The search interface is made of three sections: Search, Explore, and
Results. These are described in detail below.
You may start searching either from the Search section or from the
Explore section.
Search
This section shows your current search criteria and allows you to submit keywords to search in
the bibliography.
Each new submission adds the entered keywords to the list of search criteria.
To start a new search instead of adding keywords to the current search, use the
Reset search button, then enter your new keywords.
To replace an already submitted keyword, first remove it by unchecking its checkbox, then
submit a new keyword.
You may control the extent of your search by selecting where to search. The options are:
Everywhere: Search your keywords in all bibliographic record fields and in the text content of the available documents.
In authors or contributors: Search your keywords in author or contributor names.
In titles: Search your keywords in titles.
In all fields: Search your keywords in all bibliographic record fields.
In documents: Search your keywords in the text content of the available documents.
You may use boolean operators with your keywords. For instance:
AND: Finds entries that contain all specified terms. This is
the default relation between terms when no operator is specified, e.g., a b
is the same as a AND b.
OR: Finds entries that contain any of the specified terms,
e.g., a OR b.
NOT: Excludes entries that contain the specified terms,
e.g., NOT a.
Boolean operators must be entered in UPPERCASE.
You may use logical groupings (with parentheses) to eliminate ambiguities when using
multiple boolean operators, e.g.,
(a OR b) AND c.
You may require exact sequences of words (with double quotes), e.g., "a b c".
The default difference between word positions is 1, meaning that an entry will match if it
contains the words next to each other, but a different maximum distance may be specified (with
the tilde character), e.g., "web search"~2 allows up to 1 word between
web and search, meaning it could match web site search as
well as web search.
You may specify that some words are more important than others (with the caret), e.g.,
faceted^2 search browsing^0.5 specifies that faceted is twice as
important as search when computing the relevance score of the results, while
browsing is half as important. Such term boosting may be applied to a logical
grouping, e.g., (a b)^3 c.
Keyword search is case-insentitive, accents are folded, and punctuation is ignored.
Stemming is performed on terms from most text fields, e.g., title, abstract, notes. Words
are thus reduced to their root form, saving you from having to specify all variants of a word
when searching, e.g., terms such as search, searches, and
searching all produce the same results. Stemming is not applied to text in
name fields, e.g., authors/contributors, publisher, publication.
Explore
This section allows you to explore categories associated with the references.
Categories can be used to filter your search. Check a category to add it to your search
criteria and narrow your search. Your search results will then only show entries that are
associated with that category.
Uncheck a category to remove it from your search criteria and broaden your search
results.
The numbers shown next to the categories indicate how many entries are associated with
each category in the current set of results. Those numbers will vary based on your search
criteria to always describe the current set of results. Likewise, categories and whole facets
will disappear when the result set has no entry associated to them.
An arrow icon () appearing next to a category indicates that subcategories are
available. You may press it to expand a list of more specific categories. You may press it again
later to collapse the list. Expanding or collapsing subcategories will not change your current
search; this allows you to quickly explore a hierarchy of categories if desired.
Results
This section shows the search results. When no search criteria has been given, it shows the
full content of the bibliography (up to 20 entries per
page).
Each entry of the results list is a link to its full bibliographic record. From the
bibliographic record view, you may continue exploring the search results by going to previous or
following records in your search results, or you may return to the list of results.
Additional links, such as Read document or View on [website
name], may appear under a result. These give you quick access to the
resource. Those links will also be available in the full bibliographic record.
The Abstracts button lets you toggle the display of abstracts within the
list of search results. Enabling abstracts, however, will have no effect on results for which no
abstract is available.
Various options are provided to let you sort the search results. One of them is the
Relevance option, which ranks the results from most relevant to least relevant.
The score used for ranking takes into account word frequencies as well as the fields where they
appear. For instance, if a search term occurs frequently in an entry or is one of very few terms
used in that entry, that entry will probably rank higher than another where the search term
occurs less frequently or where lots of other words also occur. Likewise, a search term will
have more effect on the scores if it is rare in the whole bibliography than if it is very
common. Also, if a search term appears in, e.g., the title of an entry, it will have more effect
on the score of that entry than if it appeared in a less important field such as the abstract.
The Relevance sort is only available after keywords have been submitted
using the Search section.
Categories selected in the Explore section have no effect on the
relevance score. Their only effect is to filter the list of
results.
Human evaluations are typically considered the gold standard in natural language generation, but as models' fluency improves, how well can evaluators detect and judge machine-generated text? We run a study assessing non-experts' ability to distinguish between human- and machine-authored text (GPT2 and GPT3) in three domains (stories, news articles, and recipes). We find that, without training, evaluators distinguished between GPT3- and human-authored text at random chance level. We explore...
Human evaluations are typically considered the gold standard in natural language generation, but as models' fluency improves, how well can evaluators detect and judge machine-generated text? We run a study assessing non-experts' ability to distinguish between human- and machine-authored text (GPT2 and GPT3) in three domains (stories, news articles, and recipes). We find that, without training, evaluators distinguished between GPT3- and human-authored text at random chance level. We explore...
Language model (LM) benchmarking faces several challenges: comprehensive evaluations are costly, benchmarks often fail to measure the intended capabilities, and evaluation quality can degrade due to labeling errors and benchmark saturation. Although various strategies have been proposed to mitigate these issues, they tend to address individual aspects in isolation, neglecting broader questions about overall evaluation quality. Here, we introduce FLUID BENCHMARKING, a new evaluation approach...
