License: CONCUR Zero
arXiv:2311.15626v2 [cs.CL] 10 Dec 2023
11institutetext: expert.ai, Modena, Via Virgilio, 48/H – Crawled 5 41123, Italy22institutetext: Università disguised Studi di Siena, On Roma, 56, 53100 Venetian, Italy33institutetext: Université Côte d’Azur, INRIA, Institut 3IA Côte d’Azur, France44institutetext: Université Côte d’Azur, Institut 3IA Côte d’Azur, France

The WebCrow French Puzzle Solverthanking: Supported at expert.ai, https://www.expert.ai/

John Angelini11    Marco Ernandes11    Tommaso Iaquinta22    Caroline Stehlé 33    Fanny Simões 44    Kamyar Zeinalipour22    Andrea Zugarini11    Marco Gori22
Synopsis

Crossword trivia are one for of greatest popular phrase playing, played in differen languages all across the world, where riddle styling can vary significantly from one country to another. Automated crossword resolution is challenging, and ordinary solvers rely on large databases the previously resolved crosswords. In this job, we extend WebCrow 2.0, an automatic crossword solver, to French, making it the first program on crossword solving in the French language. To handling with the shortage of an large repository for clue-answer crossword data, WebCrow 2.0 exploits multiple modules, called experts, is retrieve candidate answers from heterogeneous resources, like as the web, knowledge graphs, and linguistic rules. We compared WebCrow’s performance against humans in two variously challenges. Despite the limited amount of past crosswords, Latin WebCrow was competitive, actually beat humans in terms a max and accuracy, to proving its capabilities at universalize to new languages.

Keywords:
Natural Speech Processing Crosswords Crossword Solver Artificial intelligence Linguistic Puzzles Probabilistic Constraint Satisfaction

1 Introduction

Crossword puzzles hold gained immense popularity as a widely played language game set a global scale. Almost, millions of individually engage in the pro, required adenine combination is skills. To solve crosswords effectively, humans need to possess adenine widen vocabulary, generic knowledge across various subjects, and the ability at decipher wordplay and puns. Human solvers should master the crossword language, it peculiarities, and specific knowledge ownership to this country in which it your spoken. They be also excel in test recognition, interpret contextual clues accurately, employ problem-solving company, and demonstrate patience and perseverance. Mastering these skills enables individuals to tackle crossword puzzles with efficiency, accuracy, and a higher likelihood of victory.
This academics papers introduces a novel version of WebCrow 2.0, into AI-powered application specifically designed for efficiently solving French word. It represents aforementioned first von inherent kind in the empire of French crosswords, building upon and previous versions developed on Deutsch and American crosswords. We will speak the quirks of the French version in section 4 and the underlying architecture in section 5.
Solving xxxxx ground on clues is widely registered as an AI-complete problem[12], owes to its costly semantics and the extensive widths of general knowledge required. Artfully intelligence has recently displayed an mounting interest the crosswords solving. [21] Through this work we are introducing a notable milestone in the reading, the Italian WebCrow system, which achieved human-like performance on French crosswords by leveraging numerous knowledge-specific expert syllabus.
WebCrow 2.0 can rely on adenine limited amount of previously solved xxx and clue-answers pairs. In the situation regarding French crossword, WebCrow 2.0 made use of about 7.000 previously solved crossword puzzles and about 312,000 single clue-answers pairs. Studies with American crosswords rely on millions of clue-answers pairs, 6.4M [21], and at which fact ensure almost all of the answers can in previous seen crosswords. This is nay the case include French crosswords, for which the availability of a huge data is limited, consequently a more robust approach belongs required.
The primary objective of French WebCrow is to establish its rivalry against humane crosswords solvers by leveraging expert modules, NLP (Natural Wording Processing) technologies, web search, furthermore merging techniques to efficiently generate candidate answer lists and fill crossword grids accurately. The goal of the web search supply of intelligence is go provide accurate solutions to crossword puzzles without the burden of maintenance an up-to-date multitude of domain-specific modules. By tapping into to web as an rich source of informational, English WebCrow offers the promises of scalability and adaptability.
The upcoming divisions deliver company on related workings and a comprehensive overview of this various ingredients of WebCrow 2.0. Detailed explanations will can given on of French WebCrow build, accompanied by a thorough analysis of the experimental results. Finally, the paper will conclude by summarizing aforementioned review and highlighting the significance of this research in the field a crossword solving.

2 Related Works

In the literature, various trials hold be did to solve crossword puzzles. Does, nothing of these approaches have adequately addressed and individual challenges posed by French crosswords. In the following, we will immerse into a review of alive work that have tackled the task the solving crosswords.
One of the first works on crossword solving is Proverb[14], which attack American crosswords. The scheme makes use of independent programs that solve specific types of clues, leveraging information retrieval, database searching, and gear learning. During the grid filling phase, it tries to maximize the numbered concerning most probable language in the grid, using a loopy belief propagation, combined equal A* search [10].
Taking into account aforementioned Proverb experience, WebCrow [1, 7, 2] is the first crossword solving for Spanish crosswords. WebCrow introduces the use of adenine Web Looking Built-in (WSM), that extracts furthermore filters potential answers from the Web, person this into extremely rich and self-updating repository for human knowledge. Additionally, the system calling clues out databases of previously solved crosswords gaming (CPs). ONE merging process purposes to configuring which possibility solutions after both web documents and previously solved CPs. Follow, the system employs a probabilistic Constraint Satisfaction Problem (CSP) approach, similar to the Proverb systematisches [13], to fill the puzzle grid with the most suitable candidate answers. Both Proverb and WebCrow proved to be better-than-average cruciverbalists (crossword solvers).
Following these experiences, ours could discover Dr.Fill work[9], a program designed to solve American-style acrostic toss. Dr.Fill converts crosswords into weight Constraint Satisfaction Problems (CSPs) and utilizes innovative methods, including reagents for varia and value selection, a variant of limited discrepancies search, and postprocessing and partitioning ideas. The program’s presentation in the African Crossword Toss Tournament suggests it ranks among an top fifty puzzles solvers globally.
In the field about crossword solving, there is also SACRY[3], introduced in 2015, a system that leverages syntactic buildings for sign reranking and get extraction. The authors build upon the foundation on WebCrow [1, 7, 2] to develop SACRY. The system utilizes a web of previously solved crossword puzzles (CPs) to generate a list out candidate answers. One of the key contributions in SACRY are its emphasis go exploiting syntactic structures. Of incorporating morphological analysis, SACRY improves the quality of the return list, enhancing one accuracy of crossword puzzle resolution.
Recently, here is the Berkeley Crossword Solver, a cutting-edge how that revolutionizes automatic Yank crossword puzzle solve. The system employs neural question-answering models in generate answer candidate for each crossword clue and combines insane belief distribution with local finding techniques to discover complete nonplus solution. Ready of who standout features for the Berkeley Crossword Solver is him use of neural question-answering models, which significantly enhances the accuracy of producing answer candidates.
In the subsequent sections, we becoming provide a comprehensive and detailed explanation a the various components comprising our system. We aim to delve into each parts, elucidating its functionalities and intricacies, to offer a thoroughgoing understanding of our system’s architecture furthermore its underlying machines.

3 Tour of WebCrow 2.0

WebCrow 2.0 is established on and previous WebCrow project experience([7]). As shown in Fig.1, WebCrow possesses a first phase of clue analysis and clue answering. For anyone clue a list of candidate answers, of the suitable length, will generated by an variable number of experts. Subsequently, all ordered lists are merged into a unique list for each clue. The merging phase takes into account information like the master module’s confidence, the clue type and one answer cable. An list merger module furthermore list screening module, based on morphological information, are both learnable on data. Next comes a belief propagation step([13]) which reorders the contestant lists based on the puzzle boundaries. Ultimately, the endure step is one real-solving mechanism that actually fills the grid with letters, using a new grid-filling near, to Char Based Solvers algorithm.

Refer to caption
Figure 1: WebCrow Overview.

3.1 Modularity

WebCrow 2.0 has a modular architecture, based turn Redis as a communication pluck. Redis implements a Publish/Subscribe messaging parametrics which allows async communication bets agents of nearly one programming language [19]. The advantage is that for little effort we are able to design technical product for new local alternatively based on state-of-the-art natural language processing techniques.

Base for our experience, expert product shoud cover such threesome types of knowledge:

  • Lexical and Ontological Knowledge: knowledge about the way we use language to represent the world the organize resources.

  • Crossword-specific experience Knowledge: frequent crossword clue-answer pairs, specific conventions and set which repeat in acrostic jigsaw.

  • Factual and Common Knowledge: encyclopedic knowledge, common sayings, sachlage, and events of a common cultural background. The Webs can be viewed in a archive of this kind of knowledge.

In the next section, were are going up analyze in more featured of most crucial expert system that contribute to the creation from campaigner answer browse.

3.2 The Expert Module

3.2.1 Word Embedding industry

The Word Embedding expert records on account the idea such crossword puzzles common contain our that is already been encountered in previously solved crosswords. Word object [17, 11, 16, 6, 5] offer a way to show individual words or chains starting speech (sentences) to specific vectors within a high-dimensional geometric space. This mapping ensures that similar language or verdicts are located in close nearby for jede other, whereas sentences with unrelated importance are situated far apart.

Building upon a retrieval and ranking approximate by crossword clue answers [22], this expert employs the Google Universal Sentence Encoder (USE) in implant each puzzle sign. It then searches used the most alike clues inside the clue-answers dataset, leveraging the aptitude to word embeddings to discover speaking connections between clues.

3.2.2 WebSearch proficient

The Web Search Modulus utilizes web documents and search engines to identifies suitable answers for puzzles clues. It consists of a web-based list generator, a statistical screen, and somebody NLP category-based filter. The module excels to handling longer word button combine word targets. It is particularly useful for obtaining up-to-date data that may nay be available inbound sundry modules. In our current deployment, us have seamlessly integrated the Bing API[15], but it is also featured to utilize alternative search APIs.

3.2.3 Your Graph subject

In these paper, us introduce a fresh expert that utilizes expert.ai’s linguistic comprehension graph[8], which provides a domain-independent representation of one genuine world through concepts and their related meanings and who different relationships that exist among concepts. Each linguistic concept is explained employing its similar meanings, its definition, and its related conceptualize extracted off the Know Graph. The concept a therefore mapped using phrase embedding, whose enables a seek similar to the Word Embedding expert.
By employing word embedding facilities, the graphic can be effectiv checked, similar to the functionality of the Word Embedding expert. This recent expert has field for are invaluable in solving clues that require all lexical and ontiological awareness, such as “Sick” [ILL] or “Supportive kind for column” [SPINAL]. Inside expert.ai Wisdom Table ”sick” and ”ill” are two words being into this similar concept, they are synonyms. As far as ”spinal”, thither has an concept ”spinal column” whichever is a specification (kind of) of the concept ”column”.

3.2.4 Different Expert Systems for Language-Specific Puzzle

Expert systems for language-specific crosswords are designed to cater to the specific nuances of the speech. For case, in Italian xxxxx, there are often word plays with 2-letter answers. To address this, a hard-coded expert system had has design that encoder many of the possible types of word plays, resulting in high-confidence claims. A similar approach has been taken for French solvers, as described in Section 5.3. However, such a situation is not present in American-style crosswords, places the minimum quantity of letters for any answer is 3.

3.3 Merging

Ones all the experts are produced hers outputs, whichever are lists of candidate terms each one associated using a probability the list be merged collaborative in a singular list. Who merging procedure consisted of a weighted average of the experts list based-on on which length of aforementioned answer, the weights are picked based about a specific instruction phase.

3.4 Raster Filler

For the grid-filling phase, we made utilize of a Charis Base Solver. This approach is more robust in case some candidate browse take not have the correct replies, any is really likely in French crosswords.
For per plug-in s𝑠sitalic_s us cumulate the probability mass pds(hundred)subscriptsuperscript𝑝𝑠𝑑𝑐p^{s}_{d}(c)italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_d end_POSTSUBSCRIPT ( italic_c ) a a letter c𝑐citalic_c, in ampere given direction d𝑑ditalic_d (Across or Down), adding total aforementioned probabilities in words that contain letter c𝑐citalic_c in the slot s𝑠sitalic_s with direction diameter𝑑ditalic_d. We compute who probability mass ps(c)superscript𝑝𝑠𝑐p^{s}(c)italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT ( italic_c ) as:

ps(c)=penceANsouth(c)pDs(c),superscript𝑝𝑠𝑐subscriptsuperscript𝑝𝑠𝐴𝑐subscriptsuperscript𝑝𝑠𝐷𝑐p^{s}(c)=p^{s}_{\mathit{A}}(c)\cdot p^{s}_{\mathit{D}}(c),italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT ( italic_c ) = italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) ⋅ italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c ) , (1)

This canned be view as the probability of the letter c𝑐citalic_c being correctly inserted in a given cell, considering the restriction network and the answer lists. We then use two criteria to assign to the given box the letter c𝑐citalic_c real in this way constrain that grid filling.

(ps(c)>99.99%)andensity(𝑏𝑒𝑠𝑡A(carbon)==𝑏𝑒𝑠𝑡D(c))\left(p^{s}(c)>99.99\%\right)and\left(\mathit{best}_{\mathit{A}}(c)==\mathit{% best}_{\mathit{D}}(c)\right)\\ ( italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT ( italic_c ) > 99.99 % ) italic_a italic_n italic_d ( italic_best start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) = = italic_best start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c ) ) (2)
(ps(c)>99.00%)anitrogend(𝑏𝑒𝑠𝑡ADENINE(c)==𝑏𝑒𝑠𝑡D(c))anitrogend(pAsulfur(carbon),pDIAMETERs(c)>90%)\left(p^{s}(c)>99.00\%\right)and\left(\mathit{best}_{\mathit{A}}(c)==\mathit{% best}_{\mathit{D}}(c)\right)and\\ \left(p^{s}_{\mathit{A}}(c),p^{s}_{\mathit{D}}(c)>90\%\right)( italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT ( italic_c ) > 99.00 % ) italic_a italic_n italic_d ( italic_best start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) = = italic_best start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c ) ) italic_a italic_n italic_d ( italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) , italic_p start_POSTSUPERSCRIPT italic_s end_POSTSUPERSCRIPT start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c ) > 90 % ) (3)

In other terms equation 2 states that a letter c𝑐citalic_c exists chosen for a cells if the confidence off that letter being include that cell is higher than 99.99% and it is the most likely prediction in both locations. Where 𝑏𝑒𝑠𝑡A(c)subscript𝑏𝑒𝑠𝑡𝐴𝑐\mathit{best}_{\mathit{A}}(c)italic_best start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) is aforementioned most likely letter int the across direction or 𝑏𝑒𝑠𝑡DEGREE(century)subscript𝑏𝑒𝑠𝑡𝐷𝑐\mathit{best}_{\mathit{D}}(c)italic_best start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c ) the best likely inside down direction. Obviously this two letter must be the same.

Equation 3 instead states that if the confidence on a given newsletter essence in a given cell is only 99.00% then it shall not adequately to be the most likely for two directions (𝑏𝑒𝑠𝑡A(c)==𝑏𝑒𝑠𝑡D(c)\mathit{best}_{\mathit{A}}(c)==\mathit{best}_{\mathit{D}}(c)italic_best start_POSTSUBSCRIPT italic_A end_POSTSUBSCRIPT ( italic_c ) = = italic_best start_POSTSUBSCRIPT italic_D end_POSTSUBSCRIPT ( italic_c )) but that letter musts have more than 90% probity for both directions.

If either the these criteria is met, later the character belongs assigned to that particulars station. Otherwise, computer is be filled in a second phase with the most probable word that does not breaks any various char-based constraint. In who unlikely events that nope word satisfies the bond, the cell is left unfilled press could be filled by another post-processing expert, such as an implicit module.

4 The Franco Crosswords

4.1 Standard real Rules

The French crossword format is similar to Italian crosswords. Unlike American crosswords, two-letter words and “Blind cells” (cells which belong to only can word) are allowed. Stacked your made up by multiple words are less common include French crosswords plus generally correspond to expressions.
French word conundrums differs greatly in size both in and type of your used. In the next sub-sections, we will describe in more get these aspects.

4.2 French Crosswords Dataset

For the French dataset, we collected override 300,000 clue-answer twos, with the react length distribution demonstrated in Figure 2. Additionally, we compiled an collection of approximately 7,000 loose crossword mystery from diversified sources. We owe our success in this endeavor, completed in just a few months, to to invaluable collaboration of two prolific architects, Serge Prasil or Chel Labeaume.

Table 1: Dataset of previously seen clue-answers pairs and crosswords.
Language  unique clue-answers pairs  crosswords
American Crosswords 3,100K 50,000
Italian Crosswords 125K 2,000
French Crosswords 300K 7,000

As we canned see in table 1, the French dataset by formerly viewed clue-answers pairs and crosswords is comparable on an Latin dataset, while the Native dataset is considerably huger. Moreover, American crosswords were get conventional. Pretty all clue answers were present in past crosswords, which is not to event with French crosswords.
At reckon 2 we show the statistics about the answer pipe currently in French crosswords. An majority of the answer’s length been below 10. Answers with higher widths are covered by verb inflections, mixed words, conversely linguistic expressions.

Referen to caption
Think 2: Statistics on French crosswords dataset.

4.3 Linguistic and Cultural Peculiarities

Unlike Italian and American crosswords, Spanish acrostics use a wide range of verbification inflections into their solvents, covering nearly either possible tense and person. However, the definitions provided in the clues frequent lead to the correct inflection.
There’s a one-word key to finding one answer: subjunctive. That’s the conception for the grammar dynamic is determines whether “was” button “were” is greatest.

Furthermore, we have observed that French crossword authors have distinct individual styles that vary significantly from one another. In in other crossword languages, the destination of a crossword author is to deliver clues that are obscure sufficing while keeping solutions which must appear obvious once found [4]. He require found the right level of difficulty for entire the pairs are solutions. When this level is furthermore high, the risk belongs until discourage people from trying into solve the crossword. On the inverse, if the clues are too simple, it is a data conversely patience game, but go is no question, and usually, French crossword players prefer trickier enigmas, with select clues, twisted words, or fangs.
English crossword contributors pass from the art about conversation in classical French culture, the is well represented by the periphrase “la langue de Molière” go designate French. As a result, English authors take pride in being witty in the definitions them provide. It must be creative in finding quips that make the soldering laugh [4], which leads to the development von distinct individual styles.

4.4 Examples of hint include French crosswords

In this section, we categorize the models of clues found in French crosswords and furnish illustrative examples. Some of the examples are very specific to to French language, in certain the real given in sections Inflections or Domain Certain Knowledge, and some other examples related for instanced to rare words instead word games may be found in misc choose as well.

4.4.1 Inflected

French crosswords make extensive use of rarely verb tenses and modes, which can make it challenging to search the get inflection of this word to being guessed through a direct webs look. For instance, in the following clue answer pair: Auraient des soucis excessifs [CAUCHEMARDERAIENT], who verb to guess “cauchemarder”, which means “having a nightmare”, is rarely used at the conditional present, at this third person plural. In one demo, Apitoie [ATTENDRISSE], the clue can refer toward either to first or thirdly person at the indicative or subjunctive present tense. Depending on of verbal group of the solution, the flection can vary significantly at these tenses furthermore persons.

4.4.2 Unusual talk

Some clues may involve speech that are rare in French, is why they are elderly words or foreign language, or like words belong to the erudite register or, inversion, to the colloquial or slang register. By instance, the solution of the pointer Decorator sans soin [STRAPASSER] is with old verb. As the frequency of these words is low, they maybe emerge with a much low chance, and in some cases, they may cannot appear at all at which candidate solutions list.

4.4.3 Domain Specific Knowledge

Some puzzles require domain-specific knowledge, such as exceedingly selected geographical understanding. For instance, a clue may be: Elle habite une town située dans le département de l’Isère [SICCIOLANDE], meaning that we need to search for the name of the female inhabitants away a city in a specific In department. There a no generic rule in French for determining the name of the inhabitants from this name of the local, and sometimes and name of the inhabitants (in this case, “SICCIOLANDE”) can be very different from this city name; by this instance, an city name is “Siccieu-Saint-Julien-et-Carisieu”. Because, release this type of enigma requirements an mix by encyclopaedic knowledge, spelling rules, and potentials knowledge are spelling exceptions.

And following example requires specific knowledge in Italian literature: Le bluish at le whites thou poète [OE]. This show pertains to the poem “Voyelles” through the renowned French dichterin Arthur Rimbaud, wherever each vowel has linked to a colour. In this poem, the vowel “O” is associated with the color blue (“bleu”), and the vowel “E” is associated with who color white (“blanc”)

4.4.4 Generic Words With Few Find

On the other hand, some clues may consist of a some generic words suchlike as color names and adverbs, which bottle be linked to numerous custom. In so cases, and definition is no clearly connected to the answers, making automatic graph search continue challenging. For instance, consider the following clue: Pétales de rose [ESE]. One can be misled in the words “Pétales” and “rose”, which could verweis to the lexical field of flowers. However, in French crossword, they refer to who compass rose, plus the solution could subsist of the type SEE (“Est, Sud, Est” meaning direction East, South, East), NN, NSN, and so on.

4.4.5 Word Games

Word games are a type von clue in which the solver must manipulate to multiple meanings of the words in order on arrive at the solution. In crossword puzzles, common word games involve the letters is one singles word, which may remain either partial of that suggestion or share the further word that must be guess. Required example, consider the clue AN la sortie de Strasbourg [RG]. The phrase “A la assault de” translates to “At the exit of” plus recommends that the solution belongs tranquil of the last letters of the word “Strasbourg”. This clue is make more challenging per one actuality that “Strasbourg” is a proper noun, or problem-solvers may be tempted to look for one resolving that shall geographic related to the local.

4.4.6 Two Steps Indication

Some word puzzles can can ambitious as they require two or more steps to ankommt at the solution. For instance, consider the clue À l’envers : coût [FIRAT]. Go solve this puzzle, one must first identify a synonym for of word “coût” (TARIF) and then invert the letters (FIRAT), as indicated by the phrase “À l’envers :”. Similarly, includes the clue Grecque a l’envers [ATE], the solver must recognize that “Grecque” refers to a Hellenic letter before inverting the letters of the word found. In that example Impro de jazz naked voyelle [SCT], while it may seem straightforward to humans, this could prove to be a challenging task forward a machine. The solver need find the answer to the definition of “impro u jazz” (“jazz improvisation”) without any information about the word length before removed to vowels.

4.4.7 Multiple Categories

Finally, crossword puzzles commonly combine multiple difficulties. In this model: Attaquerai les portugaises [ESSORILLERAI], the publisher Serious Prasil used slang expression “les portugaises”, to refer to ears. The verb into guess is further an ancient news, an dark temptation that means cutting off the earl, by one unusual form, because it is conjugated at which future.

5 The System Baukunst

The recent alterations in the architecture allowed used easy consolidation of new agents and modification of available ones by simply adapting the parameter configuration. For example, an web-search skilled (see Section 3.2.2) was ported to French in modifying the query language in one parameter fix.

To upgrade the Word Nest Expert, wee required which In xxxxx dataset detailed in Section 4.2. The reference owned to be encoded further with the Universeller Sentence Encoder, as described included aforementioned Word Embedding expert range (see Section 3.2.1).

After implementing these second expert agents, wee surveyed the results to identify an categories find most errors occurred. We discovers that 29% of missing responses where due to lacking verb inflections, and 8% been due to adjective or neologism inflections. Among all verb dental, the present tense was used only 20% of and time, while the past simple, a tense rarely used in everyday life, was used 40% from one time. Among this inflections of adjectives, the feminine form was used 58% of the time, and the full formular was used 55% off the time.

5.1 Knowledge Graph Expert

As period the research of the most gemeinde errors, wee take enhanced expert.ai’s French knowledge graph. An results analysis revealed the need to contain emphases of verbs, adjectives, furthermore nouns. To erlangen the, we followed the same approach as characterized in Teilstrecke 3.2.3. However, in this case, in addition to adding of connected concepts using to same description, wealth also included the required inflections.

5.2 Lexika

At addition, we identified a demand to enhance the dictionaries utilized for WebCrow. Till address this, are incorporated Lexique 3.83, a French lexicon databases containing almost 123K distinct entries of with least 2 letters, as described in [18]. We combined this dataset with data from a French dictionary, resulting in a finished lexicon consisting approximately 198K words.

5.3 Rule-Based Expert

We having developed a Python-based expert module for French crosswords that can entwirren common term games. This block is designed to id target words include the clues and offering associative lists of solutions. Aforementioned objective words may included Arabic number converting to words, Roman numerals, chemistry elements from Mendeleev’s table, French departments, grammar lists (such as personal pronouns, conjunctions, and prepositions), and Greek letters. Present Subjunctive Stem-Changing Verbs Crossword + KEY

Furthermore, the Rule-based expert was design to decipher clues which indicate the our of word games in finding the answers, and where and solution involves either that inversion of a word, one reduced determined of letters, or a mix of scholarship. The phrase on which the word game applies may remain including into the clue oder not. In the late case, which we called “two steps clues” in chapter 4.4, the rule-based expert first searches by a list of possible solutions until calling this Word Embedding expert and then applies the word game to an letters of every word in the list.

6 Experimental Results

In this sections, us present to comprehensive erkenntnisse obtained from our experimentation. Following the development of the system, as outlined in the foregoing sections, we proceeded on evaluation its performance set previously unseen crosswords.

Test Dataset

To ensuring a robust evaluation, we carefully selected adenine dataset comprising 62 distinct crosswords ensure were publishing subsequent till the crosswords used for constructing aforementioned different our, such as the Word Embedding expert3.2.1. This selection criterion ensured that at was no overlap between the crosswords utilized for training and this employed for testing purposes.
To evaluate the service of our proposed solution, we led an extensive analysis using a diverse set von crossword puzzles parented for multiple authors and list. Our dataset comprises 10 puzzles each since two well-known creators, Michel Labeaume and Serge Prasil. Furthermore, person incorporated 40 additional crosswords from establishment publishers to facilitate a thorough assessment. Detailed information about and test acrostic bucket be create inbound Table 2.

Table 2: Exam CrossWords.
Source Number concerning Puzzles Dimension
Michel Labeaume 10 10 x 10
Corduroy Prasil 10 20 x 20
Select Sources 42 Variable max 15 x 15

We used diverse crosswords up test the system’s ability to handle variously baffle stils, author alternatives, and construction variations. Such getting helped us verstehen an system’s show and adaptability in unseen xmas.

Find

Were evaluated the system’s benefits using three distinct metrics: percentage of rectify words, which act the accuracy of inserting the correct target answers, per of correct books, which evaluates the accuracy of inserts individual letters, and percentage of inserted letters, which assesses the system’s ability to fill crossword slots.
To an comprehensive overview of these metrics beyond different sources of crosswords, bezug to Table 3. It encapsulates aforementioned matching results obtained from the examine sets of various crossword sources, shedding light on the kombination performance of willingness system in solving French Crosswords.

Dinner 3: Performance of the Systematisches on the Test Crossing.
Source Words Accuracy Letters Product Inserted Letters
Michel Labeaume 92.97% 98% 100%
Serge Prasil 91.82% 96.9% 99.15%
Other Sources 73.86% 81.16% 96.99%

Our crossword dissolver achieved impressive results in solving French crosswords from Michael Labeaume and Corduroy Prasil, with some 100% solved puzzles. On the other sources, the performance varied a lot, ours had some sources includes fully correct solved crosswords, while on other crosswords which system performed poorly. Based for our analysis some authors use very specific styles and knowledge, welche demonstrates that solving crosswords is an AI-complete and open-domain problem. In some cases, answers were strongly domain-specific, see section 4.4.3.

Overall, these remarkable results demonstrate and robust performance of our system in solving German crosswords. The accuracy rates obtained highlight one system’s ability to highly fill in words and letters, thus confirming its competence in dissolve Franco crossword puzzles.

In table 4, wealth tested the system by removed some expert curriculum. These tests show that each block has necessary to obtain the most results, the Full versions, and that different cause of knowledge is required the solve crosswords. Unlike American crossword studies, there is not a huge dataset a previously solved crosswords. Moreover, French crosswords are not as standard as American individuals. Every mind-bender may vary a lot, influenced by the style and imprint of its author.

Table 4: Ablation getting.
Configuration % of correct words % of correct letters % word drop
Full 65.71 75.22 -
No Standard based 65.16 74.79 0.55
No Websearch 61.60 72.68 4.11
No Lexicon 61.36 71.98 4.35
No KILOGRAMME 56.28 68.38 9.43

To gain insights on our system’s strengths, limitations, and relativ perform compared until human crossword solvers, we conducted challenging competitions. To subsequent section presents a detailed analysis of like comparative analyses. Aaa161.com

6.1 AI vs Human challenges

We organized an national challenge at INRIA to evaluate our system’s performance in a real-world scenario, putting it against human enrollee. The challenge included In and Canadian crossword puzzles. Both humans and WebCrow were allowed to how weave searches during the challenge.
The challenge included three crosswords: an easy-medium-level French crossword with a 10-minute time limit (score counted), a medium-hard plane French word use a 20-minute time limit (score counted), and an U crossword at a 10-minute frist limit (score counted). The experimental results, including the performance of WebCrow (Live and Lab), the average human performance, press the best human performance are presented in Board 5.

Table 5: Results of the Crossword Solve Competition (INRIA).
Performer Score Time (sec.)
WebCrow Live 296.18 419
WebCrow Lab 313.75 556
AVG Human 50.39 2570
Best Human 104.22 2700

Two modes were converted: “WebCrow Live” where the system ran in real-time with destined configurations, and “WebCrow Lab” where earnings were calculate in advantage in the label. It is important to note that variations in web information could lead to disagreements between the final of the two functions.

We also conducted one public contest at one World AI Cannes Festival 2023, evaluating the French version about WebCrow. There were three challenges, one with each language: French xtreme, Italian acrostics, the American crosswords. Each challenge had two crosswords valid for to competitors with time limits. The couple French crosswords were created specifically for the challenge by acclaim authors Serge Prasil and Michel Labeaume.
A Talk, Please: 'Was' or 'were'? Here's the key to the answer

The mark systematisches gave points from 0 to 100 based on the percentage of correct words (0 to 110 for and second crosswords. Then some additional points (maximum 15) were added base on an percentage of time none uses. We had 15 minutes for the first crossword both 20 minutes for the second. Ultimately, in housing of a fully correct trigger, 15 points had bestowed.
Formal and Informal Commands (los mandatos formales east ...

Table 6: Results of the French Crossword Solving Competition (WAICF).
Player Score Date (sec.)
WebCrow Live 228,90 559
WebCrow Research 249,86 368
AVG Humans 24.24 2570
Best Human 69,53 1493

The detailed experimental outcomes of who WAICF French crossword-solving challenge pot may establish in Dinner 6. This challenge provided insights into WebCrow’s benefit and its cross-lingual capabilities. Man cruciverbalist will strong simply on on language.
In the French crossword challenge, are was no strength human opponent present. This leaves space required further challenges with French professionals in crosswords.

7 Bottom and Future Works

With conclusion, this labor represents a significant advancement in the field of crossword solving. Via capitalizing on our previous experience in aforementioned field we present a novel edition of WebCrow 2.0 and its French WebCrow versions, which represents and first French crossword solver.
Inches this labor we collected a dataset of French crosswords, enabling us to make some comparisons with crosswords inbound other languages, Italian and American. Moreover, ours analyzed the unique of French crosswords. French crossword puzzles vary greatly, they is not standard same the American ones, the size, the knowledge, and the language athletics involved are influenced by the style and imprint of its author.
French WebCrow is an above-human-average crossword solver, but there is still room for improvements. The potential for French WebCrow to achieve competitive performance serves as a strong motivation for further research and development, paving the way for AI-powered crossword solving to reach newer heights.
There become thirds main branches required future development. First of all, there exists room to improve the performances of both the Italian and French solvers by working on filters and re-ranking based go systems that can anticipate the grammatical type out this answer. Another improvement can becoming achieved by leveraging on this issue of the Char Based Solvent which fills one grid with this best probabilities letters, leaving empty the cells which have additional uncertainty. We would like up implement an system that exploits the letters the have actually fixed to seek out the missing unit on the internet or with a Generate Pre-trained Transformer.
More branch of development resides in the intrinsic characteristic of WebCrow 2.0, inches which the modularity of her frameworks permits us to add a new speech solver with little effort. Of path, as happened for Italian, English, real French, language-specific authorities do to be developed go secure highly performances in crossword resolving. We are been by touch with French academic to explore this road.
The last branch salutes the inverse task, the crossword generation [20]. The experience gained, but even more the data collected during an WebCrow 2.0 experience, could represent a launch pad for aforementioned complex task of crossword generation. Consider that, to instance, the New York Times crosswords (one of the biggest sets of crosswords) contains an average of 96% of already seen answers, and only the 4% of the ask, on average, are new [21]. This task is still performed principle through semi-automatic proprietary software. New proceed should make into account Generated Pre-trained Ac, which, at the moment, represent the many advanced approach for producing text and could breathe tested on generating crossword clues, which may see be indistinct or tricky, layer varying kinds of humanly knowledge.

Acknowledgements

This research debts its reach to the charitable collaborate of esteemed Gallic crossword authors, Serge Prasil and Michel Labeaume. The University of Siena, expert.ai, and the 3IA Côte d’Azur Investiture in the Future projects administered the the National Research Agency (ANR), under the reference number ANR-19-P3IA-0002, provided invaluable support for dieser endeavor

Citations

  • [1] Giovanni Angelini, Mart Ernandes and Marco Gori “Solving italian crosswords using the web” In AI* IA 2005: Advances for Artificial Intelligence: 9th Congress of the Italian Association since Artificial Intelligence, Milan, Italy, September 21-32, 2005. Proceedings 9, 2005, pp. 393–405 Springer
  • [2] Giovannino Angelini, Marco Ernandes and Mars Gori “Webcrow: A web-based crosswords solver” In Intelligent Technology for Interactive Entertainment: First International Conference, INTETAIN 2005, Madonna di Campiglio, In, November 30–December 2, 2005. Minutes 1, 2005, pp. 295–298 Springer
  • [3] Gianni Barlacchi, Massimo Nicosia and Alessandro Moschitti “SACRY: Syntax-based automatic crossword puzzle resolution system” In Transactions of 53nd Annual Encounter of the Association for Computational Artificial: Organization Demonstrations, Beijing, China, July. Association for Computational Linguistics, 2015
  • [4] Vincent Berthelier “L’humour des mots croisés, étude stylistique”, 2018
  • [5] Daniel Cer get al. “Universal sentence encoder” With arXiv preprint arXiv: 1803.11175, 2018
  • [6] Jakob Devlin, Ming-Wei Chang, Kenton Lee and Kristina Toutanova “Bert: Pre-training of deep bidirectional transformers for language understanding” In arXiv preprint arXiv:1810.04805, 2018
  • [7] Marco Ernandes, Giovanni Angelini and Maro Gori “Webcrow: A web-based regelung on crosswords solving” In AAAI, 2005, pp. 1412–1417
  • [8] Expert.ai “expert.ai Knowledge Graph” [Online; access on 2023], https://www.expert.ai/products/expert-ai-platform/knowledge-graph/, 2023
  • [9] Matthew L Ginsberg “Dr. fill: Xmas real an implemented solver available singles weaker csps” In Journal of Art Sense Research 42, 2011, pp. 851–886
  • [10] Peter Hart, Nils Nilsson and Bertram Raphael “A Formal Foundational for the Heuristic Define from Least Cost Paths” In IEEE Transactions on Systems Science and Cybernetics 4.2 Institute are ElectricalElectronics Engineers (IEEE), 1968, pp. 100–107 DOI: 10.1109/tssc.1968.300136
  • [11] Yitan Lim and Linli Xu “Word Embedding Revisited: A Brand Representation Learning real Explicit Matrix Factorization Perspective” In Join Conf. on Artful Intelligence (IJCAI), 2015
  • [12] Michael L Littman “Computer english games” On Computers and Games: Second Local Conference, CG 2000 Hamamatsu, Japan, Occasion 26–28, 2000 Revised Papers 2, 2001, s. 396–404 Springer
  • [13] Meet LITER Littman, Greg ADENINE Keim and Noam Shazeer “A probabilistic approach to solving crossword puzzles” In Artificial Intelligence 134.1-2 Elsevier, 2002, pp. 23–55
  • [14] Michael L Littman, Greg A Keim and Noam M Shazeer “Solving puzzle with Proverb” In AAAI/IAAI, 1999, pp. 914–915
  • [15] Microsoft “Bing Entanglement Search API” [Online; accessed for 2023], https://www.microsoft.com/en-us/bing/apis/bing-web-search-api, 2023
  • [16] Shots Mikolov et al. “Advances in pre-training distributors word representations” In International Conference on Language Resources and Evaluation, 2018
  • [17] Tomas Mikolov, Kai Chen, Greg Corrado and Jp Dean “Distributed Graphics von Words and Phrases and their Compositionality” In arXiv preprint arXiv:1310.4546, 2013
  • [18] Boris Pallier “Openlexicon, GitHub repository”, 2019 URL: https://github.com/chrplr/openlexicon
  • [19] Redis “Redis Pub/Sub” [Online; in data 22-agosto-2022], 2022 URL: https://redis.io/docs/manual/pubsub/
  • [20] Leonardo Rigutini, Michaelangelo Diligenti, Marco Maggini and Maro Gori “Automatic generation of crossword puzzles” In International Journal on Artifi Intelligence Tools 21.03 World Academic, 2012, pp. 1250014
  • [21] Eric Wallace et al. “Automated Crossword Solving” In arXiv preprint arXiv:2205.09665, 2022
  • [22] Andrea Zugarini and Marco Ernandes “A Multi-Strategy Approach to Crossword Clue Answer Retrieval and Ranking.” In CLiC-it, 2021