Summarisation Infrastructure and Baselines: Summary
This page summarises the public deliverable “D6.1 Summarisation infrastructure and baselines,” which describes the first summarisation infrastructure developed for the MULTISENSOR project and the datasets prepared to support future summarisation work.
The report covers two main areas: an initial web-based summarisation service built around extractive methods, and the early construction of corpora intended to support more advanced abstractive summarisation techniques in later stages of the project.
What the Report Covers
The deliverable presents the first-year progress of WP6, focused on summarisation and content delivery. In practical terms, it reports:
- a baseline extractive summarisation system for single-document and multi-document summarisation
- the integration of that system into the wider project architecture
- the collection and preparation of datasets for extractive and abstractive summarisation
- plans for later abstractive summarisation based on information extraction, semantic repositories, and natural language generation
Main Goal of the Infrastructure
The core purpose of the infrastructure was to provide an operational baseline. Instead of beginning directly with abstractive summarisation, the project first implemented an extractive approach that could identify relevant sentences and compose summaries by selecting and joining them verbatim from the source text.
This baseline was intended to serve as a reference point for later, more advanced summarisation methods.
Extractive Summarisation Baseline
The report explains that the initial infrastructure was based on a multilingual text-to-text extractive summarisation module. Extractive summarisation was defined here as selecting relevant fragments from the source documents rather than generating new text from scratch.
The implementation used a state-of-the-art extractive summarisation toolkit and supported both:
- single-document summarisation
- multi-document summarisation
According to the report, two SUMMA-based pipelines were created for these two settings.
How the Baseline Worked
For the single-document pipeline, the report describes a sequence of ANNIE and SUMMA modules including tokenisation, punctuation handling, sentence splitting, named-entity statistics, term-frequency filtering, position scoring, vector computation, sentence-document similarity, first-sentence similarity, and weighted sentence scoring.
The multi-document pipeline extended this process by comparing sentences across the full document set and by adding centroid-based calculations, so that sentence relevance could be assessed not only within one document but also relative to the broader set.
The pipelines could be parameterised by compression rate, metric weights, and an IDF table derived from a relevant corpus.
User Requirements and Summary Types
The report also notes that user requirements were explored together with project user partners. Several possible summary and presentation strategies were considered, including:
- lists of keywords
- extractive summaries
- abstractive summaries
- graphic visualisation of semantic repository contents
This is useful because it shows that the work was not limited to algorithm design. It also considered how summaries should be presented and what kinds of outputs different users might actually find useful.
The Summarisation Datasets
A major part of the report is devoted to the summarisation dataset work. The document says the dataset effort involved three corpora:
- a paired corpus of press articles and human-authored summaries
- a corpus of press articles automatically annotated with linguistic and semantic information
- a manually annotated corpus with linguistic features for later natural language generation work
These corpora were intended to support both the extractive baseline and the future abstractive pipeline.
Extractive Corpus
The report states that a first corpus of general press articles was provided by pressrelations and stored in the news repository. It contained 65,642 pairs of articles and human-crafted summaries, with summaries averaging about 125 words compared with about 445 words for the original documents.
This corpus was used to tune the extractive pipelines to journalistic material.
Text Planning and Lexicalisation Dataset
For abstractive summarisation, the report describes a NIF-based linked-data corpus produced through the project’s multimedia and linguistic analysis pipeline. This dataset was intended to support content determination and lexicalisation for the NLG pipeline.
The annotations included entities, relations, sentiments, context, and concepts or events depicted in images and video keyframes, represented using NIF and serialised with JSON-LD.
Multilingual Surface Generation Dataset
The report also describes a multilingual surface-generation dataset intended for sentence planning and surface realisation. This work followed layers from Meaning-Text Theory, moving from deep-syntactic structure to surface-syntactic structure, morphology, and final sentence output.
This dataset was meant to support the training or design of multilingual generation methods.
Important Limitations
The report is quite clear about limitations at this stage. It says the baseline was only an initial implementation and still required improvements to reach true state-of-the-art performance. Planned improvements included using better linguistic annotations from Linguatec, tuning separate pipelines per use case, adding richer metrics such as title similarity and named-entity-based analysis, and setting weights empirically using human summaries.
It also notes that multilingual summarisation beyond English would be delayed because annotated corpora were not yet available, and machine translation would be used as a workaround once the necessary MT components were ready.
Licensing Constraints
An important practical point in the report is that publication of the summarisation datasets was constrained by licensing. The extractive corpora were to be published via URLs pointing to original texts, while the multilingual surface-generation dataset could not be released because rights to the underlying third-party corpora were not owned by the project.
As a workaround, the report mentions the possibility of releasing tools that could transform licensed corpora into the deeper representations needed by the system.
Why This Deliverable Matters
This report is useful because it shows how summarisation was approached as both an engineering and data problem. It did not only implement a baseline service; it also built the corpus foundations for later abstractive and multilingual summarisation work.
For readers interested in speech and language systems, multimodal AI, summarisation, NLG, and semantic content processing, it provides a clear example of how a real project moved from a practical extractive baseline toward a richer semantic and generation-based pipeline.
Key Takeaway
The main takeaway is that the report established two things at once: a working extractive summarisation baseline, and the data infrastructure needed for more advanced abstractive summarisation later on. Its importance lies in creating that bridge between an immediately usable summariser and a more ambitious semantic-NLG architecture.