Wikianswers/Technical discussion

Natural-language processing

edit
For further information, see Natural-language processing.

Question paraphrasing

edit

Paraphrases of questions should each route end-users to the same wiki content which presents them with a set of justified answers.

Paraphrase-handling algorithms could utilize wiki-style redirections from phrasing-specific wiki pages to phrasing-generic wiki pages. In this way, users could correct the outputs of paraphrase-handling algorithms. These algorithms could improve as a result of users’ corrections.

Wikidata could also be of use for storing data with respect to questions and question paraphrases.

There may be other ways to develop user-correctable question-paraphrasing algorithms.

Question transformation

edit

Questions could also be transformed while preserving user intent so that resultant answers would imply or entail the answers to user-provided phrasings of questions. That is, while paraphrasing approximately preserves semantics, question transformations may result in different semantics, e.g., replacing a specific instance with a universally quantified variable.

Transformation-handling algorithms could utilize wiki-style redirections. In this way, users could correct the outputs of question transformation algorithms. These algorithms could improve as a result of users’ corrections.

There may be other ways to develop user-correctable question-transformation algorithms.

Question suggestion

edit

Either as users type their questions or after having entered their questions, users could be presented with a list of similar questions which have already been answered.

Question categorization

edit

Questions can be categorized into one or more domains for delegation to one or more question-answering systems.

Implementations of question categorization handling algorithms could utilize wiki-style categories on content. In this way, wiki users could correct the outputs of question categorizing algorithms and these algorithms could, similarly, improve as a result of users’ corrections.

There may be other ways to develop user-correctable question categorization algorithms.

Question evaluation

edit

Questions could be evaluated in terms of their predicted value to "Wikianswers", both as a resource for end-users and as training data for multiple interoperating artificial intelligence systems.

Algorithms should be able to detect and deflect mischievous or malicious questions, a form of vandalism. A list of these algorithmically-deflected questions could be archived and made available for review. Each deflected question could also be provided with a canonical URL, this displayed for an end-user upon the event of a deflection (perhaps alongside social media sharing buttons), so that, in the event of an error or dispute, the end-user could, for instance, make use of social media to discuss the performance of the question evaluation algorithms.

Question prioritization

edit

As discussed in § Question-asking systems, other systems could be awaiting human-provided answers to machine-generated questions in order to answer users' questions. For these and other reasons, algorithms may be able to score or to prioritize enqueued questions.

Argument technology

edit
For further information, see Argument technology.

Verification and validation

edit

Justifications and arguments can be algorithmically verified and validated.

As envisioned, verification and validation systems could annotate justifications and arguments with visual indicators for any informational, warning, or error messages. These visualized annotations could be hyperlinks to explanatory content.

Edits to justifications and arguments could be algorithmically verified and validated. It would be convenient for users to be able to preview provisional annotations from verification and validation systems while previewing their edits.

Vandalism detection

edit

In particular as the wiki pages in "Wikianswers" are structured, each page containing a set of answers, each answer supported by one or more justifications, it should be easier to implement vandalism-detection algorithms.

Ranking answers and justifications

edit

Solutions can be devised – beyond crowdsourced upvoting and downvoting – for scoring, ranking, and sorting candidate answers and their justifications for purposes of presentation and display.

edit
For further information, see Argumentation framework.

Described, thus far, is an abstract framework where each question can have multiple answers and each answer can be supported by multiple justifications. In this section, some related argumentation frameworks are indicated. This project proposal, "Wikianswers", intends to be compatible with any argumentation framework.

It appears to require multiple editing passes from question-answering systems to produce more complex and interwoven argument structures, or for multiple agents or systems to engage in debate on a wiki question-and-answer platform.

DebateTree algorithm

edit
For further information, see DebateTree algorithm.

Argument maps

edit
For further information, see Argument map.

Argument Web

edit
For further information, see Argument Web.
edit

Argument Interchange Format

edit
For further information, see Argument Interchange Format.
edit
For further information, see Legal Knowledge Interchange Format.

Argdown

edit
For further information, see https://argdown.org/ and MediaWiki Argdown extension.

Other algorithms

edit

Cache management

edit

As envisioned, end-users will ask natural-language questions and either will be routed to view existing wiki pages containing one or more justified answers or will briefly await the generation of new such content by one or more question-answering systems.

Machine-generated content which has not been edited by human users, as this content can be computationally regenerated, can be viewed as content in a type of cache. Cache-management algorithms can be of use, then, for a number of important scenarios including when artificial intelligence question-answering systems are versioned.

Also, for scenarios where storage resources are limited, algorithms, e.g., traffic-based, could determine which wiki pages are more relevant than others. Algorithms can be devised to prune, or to garbage-collect, barely used or otherwise superfluous content.

Dependency graphs

edit

As machine-generated answers and their justifications can depend upon wiki content (including other answers, as discussed in § Question-asking systems), question-answering systems might watch certain content for any edits or updates, subscribing to receive pings as changes occur, to subsequently update their machine-generated content and/or to make use of answers' discussion areas or threaded forums.

Question-asking systems

edit

In addition to providing justified answers for end-users’ questions, some artificial intelligence systems could also produce questions, e.g., while working on tasks or reasoning. Question-asking systems could make use of an API to enqueue questions for end-users to review, prioritize, and answer using the "Wikianswers" platform.

Some question-answering systems could also function as question-asking systems during the course of their reasoning processes. When a system must consult human end-users, the system could present its questions to the wiki question-and-answer platform. This is one scenario where an answer to an end-user would not be instantaneous or where an instantaneously provided machine-generated answer might include hyperlinks to enqueued and awaited questions.

Decision-making systems

edit
For further information, see Automated decision-making and Debate.

As envisioned, automatically-generated and subsequently human-editable wiki pages can contain multiple answers for a question, each answer supported by one or more justifications. In this way, end-users can decide for themselves while having opportunities for learning and self-reflection.

One or more automated decision-making systems – systems capable of weighing justifications and arguments – could interoperate with a "Wikianswers" platform. These systems could decide from and select certain answers from a set of justified candidate answers and annotate their selected answers. These annotations could be visualized by end-users in published wiki documents, e.g., with gavel symbols. Artificial intelligence decision-making systems should be able to explain their decisions and, so, these visualized annotations could be hyperlinks.

It could occur that a decision-making system might decide to abstain from concluding and, in such an event, it could annotate the wiki page such that a visual symbol could serve as a hyperlink to an explanation of why it abstained.

It would be convenient for users to be able to preview provisional annotations from decision-making systems while previewing their edits.

Each time that an update to a wiki page occurs, annotations from decision-making systems could be removed and decision-making systems could be notified to revisit the page to reevaluate any updated justifications or arguments.

Users might enjoy a new type of notification pertaining to the annotations of decision-making systems and the capability to receive email notifications when decision-making systems change their decisions about particular wiki articles.

Search and discovery

edit

The "Wikianswers" project could enhance search and discovery capabilities across Wikimedia projects, e.g., allowing users to enter natural-language questions into search boxes on Wikipedia.

Computer-aided document authoring

edit

For technical discussion about computer-aided document authoring, please visit: https://meta.wikimedia.org/wiki/Wikianswers/Technical_discussion/Computer-aided_document_authoring .

See also

edit

Bibliography

edit
  • Anderson, Michael, and Susan Leigh Anderson, eds. Machine ethics. Cambridge University Press, 2011.
  • Besnard, Philippe, and Sylvie Doutre. "Checking the acceptability of a set of arguments." In NMR, vol. 4, pp. 59-64. 2004.
  • Campbell, Richmond, "Moral epistemology", The Stanford Encyclopedia of Philosophy (Winter 2019 Edition), Edward N. Zalta (ed.), 2019.
  • Chen, Danqi, Adam Fisch, Jason Weston, and Antoine Bordes. "Reading wikipedia to answer open-domain questions." arXiv preprint arXiv:1704.00051 (2017).
  • Graesser, Arthur C., Vasile Rus, Zhiqiang Cai, and Xiangen Hu. "Question answering and generation." In Applied Natural Language Processing: Identification, Investigation and Resolution, pp. 1-16. IGI Global, 2012.
  • Hirschman, Lynette, and Robert Gaizauskas. "Natural language question answering: the view from here." natural language engineering 7, no. 4 (2001): 275-300.
  • Jiang, Liwei, Jena D. Hwang, Chandra Bhagavatula, Ronan Le Bras, Maxwell Forbes, Jon Borchardt, Jenny Liang, Oren Etzioni, Maarten Sap, and Yejin Choi. "Delphi: Towards machine ethics and norms." arXiv preprint arXiv:2110.07574 (2021).
  • Kobbe, Jonathan, Ines Rehbein, Ioana Hulpus, and Heiner Stuckenschmidt. "Exploring morality in argumentation." Association for Computational Linguistics, ACL, 2020.
  • Slonim, Noam, Yonatan Bilu, Carlos Alzate, Roy Bar-Haim, Ben Bogin, Francesca Bonin, Leshem Choshen et al. "An autonomous debating system." Nature 591, no. 7850 (2021): 379-384.