2
$\begingroup$

I'm researching on abstractive text summarization, and has come across many recent papers. They all seem to be focusing on Sequence to Sequence models based on RNNs. Apart from RNNs, what other approaches are there when it comes to abstractive text summarization? Does ontology-based summarization revolve around the same seq2seq model?

Most of the material I have come across are research papers on this subject; what better sources are there to get an understanding of the underlining concepts of abstractive summarization?

$\endgroup$
1

1 Answer 1

1
$\begingroup$

Apart from RNNs, what other approaches are there when it comes to abstractive text summarization?

There are many other systems, I copy below some references to them:

From {1}:

Unfortunately, abstractive summarization is known to be difficult. Existing work in abstractive summarization has been quite limited and can be categorized into two categories: (1) approaches using prior knowledge (Radev and McKeown, 1998) (Finley and Harabagiu, 2002) (DeJong, 1982) and (2) approaches using Natural Language Generation (NLG) systems (Saggion and Lapalme, 2002) (Jing and McKeown, 2000). The first line of work requires considerable amount of manual effort to define schemas such as frames and templates that can be filled with the use of information extraction techniques. These systems were mainly used to summarize news articles. The second category of work uses deeper NLP analysis with special techniques for text regeneration. Both approaches either heavily rely on manual effort or are domain dependent.

You may also want to have at section 2.3.2 Toward full abstraction and table 1 from {2}:

enter image description here

2.3.2 Toward full abstraction

Fully abstractive summarization attempts to understand the input and generate the summary from scratch, usually including sentences or phrases that may not appear in the original document. It actually involves multiple subproblems, each of its own can be made a relatively independent research topic, including: simplification, paraphrasing, merging or fusion. Cheung and Penn [29] conduct a series of studies comparing human-written model summaries to system summaries at the semantic level of caseframes, which are shallow approximations of the semantic role structure of a proposition-bearing unit like a verb and are derived from the dependency parse of a sentence. They find that human summaries are: (1) more abstractive, using more aggregation, (2) contain less caseframes and (3) cannot be reconstructed solely from original documents but are able to if in-domain documents are added.

Due to the inherent difficulty and complexity of full abstraction, current research in abstractive document summarizationmostly restricts in one or a fewof the subproblems. It is also less active compared with compressive summarization, since merely considering compressions has already boosted system performance, as discussed in the last section.9

Woodsend and Lapata [195] propose a model that allows paraphrases induced from a quasi-synchronous tree substitution grammar (QTSG) to be selected in the final ILP model covering content selection, surface realization, paraphrasing and stylistic conventions. For document summarization that involves paraphrasing and fusing multiple sentences simultaneously, other than grammar-based rewriting, one simpler more typical approach is to merge information contained in sub-sentence-level units. For instance, one can cluster sentences, build word graphs and generate (shortest) paths from each cluster to produce candidates for making up a summary [6,51]. More sophisticated treatments can also be built on syntactic or semantic analysis. One may build sentences via merging consistent noun phrases and verb phrases [13] or linearizing graph-based semantic units derived from semantic formalisms such as abstract meaning representation (AMR) [121].

There also exist psychologically motivated studies [48] trying to implement cognitive human comprehension models based on propositions, which are elements extracted from an original sentence, each containing one functor and several arguments. Propositions form a tree where a proposition is attached to another proposition with which they share at least one argument. Summaries are then generated from selected important propositions. Currently the systems have mostly been evaluated on over-specific datasets and rely heavily on various components including parsing, coreference resolution, distributional semantics, lexical chains [49] and natural language generation from semantic graphs [50

In order to better guide alignment and merging processes, supervised learning-basedmethods have been investigated [46,178]. A later work [30] expands the sentence fusion process with external resources beyond the input sentences by combining the subtrees of many sentences, allowing for relevant information from sentences that are not similar to the original input sentences to be added during fusion.

Abstractive summarization has also been studied in information extraction (IE) perspective, for example, IE-informed metrics have also been shown to be useful to rerank the output of high performing baseline summarization systems [83]. In the context of guided summarization where predefined categories and readers’ intent have been predefined, preliminary full abstraction can be achieved by extracting templates using predefined rules for different types of events [59,166].

In order to better guide alignment and merging processes, supervised learning-basedmethods have been investigated [46,178]. A later work [30] expands the sentence fusion process with external resources beyond the input sentences by combining the subtrees of many sentences, allowing for relevant information from sentences that are not similar to the original input sentences to be added during fusion. Abstractive summarization has also been studied in information extraction (IE) perspective, for example, IE-informed metrics have also been shown to be useful to rerank the output of high performing baseline summarization systems [83]. In the context of guided summarization where predefined categories and readers’ intent have been predefined, preliminary full abstraction can be achieved by extracting templates using predefined rules for different types of events [59,166].


References:

$\endgroup$
1
  • $\begingroup$ I have been skimming over this paper too.. Thank you for the pointers :) $\endgroup$
    – Dee
    Commented Jan 14, 2018 at 15:00

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.