.g., "centers for disease control and prevention") are split when generating bigrams, sumgram strives to remedy this. Generating six-grams surfaces non-salient six-grams. ebola virus ngrams vs sumgrams

Fig. 2: Comparison of top 20 (first column) bigrams, top 20 (second column) six-grams, and top 20 (third column) sumgrams (conjoined ngrams) generated by sumgram for a collection of documents about Hurricane Harvey. Proper nouns of more than two words (e.g., "federal emergency management agency") are split when generating bigrams, sumgram strives to remedy this. Generating six-grams surfaces non-salient six-grams. hurricane harvey ngrams vs sumgrams

Counting Term Frequencies

It is important to note that because sumgram was designed to generate top ngrams (summary) in a collection of text documents, it uses document frequencies (if the collection has more than one document). For example, consider the document frequency of "ebola virus" in the following collection of three documents,

Collection of 3 documents:

Document 1: "ebola virus" occurs 50 times 
Document 2: "ebola virus" occurs 15 times 
Document 3: "ebola virus" occurs 5  times 

According to sumgram, the document frequency (DF) of "ebola virus" is 3 NOT 70

According to sumgram, the document frequency of the term "ebola virus" is 3 NOT 70! Since the goal is to summarize the collection, documents are given a single vote for a single term, so as not to favor any (e.g., long) document or any term (e.g., very popular term within a few documents). However, if the collection contains a single document, term frequencies are used:

Collection of 1 document:
Document 1: "ebola virus" occurs 50 times 
According to sumgram, the TF of "ebola virus" is 50 NOT 1

Irrespective of the case (single-document or multi-document collection) sumgram uses term_freq to count the frequency of terms.

Additional Features

In addition to generating top sumgrams, sumgram ranks sentences and documents.

Ranking documents (--no-rank-docs to switch off)

get_ranked_docs() ranks documents by giving credit to documents that have highly ranked terms in the ranked list of ngrams. A document's score is awarded by accumulating the points awarded by the position of terms in the ranked list of ngrams. Please note that documents without terms in ranked list of ngrams are not awarded points. Therefore, some documents may not be ranked because they performed poorly - did not have any term in the ranked list of ngrams.

Ranking sentences (--no-rank-sentences to switch off)

rank_sents_frm_top_ranked_docs() ranks sentences in the top ranked documents exclusively, and gives credit to sentences with a high average overlap between the sentence tokens and the tokens in the top ngrams. For all sentences in a top ranked documents, a sentence's score (average overlap) is measured by calculating the average overlap between the terms in the top ngrams and the given sentence. This accounts for how many different tokens in the top ngrams that are present in a sentence.

Installation

Just type

$ pip install sumgram

OR

$ git clone https://github.com/oduwsdl/sumgram.git
$ cd sumgram; pip install .; cd ..; rm -rf sumgram;

Recommended Requirement and Performance Considerations

Recommended Requirement

For the best results, we recommend installing and running Stanford CoreNLP Server for two reasons. First, the "pos" in pos_glue_split_ngrams stands for Parts Of Speech (POS). This algorithm needs a POS annotator in order to "glue" split ngrams, hence the need for Stanford CoreNLP server. However, if you do not install Stanford CoreNLP Server, sumgram is robust enough to attempt to glue split ngrams with the second algorithm mvg_window_glue_split_ngrams.

Second, as part of ranking sentences, sumgram needs to segment the sentences in the documents. Stanford CoreNLP's ssplit annotator splits sentences after tokenization, and exploits the decisions of the tokenizer. Probabilitic methods (such as ssplit) for segmenting sentences often outperform rule-based methods that use regular expressions to define sentence boundaries. If you do not install Stanford CoreNLP however, sumgram will adopt a regular expression rule ([.?!][ \n]|\n+) to mark sentence boundaries. This rule can be passed (--sentence-pattern - command line, sentence_pattern - python) as an argument to sumgram.

Performance Considerations - ssplit and Named Entity Recognition

pos_glue_split_ngrams imposes additional runtime overhead on sumgram. You may choose to force sumgram to avoid using the ssplit annotator (implicitly switching off pos_glue_split_ngrams) by setting --sentence-tokenizer=regex (Python: params['sentence_tokenizer'] = 'regex'). Please note that the command line argument --no-pos-glue-split-ngrams does not switch off Stanford CoreNLP's ssplit, it merely avoids the use of the pos_glue_split_ngrams.

We considered leveraging Stanford CoreNLP's Named Entity Annotator as a means to find additional multi-word proper nouns in order to conjoin split ngrams. With a Named Entity Recognition (NER) system, one could easily label a text collection with entity labels (e.g., PERSON, LOCATION, and ORGANIZATION), and instruct the ngram generator to avoid splitting ngrams that have those labels, as a means to remedy the split ngrams problem. However, we decided not to apply NER to resolve split ngrams because NER would impose additional performance overhead upon sumgram. It was important to keep sumgrams as lightweight as possible without compromising the quality of results. There are some phrases such as "direct contact with" and "health care workers," that sumgram could generate unlike NER. However, NER unlike sumgrams provides the benefit of labeling ngrams (e.g, "CDC" - Organization) although with additional performance cost. Even though we recommend using sumgrams with the POS and ssplit annotators from the Stanford CoreNLP suite, and even though they impose additional overhead (while producing better conjoined ngrams and sentence segmentation), sumgrams does not require them to work, and empirical evaluation of sumgrams generated without POS or ssplit have been satisfactory.

Performance Considerations - size of output

By default --no-parent-sentences is switched off, this means that the sentences that mention the top sumgrams are included in the final dictionary output of sumgram (output of get_top_sumgrams()), thus increasing the size of the output. To avoid this, include the --no-parent-sentences option.

Performance Considerations - size of vocabulary (manipulating min_df)

See sklearn.feature_extraction.text.CountVectorizer for more details about min_df.

Sumgram begins by counting the document frequencies (DF: number of documents that include a term) for each term (ngram) in the vocabulary. The size of the vocabulary significantly affects the runtime of sumgram. The larger the vocabulary, the longer the runtime, and vice versa. Fortunately, we can take advantage of the fact that sumgram favors ngrams with DF > 1, to remove terms that occur once or a "few" times. The challenge is that the definition of "few" is subjective. min_df defines what we consider as few: when building the vocabulary, terms with document frequencies less than min_df (DF < min_df) are removed. min_df influences the size of the vocabulary by eliminating terms with DF < min_df. See the following example about how min_df affects the size of the vocabulary.

Document count: 20
Previous vocabulary size: 9,252
New vocabulary size: 9,252
Last term DF rate: 0.0500
Result: Vocabulary size does not shrink since last DF > min_df (0.05 > 0.01). Therefore, no term removed.

Document count: 593
Previous vocabulary size: 127,169
New vocabulary size: 1,321
Last term DF rate: 0.0101
Result: Vocabulary size shrinks by ~98%, as a result 125,848 (with DF < 0.01) terms removed.

These results suggest that if your collection consists of thousands of documents, you might need to increase the min_df (default is: 0.01) threshold. When sumgram is run from the command-line, the last line indicates the rank of the ngram with the least DF (last ngram). For example, the following line

last ngram with min_df = 0.01 (index/DF/DF-Rate): release transcript (1321/6/0.0101)

indicates that the last ngram ("release transcript" - 1,321st ngram) occurred in 6 documents (DF = 6, DF-Rate = 0.0101). So terms with DF < 0.01 were discard from the vocabulary. Therefore, the user could increase min_df if a DF of 6 is still considered small. In contrast, the user could decrease min_df if a DF of 5 is considered big.

Usage

Basic usage:

$ sumgram path/to/collection/of/text/files/

e.g., sumgram tests/sample_cols/harvey

Python script usage:

Command line options may be activated by setting the argument in the params dictionary passed as an argument to get_top_sumgrams(). To set a command line argument, consider the following transformation example:

params = {}
params['sentences_rank_count'] = 20  #For command line argument --sentences-rank-count

The following is a Python script example illustrating the use of sumgram done by calling the get_top_sumgrams() function.

import json
from sumgram.sumgram import get_top_sumgrams

doc_lst = [
    {'id': 0, 'text': 'The eye of Category 4 Hurricane Harvey is now over Aransas Bay. A station at Aransas Pass run by the Texas Coastal Observing Network recently reported a sustained wind of 102 mph with a gust to 132 mph. A station at Aransas Wildlife Refuge run by the Texas Coastal Observing Network recently reported a sustained wind of 75 mph with a gust to 99 mph. A station at Rockport reported a pressure of 945 mb on the western side of the eye.'},
    {'id': 1, 'text': 'Eye of Category 4 Hurricane Harvey is almost onshore. A station at Aransas Pass run by the Texas Coastal Observing Network recently reported a sustained wind of 102 mph with a gust to 120 mph.'},
    {'id': 2, 'text': 'Hurricane Harvey has become a Category 4 storm with maximum sustained winds of 130 mph. Sustained hurricane-force winds are spreading onto the middle Texas coast.'}
  ]

'''
  Use 'add_stopwords' to include additional stopwords not included in stopwords list (https://github.com/oduwsdl/sumgram/blob/0224fc9d54034a25e296dd1c43c09c76244fc3c2/sumgram/util.py#L31)
  'add_stopwords' expects a comma-separated string of stopwords, e.g., "image, photo, image of"
'''
params = {
    'top_sumgram_count': 10,
    'add_stopwords': 'image',
    'no_rank_sentences': True,
    'title': 'Top sumgrams for Hurricane Harvey text collection'
}

ngram = 2
sumgrams = get_top_sumgrams(doc_lst, ngram, params=params)
with open('sumgrams.json', 'w') as outfile:
  json.dump(sumgrams, outfile)

Examples (see sample collection tests/sample_cols/harvey):

Generate top 10 (t = 10) sumgrams for the Archive-It Ebola Virus Collection:

$ sumgram -t 10 cols/ebola/
 rank  sumgram                                              DF   DF-Rate
  1    in west africa                                       50    0.35 
  2    liberia and sierra leone                             46    0.33 
  3    ebola virus                                          44    0.31 
  4    ebola outbreak                                       41    0.29 
  5    public health                                        40    0.28 
  6    the centers for disease control and prevention       23    0.16 
  7    the united states                                    23    0.16 
  8    the world health organization                        22    0.16 
  9    ebola patients                                       20    0.14 
  10   health workers                                       20    0.14 

Generate top 10 (t = 10) sumgrams for the Archive-It Hurricane Harvey Collection:

$ sumgram -t 10 cols/harvey/
rank   sumgram                                              DF   DF-Rate
  1    hurricane harvey                                     20    0.47 
  2    tropical storm harvey                                10    0.23 
  3    2017 houston transtar inc.                           9     0.21 
  4    2017. photo                                          9     0.21 
  5    corpus christi                                       9     0.21 
  6    image 28 of                                          9     0.21 
  7    image 29 of                                          9     0.21 
  8    image 30 of                                          9     0.21 
  9    image 31 of                                          9     0.21 
  10   image 32 of                                          9     0.21

This collection has lots of images, but the "image" term might obscure more salient ngrams, so let's rerun the command, but this time consider "image" a stopword (--add-stopwords="image"). As seen below such modification exposed more salient bigrams such as "buffalo bayou" and "coast guard". The argument of --add-stopwords is a comma-separated string of stopwords (e.g., "image, photo, image of"). Use this parameter to add domain specific stopwords not included in sumgram's default stopwords list.

$ sumgram -t 10 --add-stopwords="image" cols/harvey/
 rank  sumgram                                              DF   DF-Rate
  1    hurricane harvey                                     20    0.47 
  2    tropical storm harvey                                10    0.23 
  3    2017 houston transtar inc.                           9     0.21 
  4    2017. photo                                          9     0.21 
  5    corpus christi                                       9     0.21 
  6    texas photo                                          9     0.21 
  7    27, 2017                                             8     0.19 
  8    buffalo bayou                                        8     0.19 
  9    coast guard                                          8     0.19 
  10   harvey photo                                         8     0.19 

Use application from Python Docker container to generate top 10 (t = 10) sumgrams for the Archive-It Hurricane Harvey Collection:

$ docker run -it --rm --name my-running-script -v "$PWD":/usr/src/myapp -w /usr/src/myapp --network=host python:3.7-stretch bash
$ pip install sumgram
$ sumgram -t 10 --add-stopwords="image" cols/harvey/
 rank  sumgram                                              DF   DF-Rate
  1    hurricane harvey                                     20    0.47 
  2    tropical storm harvey                                10    0.23 
  3    2017 houston transtar inc.                           9     0.21 
  4    2017. photo                                          9     0.21 
  5    corpus christi                                       9     0.21 
  6    texas photo                                          9     0.21 
  7    27, 2017                                             8     0.19 
  8    buffalo bayou                                        8     0.19 
  9    coast guard                                          8     0.19 
  10   harvey photo                                         8     0.19 

Generate top sumgrams from collection of URLs:

This requires dereferencing the URLs and removing the HTML boilerplate. This example requires the installation of NwalaTextUtils.

import json
from NwalaTextUtils.textutils import parallelGetTxtFrmURIs
from sumgram.sumgram import get_top_sumgrams

ngram = 2
uris_lst = [
  'http://www.euro.who.int/en/health-topics/emergencies/pages/news/news/2015/03/united-kingdom-is-declared-free-of-ebola-virus-disease',
  'https://time.com/3505982/ebola-new-cases-world-health-organization/',
  'https://www.scientificamerican.com/article/why-ebola-survivors-struggle-with-new-symptoms/'
]

doc_lst = parallelGetTxtFrmURIs(uris_lst)
sumgrams = get_top_sumgrams(doc_lst, ngram)

with open('sumgrams.json', 'w') as outfile:
    json.dump(sumgrams, outfile)

Sumgram output

Highlights of selected fields in the output (harvey_sumgrams.json) generated from the following command.

sumgram -t 20 -o harvey_sumgrams.json --pretty-print cols/harvey/
  • base_ngram (int)
  • top_sumgram_count (int)
  • ranked_docs (optional, default ON) array[objects]):
    • (object)
      • score: Score assigned to document (int)
      • doc_id: auto-generated identifier for document (int)
      • doc_details (object): user-supplied document details
  • ranked_sentences (optional, default ON) (array[objects]):
    • (object)
    • avg_overlap (float): Average overlap between sentence and all the different top sumgrams
    • sentence (string)
    • doc_indx (int): Integer position of document where sentence was extracted
    • doc_id (int)
    • sent_indx (int): Integer position of sentence within document sentence was extracted
    • segmenter (string): Method (ssplit or regex) used for segmenting this sentence
  • top_sumgrams: array[objects])
    • (object)
    • ngram (string): ngram or sumgram (conjoined ngram)
    • term_freq (int): The total number of times ngram occurs in the collection. Every ngram is counted once within a document if the collection contains multiple documents (term_freq = Document Frequency), but not if the collection has a single document (term_freq = Term Frequency).
    • term_rate (int): term_freq as fraction of total collection
    • parent_sentences array[objects]: sentences that mention ngram
    • sumgram_history array[objects]: historical information showing how ngram was was conjoined to form a sumgram
    • (object)
      • prev_ngram (string): previous state of ngram
      • annotator (string): annotator (pos or mvg_window) responsible for conjoining ngrams to form sumgram
      • cur_ngram (string): current state of ngram
      • cur_freq (int): occurrence frequency of cur_ngram
      • cur_pos_sequence (array[string]): Part of Speech labels of cur_ngram
      • proper_noun_rate (float): fraction of ngram tokens that are labeled NNP
  • params (objects)

Full usage

sumgram [options] path/to/collection/of/text/files/

Options:
-n=2                                      The base n (int) for generating top sumgrams, if n = 2, bigrams become the base ngram

-d, --print-details                       Print details
-o, --output                              Output file
-s, --sentences-rank-count=10             The count of top ranked sentences to generate
-t, --top-sumgram-count=10                The count of top sumgrams to generate

--add-stopwords                           Comma-separated list of addition stopwords
--base-ngram-ansi-color='91m'             Highlight (color code format - XXm, e.g., 91m) base ngram when printing top ngrams, set to empty string to switch off color
--collocations-pattern                    User-defined regex rule to extract collocations for pos_glue_split_ngrams
--corenlp-host=localhost                  Stanford CoreNLP Server host (needed for decent sentence tokenizer)
--corenlp-port=9000                       Stanford CoreNLP Server port (needed for decent sentence tokenizer)
--corenlp-max-sentence-words=100          Stanford CoreNLP maximum words per sentence
--include-postings=False                  Include inverted index of term document mappings

--log-file                                Log output filename
--log-format                              Log print format, see: https://docs.python.org/3/howto/logging-cookbook.html
--log-level=info                          Log level from OPTIONS: {critical, error, warning, info, debug, notset}

--mvg-window-min-proper-noun-rate=0.5     Mininum rate threshold (larger, stricter) to consider a multi-word proper noun a candidate to replace an ngram
--min-df=0.01                             See min_df in https://scikit-learn.org/stable/modules/generated/sklearn.feature_extraction.text.CountVectorizer.html
--ngram-printing-mw=50                    Mininum width for printing ngrams

--no-mvg-window-glue-split-ngrams=False   Do not glue split top ngrams with Moving Window method (default is False)
--no-parent-sentences                     Do not include sentences that mention top ngrams in top ngrams payload (default is False)
--no-pos-glue-split-ngrams=False          Do not glue split top ngrams with POS method (default is True)
--no-rank-docs=False                      Do not rank documents flag
--no-rank-sentences=False                 Do not rank sentences flag

--parallel-readtext                       Read input files in parallel
--pos-glue-split-ngrams-coeff=0.5         Coeff. ([0, 1]) for permitting matched ngram replacement by pos_glue_split_ngrams(), bigger means stricter
--pretty-print=False                      Pretty print JSON output
--rm-subset-top-ngrams-coeff=0.5          Coeff. ([0, 1]) for permitting matched ngram replacement by rm_subset_top_ngrams(), bigger means stricter

--sentence-pattern='[.?!][ \n]|\n+'       For sentence ranking: Regex string that specifies tokens for sentence tokenization
--sentence-tokenizer=ssplit               For sentence ranking: Method for segmenting sentences, options: {ssplit, regex}
--shift=0                                 Factor to shift top ngram calculation
--token-pattern                           Regex string that specifies tokens for document tokenization. Default = '\b[a-zA-Z\'\’-]+[a-zA-Z]+\b|\d+[.,]?\d*'
--title                                   Text label to be used as a heading when printing top sumgrams
--thread-count=5                          Maximum number of threads to use for parallel operations like segmenting sentences
--update-rate=50                          Print 1 message per update-rate for long-running tasks

Algorithms for detecting and gluing split Multi-Word Proper Noun (MWPN) ngrams

pos_glue_split_ngrams

This algorithm is the first measure to merge split MWPN ngrams. For example, the ngram fragment "emergency management" was split (base ngram = 2) from its parent MWPN: "federal emergency management agency". pos_glue_split_ngrams attempts to replace the ngram fragment with its MWPN.

The pos_glue_split_ngrams process is outlined as follows:

  • All tokens in all the sentences are labeled with their respective Parts of Speech (POS) with Stanford CoreNLP's POS annotator.

  • MWPNs are identified by this rule: a MWPN is a contiguous sequence of NNP or a contiguous sequence of NNP interleaved with CC or IN. For example, given the follow POS label descriptions

    NNP: Proper Noun Singular
    NNPS: Proper Noun plural
    CC: Coordinating conjunction
    IN: Preposition or subordinating conjunction
    

    According to pos_glue_split_ngrams, the following ngram sequences are MWPNs:

    "Hurricane harvey"                    - (NNP NNP)
    "Centers for Disease Control"         - (NNP IN NNP NNP)
    "Federal Emergency Management Agency" - (NNP NNP NNP NNP)
    
  • Let TF_C = Term Frequency of fragment child ngram (e.g., emergency management). Let TF_P = Term Frequency of MWPN (e.g., federal emergency management agency). pos_glue_split_ngrams replaces a fragment child ngram with a parent MWPN, if TF_P > TF_C * pos_glue_split_ngrams_coeff. The restriction is done in order to avoid replacing a high-quality fragment child ngram (e.g., tropical storm with TF_C: 121) with a poor-quality MWPN (e.g., ddhhmm tropical storm harvey discussion number with TF_P: 5)

mvg_window_glue_split_ngrams

This algorithm is the second measure to merge split MWPN ngrams. Unlike pos_glue_split_ngrams which relies on a rule for identifying MWPNs, mvg_window_glue_split_ngrams seeks out terms (controlled by an expanding window size) that are frequently left and right neighbors of the fragment ngrams.

For example, from the Hurricane Harvey collection, the ngram fragment emergency management, often had the term federal to its left, and agency to its right. The concatenation of the fragment ngram and its highly frequent left and right neighbors is considered a MWPN: federal + emergency management + agency. If the co-occurrence rate of the MWPN >= mvg_window_min_proper_noun_rate, mvg_window_glue_split_ngrams replaces the child fragment ngram (emergency management) with its parent MWPN (federal emergency management agency).

Snippet of mvg_window_glue_split_ngrams process:

mvg_window_min_proper_noun_rate: 0.5 (Let the MWPN occur at least 50% of the time)
window_size: 1
fragment ngram: ['emergency', 'management']
candidate sentence tokens: 

['more', 'than', '32', '000', 'people', 'have', 'been', 'housed', 'in', 'shelters', '', 
'and', 'the', 'federal', 'emergency', 'management', 'agency', 'is', 'expecting', 'nearly', 
'a', 'half', 'million', 'people', 'to', 'seek', 'some', 'sort', 'of', 'disaster', 'aid.']

#Here are the candidate MWPNs generated by adding 1 (window_size = 1) term to the left, right, and both left and right of the fragment ngram:

left  MWPN: "federal emergency management"        occurrence rate: 0.875
right MWPN: "emergency management agency"         occurrence rate: 0.625
both  MWPN: "federal emergency management agency" occurrence rate: 0.625 

mvg_window_glue_split_ngrams favors longer MWPNs as long as its occurrence rate >= mvg_window_min_proper_noun_rate, so even though the left MWPN's (federal emergency management) occurrence rate is the highest (0.875), the algorithm would select the longest MWPN (federal emergency management agency) since it fulfills the selection criteria: its occurrence rate (0.625 > mvg_window_min_proper_noun_rate = 0.625 > 0.5). For each fragment ngram, mvg_window_glue_split_ngrams searches for the longest MWPN that fulfills the selection criteria.

Sumgrams vs. Collocations

Sumgram may be likened to some of the multiple methods of detecting common phrases or collocations since it strives to identify multi-word proper nouns. But there are some significant differences.

TLDR: Sumgram generates summaries by finding conjoined ngrams that are representative of the entire collection and not merely groups of words that co-occur. The focus of collocation method is not to summarize collections but to detect word groups that frequently occur together.

First, the primary goal of sumgrams is to summarize text documents in a collection, the process of conjoining split ngrams is secondary. This primary goal is the reason sumgram uses document frequency (DF) in order to be give all documents a fair chance in deciding what terms are important, and to represent as many different (diverse) terms as possible. From Figs. 3 and 4, I claim the sumgrams produce more diverse and cohesive topics compared to the bigram collocations. Collocations unlike sumgram use raw TF counts and the primary focus of collocation is to identity groups of words (with limited compositionality) that frequently co-occur. Therefore, unlike collocation methods, sumgram simultaneously generates a summary for a collection and conjoins split ngrams (akin to extracting collocations). However, sumgrams is primarily a collection summarization method.

Second, collocation method often begin by ranking frequent ngrams (e.g., bigrams or trigrams). Variation comes with applying filters: stopwords/punctuation removal, use of Pointwise Mutual Information, Chi-squared tests, T-test, or some other statistical tests to ensure collocations are statistically significant. The key difference here between sumgram and collocation methods is: after collocation detection methods apply filters to get better results (e.g., bigrams/trigrams), they often stop. This means by design, since we restrict the calculation of collocations to specific n-grams (e.g, bigrams), we may split collocations with more that two terms. In contrast, after sumgram returns bigrams for example, it tries to expand the bigram into a k-gram (k > 2) especially if the bigram is part of a multi-word proper noun. So it is not sufficient that the bigram is frequent, sumgram strives to avoid splitting MWPNs. This gives sumgram the flexibility to return multiple ngrams (bigrams, trigrams, six-grams, etc.) as part of the list of most frequent ngrams in a collection.

Fig. 3: Comparison of top 20 sumgrams and top 20 bigram collocations generated using different methods for labeling collocations. Since collocations are calculated for fixed (n=2) ngrams, they are prone to splitting MWPNs (highlighted). The collocation output were generated by running Nicha Ruchirawat's implementation of some common methods for identifying ngram collocations. sumgrams vs collocations ebola

Fig. 4: Comparison of top 20 sumgrams and top 20 bigram collocations generated using different methods for labeling collocations. Since collocations are calculated for fixed (n=2) ngrams, they are prone to splitting MWPNs (highlighted). The collocation output were generated by running Nicha Ruchirawat's implementation of some common methods for identifying ngram collocations. sumgrams vs collocations harvey

Sumgrams vs. LDA

Fig. 5: Comparison of top LDA topics and top 20 sumgrams (conjoined ngrams) generated by sumgram for a collection of documents about Hurricane Harvey. sumgrams vs lda LDA often used to discover abstract "topics" in a large volume of documents. Fig. 5 juxtaposes the top 20 topics surfaced by a Python LDA library at top 20 sumgrams. As seen in Fig. 5 (Column 2), LDA abstract topics confirm our intuition that the collection is indeed about a Hurricane. However, I am uncomfortable with LDAs output and prefer sumgrams as a summary for the collection for the following reasons:

  • It is not easy to find set the words per topic input to the LDA system. In Fig. 5 we set words per topic = 5 to facilitate comparison with sumgrams. Also, since our collection is homogeneous (we already know most documents are about Hurricane Harvey), apply LDA to our collection is akin to finding the subtopics within our broader Hurricane Harvey topic. Consequently, if words per topic is set to small, we may abbreviate a subtopic, and if it is set too high, we may mix multiple subtopics within a single LDA abstract topic. This means LDA might be too coarse to find subtopic because it is hard to control topic boundaries in a supervised manner. Sumgrams in contrast isolate entities. For example, the first sumgram identifies only one entity hurricane harvey, as well as the second: the federal emergency management agency.
  • Similar to the first point, the LDA topics are not sequences, words in the topics are ordered according to their probability of belonging to the topic. This means the discovery of proper nouns in the LDA topics is not obvious. For example, in the 8th LDA topic puerto hurricane rico maria images, rico is not next to puerto. I could only detect this because I am familiar with the collection topic.
You can’t perform that action at this time.