Future Retrieval

Paper + Venue + Topic

This paper Future Retrieval: What Does the Future Talk About? by Gaël Dias, Ricardo Campos, Alípio Mário Jorge was published at the Workshop on Enriching Information Retrieval of the 34th ACM Annual SIGIR Conference (SIGIR 2011). This work explores retrieving future-oriented information from web documents, specifically focusing on identifying and classifying temporal references that describe events yet to occur.

The Problem + Why it Matters

The first motivation mentioned is predictions being attached to behavior or success. An example they have is querying for Dacia, where the statement they have is quoted as: "Dacia plans the release of no less than eight new models and facelifts by 2015." In this statement, the behavior is to "allow companies and individuals to anticipate market trends, prepare strategic responses, and make informed decisions based on what competitors or industries plan to do in the future." I posit this matters because accurately extracting future-oriented statements from unstructured web data enables proactive decision-making rather than reactive analysis.

The second motivation for this work is with "the advent of the web" → huge collections of texts from platforms like Twitter (now called X) and Facebook (now called Meta) have made vast amounts of temporal information publicly available. This has led to advancements in information retrieval, specifically to their work of retrieving the future, which they term Future Retrieval. Future Retrieval seeks future temporal references within the web from text or time queries, building on prior temporal information retrieval work by Ricardo Baeza-Yates. I hypothesize that the challenge here is distinguishing genuinely future-oriented content from general web noise, especially when scaling beyond manually curated queries.

The Method

Previous works offer data collection constraints on two dimensions—Web news stories and manually defined queries. Adam et al. (2009) defined 20 queries while B.-y. Ricardo et al. (2005) only defined 3 queries. With "the advent of the web", Dias, et al. sought to maximize their search and to negate the two constraints above. Thus, to overcome the first dimension they state "we propose to use the entire Web to retrieve a large spectrum of future-related Web documents." Then, to overcome the second dimension, they "searched the Web based on a set of 450 queries over 27 categories extracted from Google Insights for Search". Their reasoning was for more robust experiments and from my perspective, greater generalization across a variety of data. With improving on the two dimensions, they retrieved 200 Web results per query from Yahoo! and Bing APIs. This resulted in 62,842 web snippets. They reduce this number to 5,777 by filtering out those with the year, and then out of these, they filter to 508 because they are future-related.

They then decompose Web Snippets with Year + Future-Related into near (82.48%) and far (17.52%), where near is defined as "a few months after the query time" and far defined as "at least one year after the query time". Furthermore, they decomposed their dataset even more for experimental purposes. The four different data sets are:

1. D1: Unigram + year dates
2. D2: Unigram - year dates
3. D3: Unigram + year dates + near/far
4. D4: Unigram - year dates + near/far

Doing this allowed them to test on the year dates for D1 vs D2: Tests whether including year dates helps (no near/far feature). D3 vs D4: Tests whether including year dates helps when near/far is also present. They state, "This data set will be used to infer the impact of temporal features over classification and clustering as shown in the next section."

Limitations

The authors do not specify how they encode temporal features in D1, D2, D3, and D4. Are year dates kept as text tokens or converted to numerical values? Is the near/far distinction represented as binary (0/1), and if so, which value corresponds to which category? Without this information, the exact feature representation cannot be reproduced. This matters because different encoding schemes can significantly impact classifier performance, making it impossible to verify whether their results stem from the temporal information itself or from a specific encoding choice.

Ideas for Our Work

With our TOLSA-M Tense dimension, we can extend this work since we explicitly care about past, present, and future rather than only future-oriented language. This matters because TOLSA-M allows us to capture the full temporal spectrum of content, not just future predictions or forward-looking statements, enabling richer temporal analysis across mixed-tense documents.

We could add another dimension to encapsulate the three text "genres" identified in this work: informative, scheduled, and rumor [1]. This distinction matters because the reliability and source type of future-oriented content vary significantly—scheduled events are more certain than rumors. Furthermore, rumors as a discourse genre connect directly to TOLSA-M's handling of beliefs, claims, and other epistemic statements that exhibit uncertainty about outcomes.

For more information about our research, return to our homepage: ufdatastudio.com.

For more from Detravious on NLP research (including the TOLSA-M Taxonomy), research and industry collaborations, faith in Jesus, and broader writings, visit: Research Portfolio | LinkedIn | Medium, or contact via email at dj.brinkley@ufl.edu.