| Title |
Classification(s) |
Description |
Reference |
Link |
| Using Lexical Chains for Text Summarization |
Extractive or Abstractive / Language Dependant |
Lexical chain :
"Sequence of words which have lexical cohesion (A group of words is lexically cohesive when all of the words are semantically related; for example, when they all concern the same topic)."
Can be built using thesaurus, or part of speech tagging. Greedy disambiguation is not necessarily good. Coherence: "macro level syntactic structure" Cohesion - non-structural connectivity of the text. Uses Wordnet. Finds related sentences (chains) and score = lenght(number of occurrences of members)*homogeneity(1-number of distinct occurrences over the length)
|
Barzilay, Regina and Michael Elhadad. 1997. Using lexical chains for text summarization. In Proceedings of the Workshop on Intelligent Scalable Text Summarization, pages 10-17, Madrid, Spain, August. Association for Computational Linguistics. |
11317796705928478939 |
| The Use of MMR, Diversity-Based Re-ranking for Reordering Documents and Producing Summaries |
Extractive / Statistical / Language Independent |
MMR – Maximum Marginal Relevance - maximize (relevance to query with minimal similarity to previously selected documents). (Enhanced vector based model)
Works well for long documents. Did quite well against many methods.
|
Carbonell, J. and Goldstein, J.. (1998). {The use of MMR, diversity-based reranking for reordering documents and producing summaries}. Proceedings of the 21st annual international ACM SIGIR conference on Research and development in information retrieval, 335--336. |
16038175471093825333 |
| The supply and demand for special education teachers: A review of research regarding chronic shortage of special education teachers. |
Special Education |
Highlights the chronic shortage of special education teachers. |
[Mcl04] McLeskey, J., Tyler, N.C, & Flippin, S.S. (2004). The supply and demand foe special education teachers: A review of research regarding chronic shortage of special education teachers. The Journal of Special Education, 38(1), pp. 5-21. |
5547426187426857161 |
| The diversity-based approach to open-domain text summarization |
Extractive |
Assumes human summaries are not a good gold standard.
Measure: in terms of how well they represent source documents in usual IR tasks such as document retrieval and text categorization
redundancy and diversity. The former relates to how repetitive concepts (or content words) are, the latter relates to how many different concepts there are in the text. (MMR is the approach that accounts for both)
minimum description length principle (MDL) .. choose the simplest model for the data
Use k-means (an updated version called x-means) clustering algorithm. “measuring the loss of information in extracts in terms of retrieval performance” |
Nomoto, T. and Matsumoto, Y.. (2003). {The diversity-based approach to open-domain text summarization}. Information Processing and Management, 39, 363--389. |
12334623528409963050 |
| The Automatic Creation of Literature Abstracts |
Statistical / Language Independent / Extractive |
Word frequency (type of statistical) (first done by Luhn in 58):
• First paper available in automatic summarization.
• Used frequent words to indicate key topics. If not too many junk words between these, the sentence is accepted.
|
Luhn H.P.,"The Automatic Creation of Literature Abstracts" in IBM journal of Research and Development, Vol 2 (2). 1958. |
17331186566995103346 |
| Summarizing Similarities and Differences Among Related Documents |
Language Dependant |
Use a graph based approach with words as the nodes. Using vector space and sentence tagging. Tag synonym and hypernyms using wordnet after finding nouns with a part of speech tagger. (better results with pos tagger vs. just assuming every word is a noun) |
Mani, I. and Bloedorn, E. 2000. Summarizing Similarities and Differences Among Related Documents. Information Retrieval, 1(1). |
16966921942334003677 |
| New methods in automatic extracting. |
Extractive / Language Dependant / Statistical |
Cue words, Position, Title and headings “New Methods in Automatic Extracting” (H. P. Edmundson):
Building on Luhn, also accounts for other text factors. Words such as "this","therefore", “since” etc. indicates sentences that rely, or are related to, other sentences.
Cue words (words that indicate the sentences is important but is itself not a keyword) can be bonus words (words that indicate important) and stigma words (words that indicate unimportant). analysis of the experimental data revealed the following classes of Cue words: Null--ordinals, cardinals, the verb "to be," prepositions, pronouns, adjectives, verbal auxiliaries, articles, and coordinating conjunctions; Bonus-- comparatives, superlatives, adverbs of conclusion, value terms, relative interrogatives, causality terms; Stigma--anaphoric expressions, belittling expressions, insignificant- detail expressions, hedging expressions; and Residue--positives, technical terms, and archaic terms.
Key words – similar to the topic words of luhn. Take most frequent words across document as a measure of sentence importance.
Title words – bonus for occurring in title/ headers, etc.
Position – Important sentences are early or late in a document (and first or last in paragraph) and often close to the title or heading.
(Also discussed by [Brandow 95][Baxendale 58])
Importance of position depends on genre of text
(Edmundson defined the location method, in which sentences
received various scores according to whether they occurred at the beginning or end
of a paragraph, near the beginning or end of the whole document, or below a heading)
|
[Edm69] H. Edmundson. New methods in automatic extracting. Journal of the Association for Computing Machinery, 16(2):264-285, 1969. |
12740448630004231561 |
| Multi-document summarization using off the shelf compression software |
Not feasible for deployment, but a potential method for testing resulting summaries. |
Use gzip to test the compressibility of attempted summaries; a possible measure of redundancy of the summary created. |
Grewal, A. and Allison, T. and Dimitrov, S. and Radev, D.. (2003). {Multi-document summarization using off the shelf compression software}. Proceedings of the HLT-NAACL 03 on Text summarization workshop-Volume 5, , 17--24. |
10803075814660046602 |
| LexRank: Graph-based Centrality as Salience in Text Summarization |
Extractive / Statistical |
Stochastic graph-based method for computing relative importance of textual units for Natural Language Processing eigenvector centrality in a graph representation of sentences.
The centroid of a cluster is a pseudo-document which consists of words that have tf×idf scores above a predefined threshold, where tf is the frequency of a word in the cluster, and idf values are typically computed over a much larger and similar genre data set.
Like in social networking.
Considers where the votes come from and takes the centrality of the voting nodes into account in weighting each vote.
*Graph-based centrality has several advantages over Centroid.
*Prefers more informative sentences in subsumption situations.
*Prevents unnaturally high idf scores from boosting up the score of a sentence that is unrelated to the topic.
|
Erkan, G. and Radev, D.R.. (2004). {LexRank: Graph-based Lexical Centrality as Salience in Text Summarization}. Journal of Artificial Intelligence Research, 22, 457--479. |
15186613774515849351 |
| Improving the Reading Comprehension of Middle School Students with Disabilities through Computer-Assisted Collaborative Strategic Reading. |
Special Education and computers |
In short, with proper application, computer software can be helpful for reading comprehension. |
[Kim06] Kim, Ae-Hwa, et al. "Improving the Reading Comprehension of Middle School Students with Disabilities through Computer-Assisted Collaborative Strategic Reading." Remedial and Special Education. 27.4 (2006): 235-249. |
16626987001119583717 |
| Hybrid Text Summarization: Combining external relevance measures with Structural Analysis |
Language Dependant / Extractive (potentially abstractive) |
Uses discourse analysis. use structural
methods based on discourse parsing to construct
a representation of the text, apply conventional
statistical methods to identify salient information
Improve systems such as summarist by further discourse tree pruning.
|
Thione, Gian Lorenzo, Martin van den Berg, Chris Culy and Livia Polanyi. 2004a. Hybrid Text Summarization: Combining external relevance measures with Structural Analysis. Proceedings ACL Workshop Text Summarization Branches Out. Barcelona. |
4672750300433024448 |
| Generating natural language summaries from multiple on line sources. |
MDS / Extractive and Abstractive / Statistical |
Uses frequency of facts (can by synthesized) . Prototype. |
[Rad97] Dragomir R. Radev. Generating natural language summaries from multiple on line sources. Technical Report CUCS-005-97, Columbia University, Department of Computer Science, New York, NY, USA, March 1997. |
3697622236017510709 |
| Extracting sentence segments for text summarization: a machine learning approach. |
Extractive / Statistical |
Uses decision trees and naive bayes to maximize the information gain on potential sentences (feature vectors are defined using the vector space model). |
[Chu00 ] W. T. Chuang and J. Yang. Extracting sentence segments for text summarization: a machine learning approach. Proceedings of the 23rd International Conference on Research in Information Retrieval (SIGIR ’00), pp. 152–159, 2000. |
15989675033255142794 |
| Evaluation challenges in large-scale document summarization |
Overview of Topic |
Provides an overview of the state-of-the-art text summarizers out there.
Relevance Correlation - measures how well a summary can be used to replace a document for retrieval purposes.
|
Radev, D. R., S. Teufel, H. Saggion, W. Lam, J. Blitzer, H. Qi, A. Celebi, D. Liu & E. Drabek (2003). Evaluation challenges in large-scale document summarization. In Proceedings of the 41st Annual Meeting of the Association for Computational Linguistics, Sapporo, Japan, 7–12 July 2003, pp. 375–382. |
8357967203515710625 |
| Discourse Trees Are Good Indicators of Importance in Text |
Extractive or Abstractive / Language Dependant |
Rhetorical analysis:
Rhetorical Structure Theory (RST)
– Mann & Thompson in 88, A theory of text organization that classes text as a nucleus and a satellite. Uses cue phrases to make a structure tree. (shoes that the nuclei of a discourse structure is commonly perceived to be the most important information in that structure. Induce a partial ordering on the parts of the text (nodes a the top of the discourse tree are more important)
|
Marcu, Daniel. 1999b. Discourse trees are good indicators of importance in text. In Inderjeet Mani and Mark Maybury, editors, Advances in Automatic Text Summarization. MIT Press, Cambridge, MA, pages 123-136. |
15552293050751971577 |
| Constructing literature abstracts by computer |
Extractive & Abstractive / Language Dependant |
Set of features noted by C. D. Paice In “Constructing literature abstracts by computer”. Frequency-Keyword (theme), title-keyword, location, indicator phrases(determine indicative or informative), cue words – bonus(“greatest”, “significant”) and stigma(“hardly”, “impossible”). Discusses difficulties in anaphora recognition and resolution to maintain cohesion of the summary. Describes abstract frames (slots for holding the subtopics of a text). |
Paice, C.. (1990). {Constructing literature abstracts by computer: techniques and prospects}. Information Processing and Management: an International Journal, 26, 171--186. |
16320631068434990296 |
| Centroid-based summarization of multiple documents: sentence extraction, utility-based evaluation, and user studies |
Extractive |
MEAD is centroid based. A centroid is a set of words that are statistically important to a cluster of documents. cluster-based sentence utility-modified vector space model for clustering - check relevance of sentence to that of the document collection.
cross-sentence informational subsumption – does what MMR does but by – identifies sentence subsumption (repetition of information)
|
Radev, D.R. and Jing, H. and Sty{\'s}, M. and Tam, D.. (2004). {Centroid-based summarization of multiple documents}. Information Processing and Management, 40, 919--938. |
17609335274889561888 |
| Automatic Text Structuring and Summarization |
Extractive / Statistical |
Summarizes at the paragraph level instead of the sentence level. Network method using Vector space model:
“It was expected that since a paragraph provides more context, the problems of readability and coherence that were seen in the summaries generated by sentence extraction would be, at least partially, ameliorated.”
Do sentence extraction in a paragraph scope. Decide the important paragraphs in one of 3 ways; bushy path (select highly connected paragraphs in order), depth-first path (select paragraph with most connections, then the next most connected, and so on (in order and favoring bushy nodes)), and segmented bushy path ((good for multi topic) do bushy path looking at all paragraphs regardless to their relation to the last, use most related ‘in between’ paragraph to create a transition).
|
Silla Jr, C.N. and Pappa, G.L. and Freitas, A.A. and Kaestner, C.A.A.. (2004). {Automatic Text Summarization with Genetic Algorithm-Based Attribute Selection}. Advances in Artificial Intelligence--Iberamia 2004: 9th Ibero-American Conference on AI, Puebla, M{\'e}xico, November 22-26, 2004: Proceedings, , . |
17306408869423949521 |
| Automatic Summarization. |
Overview of Topic |
A book covering many topics in automatic summarization. Written from Mani's perspective as an overview of the field for students. |
[Man01] Inderjeet Mani. Automatic Summarization. John Benjamin’s Publishing Company, Amsterdam/Philadelphia, 2001. |
4324463553460310266 |
| Automated text summarization in SUMMARIST. |
Abstractive / Language dependant |
Uses wordnet. Topic identification + interpretation + generation is a summary in their formula. Adjustable level of output from basic topics to dynamically generated sentences. This paper describes Optimal Position Policy.. Places where the topic is likely to be found in the text. This method to learn the positions containing relevant text
needs to have docs, abstracts, and keywords given
Score = a*cue + b*title+ c*position + d*key
Turned out cue+title+position worked best.
• Stemming should be used
• “stop words” (i.e.”the”, “a”, “for”, “is”) are ignored Count concepts instead of words. Otherwise, it’s vector space model. Will (is able to?) generate sentences on the fly from the content (i.e., abstract instead of extract). |
[Hov97] Hovy, Eduard and Chin Yew Lin. 1997. Automated text summarization in SUMMARIST. In Proceedings of the Workshop on Intelligent Scalable Text Summarization, pages 18-24, Madrid, Spain, August. Association for Computational Linguistics. |
6446408290025622442 |
| A Trainable Document Summarizer |
Extractive / Language independent / Statistical |
20% of original with success.
Not include short (<5 words) sentences.
Fixed-phrase feature : certain phrases indicate important sentences.. “in conclusion..”.
Paragraph position (first 10, last 5).
Thematic words – word frequency counts on content words.
Uppercase words – tend to be important, such as names.
Each feature F1through Fk, with probability of including sentence s in summary S.
|
Kupiec, Julian, Jan Pedersen, and Francine Chen. 1995. A trainable document summarizer. In Proceedings of the 18th Annual International ACM/SIGIR Conference, pages 68-73, Seattle, WA. |
874767616512682383 |
| Automatic Text Summarization with Genetic Algorithm-Based Attribute Selection |
GA / Extractive / Trained |
(1) The system constructs a training set where each example (record) corresponds
to a sentence of the original documents, and each example is represented by a set of
attribute values and a known class.
(2) A classification algorithm is trained to predict each sentence’s class (Summary
or Not-Summary) based on its attribute values.
(3) Given a new set of documents, the system produces a test set with predictor
attributes in the same format as the training set. However, the values of the classes are
unknown in the test set.
(4) Each sentence in the test set is classified, by the trained algorithm produced in
step (2), in one of the two classes: Summary or Not-Summary.
The classification algorithms used in the current version of ClassSumm are Naïve
Bayes and C4.5.
After classification, then ranked on a vector of:
1. Position: indicates the position of the sentence in the text, in terms of
percentile
2. Size: indicates the number of terms (words) in the sentence
3. Average-TF-ISF: the TF-ISF (term frequency – inverse sentence frequency)
4. Similarity to Title:
5. Similarity to Keywords:
6. Cohesion w.r.t. All Other Sentences:
7. Cohesion w.r.t. the Centroid:
8 The depth of the sentence in the tree, i.e, the number of nodes that are ancestors of the leaf node representing that sentence.
9. The direction of the sentence in the tree, computed by following the path
from the root towards the sentence up to depth four. At each depth level the
direction can be Left, Right ou None (in case the current level is greater than
the level of the sentence). This produces four attributes, each with one
direction value. These attributes indicate the approximate position of the
sentence in the rhetorical tree, incorporating linguistic knowledge into the set
of predictor attributes.
10. Indicators of Main Concepts: these indicators are computed by using a
morphological part-of-speech tagger that identifies nouns in the document.
The motivation for focusing on nouns is that they tend to be more
meaningful (at least as individual words) than other part-of-speech classes.
The 15 most frequent nouns in the document are selected to be the indicators
of main concepts. For each sentence, the value of this attribute is true if the
sentence contains at least one of those 15 indicators, and false otherwise.
11. Presence of Anaphors:
12. Presence of Proper Nouns: This attribute is computed directly from the
output of a part-of-speech tagger. The value of the attribute is true if the
sentence contains at least one proper noun, and false otherwise.
13. Presence of Discourse Markers: Some discourse markers, such as because,
furthermore, also tend to indicate the presence of non-essential information.
Multi-Objective GA (MOGA) was quite effective., Single Objective (error reduction) was not.
|
Silla Jr, C.N. and Pappa, G.L. and Freitas, A.A. and Kaestner, C.A.A.. (2004). {Automatic Text Summarization with Genetic Algorithm-Based Attribute Selection}. Advances in Artificial Intelligence--Iberamia 2004: 9th Ibero-American Conference on AI, Puebla, Mexico, November 22-26, 2004: Proceedings. |
10862703609584129838 |
| Man-made index for technical literature-an experiment. |
|
It was Baxendale in 1958 who first noted that, within a paragraph, the first sentence is usually the most central to the theme of the text, while in many other cases it is the last. |
Baxendale, P.B. Man-made index for technical literature-an experiment. I.B.M. J. of Res. Dev.. 2(4): 354-361; 1958. |
|
Sync