Detecting CSV file dialects by table uniformity measurement and data type inference

Abstract

The human-readable simplicity with which the CSV format was devised, together with the absence of a standard that strictly defines this format, has allowed the proliferation of several variants in the dialects with which these files are written. The latter has meant that the exchange of information between data management systems, or between countries and regions, requires human intervention during the data mining and cleansing process. This has led to the development of various computational tools that aim to accurately determine the dialects of CSV files, in order to avoid data loss at data loading stage in a given system. However, the dialect detection is a complex problem and current systems have limitations or make assumptions that need to be improved and/or extended. This paper proposes a method for determining CSV file dialects through table uniformity, a statistical approach based on table consistency and records dispersion measurement along with the detection of data type over each field. The new method has a 93.38% average accuracy on a dataset with 548 CSV files composed of samples coming from a data load testing framework, the test suite provided by the CSV on the Web Working Group (CSVW), curated experimental data set from similar tool development and some others CSV files added as verification of the parsing routines. In tests, the proposed solution outperforms the state-of-the-art tool by achieving an average improvement of 16.45%, resulting in an net increment of about 10% in the accuracy with which dialects are detected on truly messy data for this research dataset. Furthermore, the proposed method is accurate enough to determine dialects by reading only ten records, requiring more data to disambiguate those cases where the first records do not contain the necessary information to conclude with a dialect determination.

1. Introduction

The CSV files are a special kind of tabulated plain text data container widely used in data exchange, currently there is no defined standard for CSV file’s structure and a multitude of implementations and variants. Notwithstanding the foregoing, there are specifications such as RFC-41801

¹

https://datatracker.ietf.org/doc/rfc4180/

that define the basic structure of these files, while a useful addendum to this is defined in the specifications of the USA Library of Congress (LOC). According to the LOC specifications the CSV simple format is intended for representing a rectangular array (matrix) of numeric and textual values. “It is a delimited data format that has fields/columns separated by the comma character %x2C (Hex 2C) and records/rows/lines separated by characters indicating a line break. RFC-4180 stipulates the use of CRLF pairs to denote line breaks, where CR is %x0D (Hex 0D) and LF is %x0A (Hex 0A). Each line should contain the same number of fields. Fields that contain a special character (comma, CR, LF, or double quote), must be “escaped” by enclosing them in double quotes (Hex 22). An optional header line may appear as the first line of the file with the same format as normal record lines. This header will contain names corresponding to the fields in the file and should contain the same number of fields as the records in the rest of the file. CSV commonly employs US-ASCII as character set, but other character sets are permitted”2

²

https://www.loc.gov/preservation/digital/formats/fdd/fdd000323.shtml

. Furthermore, so far to the specifications, in a file may exist: commented or empty records; the tab character (\t) or semicolon (;) as field delimiter; one or more, in exceptional cases, of the characters CRLF, CR, and LF as a record delimiter; quote character escaped by preceding it with a backslash (Unix style).

Given that many public administration portals use CSV files to share information of public interest,3

³

An analysis of a 413 GB data body found CSV files available for download on 232 portals.

coupled with the reality that the process of manipulating the information contained in them requires structuring the data in tables and correcting data quality errors, it is necessary to automate tasks as much as possible to reduce the time and effort required to deal with messy CSV data [8,10]. The automation problem focuses on seeking the delimiters (also called dialect sniffing) of a given file. Dialect sniffing requires that the field delimiter, record delimiter and escape character be determined [11].

This problem seems straightforward, but it is by no means simple. If one opts to implement a simple field delimiter counter to choose the one with the most occurrences in the entire file, it is very likely that disambiguation will become impossible if the algorithm is confronted with data that have two or more delimiters with the same number of matches.

A CSV file with a structure as shown in Fig. 1 is at risk of being misinterpreted, this is illustrated in [4]. If delimiters are counted, the period or space will be selected as field delimiters because of their three constant occurrences, generating four fields, in the records, as opposed to the two occurrences and three fields generated by the comma and semicolon, as shown in Fig. 2. Although a well-defined file should have a header row, there are many files on the Internet that do not [10].

Fig. 1.

CSV that cannot be disambiguated by a simple delimiter count.

Fig. 2.

Misinterpreted data using the “most frequent char” strategy.

It is a fact that systems that work with CSV files may require the user to set the configuration with which they want the file to be processed, however, when the intention is to analyze data coming from different sources, it is very beneficial to implement a methodology that allows to automatically infer CSV dialects with minimal user intervention.

In this sense, CSV file dialect inference is a fundamental part of data mining, data wrangling and data cleansing environments [10]. Moreover, dialect detection has the potential to be embedded in systems designed for the new paradigm with the NoDB philosophy, under which it is proposed to make databases systems more accessible to users [2,7]. These trends suggest that the traditional practice of considering CSV files outside of database systems is tending to change [6].

The methodology presented in this paper approaches the problem from a new perspective, combining a set of tangible characteristics in structured data into a single metric. To achieve this goal, a mathematical model was devised that receives as input a table, whose concept can be translated into data structured in an array variable or other similar programming structure, making it easy to implement in different systems used to load information from CSV files. The method takes advantage of a table structure definition by considering the table header as a simple record. This approach makes it possible to infer CSV dialects with high accuracy, and without the need for additional configuration, over files in both cases, whether or not the header row is present. The input table structure is evaluated on the basis that a correctly structured table has persistence of fields between its records and of data types across its fields, coupled with the fact that tables tend to cluster the data rather than show scattered observations. From these key facts, the concept of table uniformity is derived, which qualifies the input table by determining the consistency between input table structure and the data present in its records, also weighing the divergence between these factors.

Traditionally, even notorious in the methodology implemented by state-of-the-art tools, heuristics focus on weighting the count of delimiters in conjunction with data detection. This approach fails when implemented on single-column tables, since the delimiter count is zero, and assumes that the more occurrences of a given character, the higher the probability that it is the delimiter of a potential CSV dialect with which the file was written. Although this observation is applicable for a good number of CSV files, as demonstrated in [11], there are many other situations where this observation results in a false positive and an inappropriate interpretation of the information contained in certain CSV files. The present research sheds light on the importance of divergence measurement, incorporated in the new concept of table uniformity, in solving this type of problem and in the consequent increase in the accuracy with which dialects are determined in CSV files.

2. Related work

Dialect detection in CSV files is an understudied field, and there are few sources on the subject. In 2017, T. Döhmen proposed the ranking decision method based on quality hypotheses for parsing CSV files. A similar method is implemented in the DuckDB system4

⁴

https://duckdb.org/docs/archive/0.9.2/

. Another treatment, based on the discovery of the table structures once the information is loaded into the RAM, is addressed by C. Christodoulakis et al. [3]. This latter methodology uses the classification of records present in CSV files with a specific heuristic applied to discover and interpret each line of data.

In 2019, G. van den Burg et al., developed the CleverCSV system as a culmination of his research, in which he demonstrated that the methodology significantly improved the accuracy for dialects determination problem compared to tools such as Python’s csv module, or the intrinsic functions of the Pandas package, also in the Python programming language. The implementation of CleverCSV is based on detection of patterns in the structure of CSV records, in addition to data types inference over the fields that compose each record. In this way, the utility applies necessary heuristics to seek the potential dialect for a given CSV file through mathematical and logical operations devised to discern between possible dialects [11].

In 2023, Leonardo Hübscher et al., presented a research project that led to the development of a software application capable of detecting tables in text files. This research considers the dialect determination of CSV files as a subproblem to be solved in order to seek the dialect that produces the best table [5].

3. Preliminaries

Properly formulating the dialect detection problem requires establishing certain fundamental definitions.

Definition 1 (CSV content).

Given a CSV file Υ, its content is defined as $\mathit{\xi }\{{\mathit{\xi }}_1\mathrm{,}{\mathit{\xi }}_2\mathrm{,}\dots \mathrm{,}{\mathit{\xi }}_{\mathit{n}}\}$ , where ${\mathit{\xi }}_{\mathit{i}}\in \mathrm{\Omega }$ and Ω represents a character set encoded using a given encoder.

As per the CSV content definition, there is a real possibility that a single CSV file contains characters encoded in more than one encoder. For the purposes of this document, it is assumed that all characters share the same encoding.

Given that each file Υ originates from a table Γ to which a format Ψ(Γ, ρ) and the helper function $\mathit{W}\mathrm{(}\mathit{\xi }\mathrm{)}$ have been applied to produce and write a sequence of human readable characters separated by lines; then from each CSV content ξ is possible to obtain a table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ so that we can verify ${\mathrm{\Gamma }}_{\mathit{\delta }}={\mathrm{\Psi }}^{-1}\mathrm{(}{\mathit{\xi }}_{\mathit{\delta }}\leftarrow \mathit{R}\mathrm{(}\mathrm{{\rm Y}}\mathrm{)}\mathrm{,}{\mathit{\rho }}_{\mathit{\delta }}\mathrm{)}$ .

Definition 2 (CSV table).

A table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ is defined as a set of records composed of a given set of fields, which share data types between corresponding fields across their records. This table can be represented as a data array of fields and records. Thus, its records are defined as $\mathrm{\Phi }\{{\mathit{\phi }}_1\mathrm{,}{\mathit{\phi }}_2\mathrm{,}\dots \mathrm{,}{\mathit{\phi }}_{\mathit{n}}\}$ ; i.e. a set of fields ${\mathit{\phi }}_{\mathit{i}}$ ; $\mathit{i}\in [1\mathrm{,}2\mathrm{,}\dots \mathrm{,}\mathit{k}]$ . Then, the table can be expressed as ${\mathrm{\Gamma }}_{\mathit{\delta }}\{{\mathrm{\Phi }}_1\mathrm{,}{\mathrm{\Phi }}_2\mathrm{,}\dots \mathrm{,}{\mathrm{\Phi }}_{\mathit{n}}\}$ ; i.e. a set of records ${\mathrm{\Phi }}_{\mathit{i}}$ ; $\mathit{i}\in [1\mathrm{,}2\mathrm{,}\dots \mathrm{,}\mathit{n}]$ .

The function $\mathit{R}\mathrm{(}\mathrm{{\rm Y}}\mathrm{)}$ is in charge of reading content from the file Υ, while the function ${\mathrm{\Psi }}^{-1}\mathrm{(}{\mathit{\xi }}_{\mathit{\delta }}\mathrm{,}{\mathit{\rho }}_{\mathit{\delta }}\mathrm{)}$ parses and transforms the CSV content ${\mathit{\xi }}_{\mathit{\delta }}$ into a table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ . The parsing and transformation processes are clearly out of this study scope, so in the following it is assumed that the selected implementation is able to process the tables obtained by parsing a CSV file with the selected tool.

Definition 3 (CSV dialect).

Let Γ be the data table from which the content ${\mathit{\xi }}_{\mathit{\delta }}$ of file Υ is generated, the dialect ρ is defined as the formatting rule to be applied to produce the output data stream.

So that, by the dialect definition, the following statement is verified: $$\begin{array}{r}\mathrm{{\rm Y}}\leftarrow \mathit{W}\mathrm{(}\mathit{\xi }\leftarrow \mathrm{\Psi }\mathrm{(}\mathrm{\Gamma }\mathrm{,}\mathit{\rho }\mathrm{)}\mathrm{)};\mathit{\rho }\{{\mathit{\upsilon }}_{\mathit{d}}\mathrm{,}{\mathit{\upsilon }}_{\mathit{q}}\mathrm{,}{\mathit{\upsilon }}_{\mathit{e}}\mathrm{,}{\mathit{\upsilon }}_{\mathit{r}}\}\in \mathrm{\Omega }.\end{array}$$

Definition 4 (CSV dialect determination).

Given a CSV file Υ determining the dialect is the act of seeking the dialect ${\mathit{\rho }}_{\mathit{\delta }}$ that satisfies the statement $\mathrm{\Gamma }\approxeq {\mathrm{\Gamma }}_{\mathit{\delta }}\leftarrow {\mathrm{\Psi }}^{-1}\mathrm{(}{\mathit{\xi }}_{\mathit{\delta }}\leftarrow \mathit{R}\mathrm{(}\mathrm{{\rm Y}}\mathrm{)}\mathrm{,}{\mathit{\rho }}_{\mathit{\delta }}\mathrm{)}$ .

Thus, it can be concluded that for a CSV file Υ, created using a dialect ρ, there exists a dialect ${\mathit{\rho }}_{\mathit{\delta }}$ that verifies the condition $\mathrm{\Gamma }\approxeq {\mathrm{\Gamma }}_{\mathit{\delta }}$ . Therefore, it is verifiable that the content of a CSV file is a function of its dialect.

3.1. Potential dialect boundaries

It should be noted that multiple potential dialects can produce similar table outputs that are equal or approximately equal to the source table Γ. Furthermore, ${\mathit{\rho }}_{\mathit{\delta }}$ shares the same character set as the contents ξ for the CSV file Υ. That is, an element from ${\mathit{\rho }}_{\mathit{\delta }}$ can be practically any character within Ω domain. Thus, it is necessary to reduce the range of candidate characters involved in dialect detection to streamline the process.

For the purposes of this research, the potential dialect is restricted to

${\mathit{\rho }}_{\mathit{\delta }}\{$

${\mathit{\upsilon }}_{\mathit{d}}[\text{“}\mathrm{,}\text{”}\text{“};\text{”}\mathit{TAB}\text{“}|\text{”}\text{“}:\text{”}\mathit{SPACE}]$ ,

${\mathit{\upsilon }}_{\mathit{q}}[\text{“}\text{” “}\text{“}\text{”}\text{“}\sim \text{”}]$ ,

${\mathit{\upsilon }}_{\mathit{e}}[{\mathit{\upsilon }}_{\mathit{q}}\text{“}\setminus \text{”}]$ ,

${\mathit{\upsilon }}_{\mathit{r}}[\mathit{CRLF}\mathit{CRLF}]\}$ 5

⁵

In most applications the record delimiter ${\mathit{\upsilon }}_{\mathit{r}}$ is not considered, as modern systems handle new lines discrepancies internally.

4. Approach

As introduced in previous sections, the methodologies currently used to determine CSV dialects share a common area for improvement that can be exploited by incorporating data divergence variables into the models. The concept of table uniformity, defined as a computable parameter, encompasses variables aimed at addressing these deficiencies and reducing uncertainty in the determination of CSV dialects. The divergence of tables produced when reading CSV files using a specific dialect can be quantified based on the dispersion of their data, and persistence can be measured in terms of the consistency of their records. This approach aims to reduce false positives.

4.1. Table uniformity

The table uniformity approach is proposed to solve the problem of dialect determination. The method is based on consistency measurement over a table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ , which has been returned by parsing a CSV file with a dialect ${\mathit{\rho }}_{\mathit{\delta }}$ , and the dispersion of records along with the inference of raw data types from fields.

Definition 5 (Table consistency).

Let ${\mathrm{\Gamma }}_{\mathit{\delta }}$ be a table generated when parsing a CSV file Υ, using a dialect ${\mathit{\rho }}_{\mathit{\delta }}$ , the table consistency, denoted by ${\mathit{\tau }}_0$ , is a ratio that describes how uniform ${\mathrm{\Gamma }}_{\mathit{\delta }}$ is across its k fields and its n records.

Definition 6 (Records dispersion).

Let Φ be the sets of records from table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ , generated when parsing a CSV file Υ using a dialect ${\mathit{\rho }}_{\mathit{\delta }}$ , the records dispersion, denoted by ${\mathit{\tau }}_1$ , is a measure describing the magnitude of the change in the records composition throughout ${\mathrm{\Gamma }}_{\mathit{\delta }}$ .

These definitions are based on the fact that tables, in general, have a defined structure with persistent k fields in its n records.

The two measurements that define the table uniformity parameter $\mathit{\tau }\{{\mathit{\tau }}_0\mathrm{,}{\mathit{\tau }}_1\}$ are related to the structure of records Φ from a table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ . Where ${\mathit{\tau }}_0$ is a direct function of the standard deviation of fields, and ${\mathit{\tau }}_1$ is a function measuring the weighted dispersion in records structures as a factor of the statistical segmented mode.6

⁶

Segmented mode refers to the use of sample segments, which are defined as the data undergoes dispersion.

$$\begin{array}{c}{\mathit{\tau }}_0={\displaystyle \frac{1}{1+2\sqrt{\mathit{\sigma }}}};{\mathit{\tau }}_1=2·\mathit{R}\mathrm{(}{\mathit{\alpha }}^2+1\mathrm{)}\mathrm{\left(}{\displaystyle \frac{1-\mathit{\beta }}{\mathit{M}}}\mathrm{\right)}\end{array}$$

Where, for a given table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ , σ is the number of fields standard deviation across records; α represents the count of times number of fields changes between records; R is the statistical range for the number of fields over records; M is the segmented mode, describing the largest number of times the record structure is sequentially preserved within the table, and $\mathit{\beta }=\frac{\mathit{M}}{\mathit{n}}$ is the records variability factor.

The definitions provided propose a concept diametrically opposed to that used in most solutions, since it discourages data dispersion, i.e. records with a higher number of fields/columns are only favored if their record structure is uniform.The parameter ${\mathit{\tau }}_0$ indicates the degree of consistency for the records in a table, while ${\mathit{\tau }}_1$ is a fine-grained measure of the dispersion and inconsistency within the records. This quality allows the new method to discern between data tables by inferring uniformity in two senses: consistent and invariant records with little dispersion in their structure. The parameter ${\mathit{\tau }}_0$ ranges from $0⩽{\mathit{\tau }}_0⩽1$ , being 1 for those tables with consistent records; while ${\mathit{\tau }}_1$ ranges from $0⩽{\mathit{\tau }}_1\mathrm{<}\infty$ , being 0 for those tables with invariant record structure and without dispersion.

4.2. Type detection

Data type detection is the core basis of the implemented methodology. Recognition of data types over fields from each record allows us to collect information about the contents of a given table. In this context, the records scoring, denoted as λ, is computed as $$\begin{array}{c}\mathit{\lambda }={\displaystyle \frac{{\mathrm{(}{\sum }_{\mathit{i}=1}^{\mathit{k}}{\mathit{S}}_{\mathit{i}}\mathrm{)}}^2}{100·{\mathit{k}}^2}}\end{array}$$

Where ${\mathit{S}}_{\mathit{i}}$ is a score for the ith field φ in $\mathrm{\Phi }\{{\mathit{\phi }}_1\mathrm{,}{\mathit{\phi }}_2\mathrm{,}\dots \mathrm{,}{\mathit{\phi }}_{\mathit{n}}\}$ from the table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ . If the type of the ith field φ is known, ${\mathit{S}}_{\mathit{i}}=100$ , ${\mathit{S}}_{\mathit{i}}=0.1$ otherwise.

It is important to highlight that the detection of data types is based on the findings presented in the research conducted at [8] on approximately 413 GB of data. The study determined that 97% of the columns loaded from CSV files were of numeric, date, and character sequence types. In the latter category, fields with IDs and Tokens predominate. Only 3% of the studied fields were determined to be empty. These conclusions suggest that inferring a few data types is sufficient to differentiate between CSV file dialects.

Data types are inferred using simple pattern matching like MM/DD/YYYY, which can be implemented with Regex engines or their predecessors, in the case of non-numeric data types. For numeric data, inference routines also use simple data conversion instructions supported by programming languages where a text string is taken as input and the programming language returns a truth value for the requested inference. That being said, it is worth mentioning that the data types detected by conversion functions can vary between programming languages.

For the purposes of this paper, the following field types are generally considered to be known:

• Time and date: matching regular dates and time format, as well stamped ones like

  MM/DD/YYYY[YYYY/MM/DD] HH:MM: SS +/- HH:MM.

• Numeric: matching all numeric data supported by the implementation language selected.

• Percentage.

• Alphanumeric: matching numbers, ASCII letters and underscore.

• Currency.

• Especial data: like “n/a” or empty strings.

• Email.

• System paths.

• Structured scripts data types: matching JSON arrays and data delimited by parentheses, curly and square brackets.

• Numeric lists: matching fields with numeric values delimited with common separator character.

• URLs.

• IPv4.

Al other fields will be scored as unknown type. A particularly true fact is that data inference is an inherently incomplete process aimed at favoring one dialect over another.

4.3. Table scoring

Once table uniformity $\mathit{\tau }\{{\mathit{\tau }}_0\mathrm{,}{\mathit{\tau }}_1\}$ for records $\mathrm{\Phi }\{{\mathit{\phi }}_1\mathrm{,}{\mathit{\phi }}_2\mathrm{,}\dots \mathrm{,}{\mathit{\phi }}_{\mathit{n}}\}$ from the table ${\mathrm{\Gamma }}_{\mathit{\delta }}\{{\mathrm{\Phi }}_1\mathrm{,}{\mathrm{\Phi }}_2\mathrm{,}\dots \mathrm{,}{\mathrm{\Phi }}_{\mathit{n}}\}$ , which has been generated by reading a CSV file Υ using a dialect ${\mathit{\rho }}_{\mathit{\delta }}$ , and the score λ are computed, the table score, denoted as ϖ, is computed as $$\begin{array}{c}\mathit{\varpi }=\mathrm{(}{\displaystyle \frac{{\mathit{\tau }}_0}{\mathrm{\Delta }}}+{\displaystyle \frac{1}{{\mathit{\tau }}_1+\mathit{n}}}\mathrm{)}·{\displaystyle \sum }_{\mathit{i}=1}^{\mathit{n}}{\mathit{\lambda }}_{\mathit{i}};\forall \mathit{n}\mathrm{>}1\end{array}$$

Where Δ is a threshold indicating the expected number of records to be imported from the CSV file Υ which contains a number of records m. For $\mathit{m}\mathrm{>}\mathit{n}$ , and an appropriate selection of ${\mathit{\rho }}_{\mathit{\delta }}$ , ${\mathrm{\Psi }}^{-1}\mathrm{(}{\mathit{\xi }}_{\mathit{\delta }}\leftarrow \mathit{R}\mathrm{(}\mathrm{{\rm Y}}\mathrm{)}\mathrm{,}{\mathit{\rho }}_{\mathit{\delta }}\mathrm{)}$ will generate a table where $\mathrm{\Delta }=\mathit{n}$ ; therefore, by the definition stated, the table score is in the range $0\mathrm{<}\mathit{\varpi }⩽200$ .

In the case $\mathit{n}=1$ we have $$\begin{array}{c}\mathit{\varpi }=\mathit{\lambda }·{\displaystyle \frac{\mathit{\eta }+\frac{1}{\mathit{k}}}{\mathit{k}-\lfloor \mathit{\eta }·\mathit{k}\rfloor +1}}\end{array}$$

Where $\mathit{\eta }=\frac{\sqrt{\mathit{\lambda }}}{10}$ is a discriminant to ensure the exclusion of false positives with a single record.

4.4. Determining CSV file dialects

Algorithm 1

Dialect determination

This section shows the core algorithms on which the methodology presented in this research is based, complementary algorithms are listed in the appendix.

The main pseudocode for dialect determination is listed in Algorithm 1. At line 2 the set of predefined dialects are initialized; then, in line 4, a table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ is created by parsing the CSV content $\mathit{\xi }$ with each $\mathit{\rho }$ dialect.

At this point, it becomes clear that the selection of a robust parser is of utmost importance in order to obtain the best results even on messy files. In line 5, the output table ${\mathrm{\Gamma }}_{\mathit{\delta }}$ is scored and this result is saved within the current dialect in the collection ℵ. At line 6, the dialect that gets the highest scored table is selected.

The table uniformity procedure is outlined in Algorithm 2 pseudocode. The method uses a set of sentinels to measure table inconsistency through monitoring table changes over parsed records.

The parameter ${\mathit{\tau }}_0$ is derived from the standard deviation that indicates how uniformly the fields count are grouped around the average number of fields contained in the parsed records, resulting in an appropriate measure to qualify the structure of a table [1]. However, when there are two or more dialects with a small variance, the ${\mathit{\tau }}_0$ parameter is not decisive. It is in this situation where the ${\mathit{\tau }}_1$ parameter provides support by penalizing tables with variations in its records structures, and whose structure resembles sparse data that do not maintain consistency.

Algorithm 2

Table uniformity

5. Evaluation setup

This section outlines the structure of comparative tests between state-of-the-art tool and the approach presented in the current research. Previous studies, which have even been the backbone in the development of tools like CleverCSV, have demonstrated that there is significant variability in the dialects present in CSV files, making it necessary to thoroughly study the available alternatives to overcome these issues. Considering this, the selected datasets aim to test edge cases where dialect detection is non-trivial.

The proposed method was tested at development phase using a simple set of 19 test files, the results of which were used to conclude the coding phase. Remaining CSV files were integrated into the global dataset without running any tests on them. It was decided to code the new method and integrate it with CSV Interface,7

⁷

https://github.com/ws-garcia/VBA-CSV-interface

a VBA CSV file parser. Thus, the new CSV dialect determination method will be available in a widespread programming language without over-investing efforts. Additionally Python code has been written to run the tests for CleverCSV. The code repository is currently available on GitHub.8

⁸

https://github.com/ws-garcia/CSVsniffer

To meet the reproducibility requirements of the experiments conducted, the code repository was linked to a Zenodo record.9

⁹

https://zenodo.org/records/11331538

This record contains everything necessary for any other researcher to replicate the results obtained in the present study.

5.1. Datasets

The experiments uses three datasets, which have been added to the Zenodo record cited earlier in this section

• provided by Gerardo Vitagliano et al., also available in GitHub. For this dataset, one or two polluted CSV file per pollution case are included for testing, all the 99 surveys having at least one pollution case as described in the aforementioned study (excluding empty ones by the fact infinite dialects can be produce no payload files [12]). In addition, the dataset was enriched with data from the OpenRefine10

  ¹⁰

  An open-source tool for working with messy data: https://openrefine.org/.

  testing, CleverCSV failure cases and other files used at development phase serves as testing samples. In total, this dataset is comprised of 148 CSV files (104 MB of data).

• provided by G. van den Burg as URLs into JSON files in the CleverCSV repository. This dataset is composed of the 256 CSV files that CleverCSV could not accurately determine when conducting the research that led to the tool development [9]. At the time of this research, 244 of these files were available online. A filter was applied to exclude from the dataset all files with a structure that did not visually look like a CSV. After filtering, the dataset ended up with 179 CSV files (79 MB of data).

• provided by the CSV on the Web Working Group.11

  ¹¹

  https://github.com/w3c/csvw/tree/gh-pages/tests

  This dataset consists of 221 CSV files (33.5 MB).

Overall, the dataset used in the experiments comprises 548 CSV files (216.5 MB).

5.2. Ground truth

The CleverCSV testing dataset were filtered to extract from them a subset of CSVs that we can call “messy”; the structure of these being unconventional and whose dialect is much more difficult to infer. This last step is required since the dataset contains files that fall under the “normal forms” classification implemented in CleverCSV, which refers to CSV files with such a simple structure that they allow the determination of their dialects using only data inference.12

¹²

https://clevercsv.readthedocs.io/en/latest/source/clevercsv.html#module-clevercsv.normal_form

Each dataset was then manually reviewed and analyzed to produce a reliable ground truth. For each dataset, a file containing the dialect for each CSV file was produced. Each dialect was determined objectively, with RFC-4180 specifications predominance.

5.3. Tools

Since the research in [11] was conclusive positioning CleverCSV as a tool with the state-of-the-art methodology for CSV dialect determination, the latter will be the only used in the comparative experiments in the present research.

a Python package for handling messy CSV files that aims to provide a direct replacement for the built-in CSV module with improved dialect detection. This package uses a set of techniques to discern between dialects by determining the data types and patterns followed by the structure of CSV file records. The tool uses some techniques to detect potential dialects in early stages, discarding the others at analysis beginning, with performance improvement as a main objective.

6. Experiments

The manually annotated files were used for returned dialects and ground truth comparison in order to validate the automatic detection. Both tools were evaluated for the accuracy with which they determine the dialects of the CSV files for each test dataset. In this context, we define the accuracy of dialect detection as the ratio of correctly detected dialects to the total number of test files with no error after execution.

The experiments with CleverCSV were conducted by running scripts from the clevercsv_test.py file. The results are stored in the “Current research” folder for the results obtained by the proposed methodology and in the “cleverCSV” folder for the results obtained by CleverCSV.

In parallel, the tests were designed to measure the execution times of each methodology, even though it is known that there is no possible comparison between the two solutions. This is because CleverCSV precision in detecting dialects is closely related to reading all the information contained in CSV files,13

¹³

https://github.com/alan-turing-institute/CleverCSV/issues/15

which is a clear disadvantage in a performance test where the competing solution only needs to load a few records to infer dialects.

6.1. Closer look

Let’s examine a preliminary behavior comparison example for the studied methodologies. The Fig. 3 shows a preview from the modified content of one file used during the testing phase. File was accessed from the CleverCSV repository on GitHub.14

¹⁴

https://github.com/alan-turing-institute/CleverCSV/issues/99

The star character has been replaced by the vertical bar “|” to include in the detection a potential dialect with this character. As the author points out, the CSV file is comma delimited, using double quotes as the quote and escape character, then this file is compliant with RFC-4180 specifications. When running dialect detection, CleverCSV gets the vertical bar “|” as the delimiter because this field pattern gets a $\mathit{P}=93.6395$ score vs a $\mathit{P}=37.647059$ from patterns with the “,” character as delimiter. This behavior is because the implemented logic heavily weights the delimiter count over the detected data types, where dialects containing the comma as delimiter obtain a type score of $\mathit{T}=0.942647$ against the type score of $\mathit{T}=0.843074$ obtained by dialects with the vertical bar as delimiter.

By executing the algorithms presented in this research, we get the following for dialects with the vertical bar as the delimiter $\mathit{\lambda }=448.2243$ , ${\mathit{\tau }}_0=0.2056$ , ${\mathit{\tau }}_1=12$ , and $\mathit{\varpi }=29.5883$ . For the comma we get $\mathit{\lambda }=897.3315$ , ${\mathit{\tau }}_0=1$ , ${\mathit{\tau }}_1=0$ , and $\mathit{\varpi }=179.4663$ . Then the comma “,” character is selected as delimiter.

Fig. 3.

Messy CSV file preview.

7. Results

In this section, the results from experiments execution are disclosed. It’s pivotal to underscore that priority will be given to the accuracy in dialects detection due to the precision diminished experienced by CleverCSV when the complete file content is not read during dialect determination. Specific details on this particular will be given upon in the discussion section.

7.1. Dialect detection accuracy

The Table 1 shows the results after running the dialect detection tests over the simple Pollock testing dataset. It can be seen that the new proposed heuristic gets a perfect score when using a table with a threshold of fifty records (50R) to be imported from the target CSV file.

Table 1

Accuracy on dialect detection in simple Pollock testing dataset. An erroneous detection implies that the method has failed to infer either the delimiter or quote character, or both

Method	Success rate %	Erroneous rate %
Actual (10R)	99.32	0.68
Actual (25R)	99.32	0.68
Actual (50R)	100.00	0.00
CleverCSV	94.59	5.41

Fig. 4.

Scoring variation of three different delimiters and their dialects when applying the uniformity heuristic over tables from the dd_Wickenburg_nobmp_623.csv file.

When using tables of ten or twenty-five records (10R, 25R) for dialect determination, the proposed method was not able to determine dialect of the “dd_Wickenburg_nobmp_623.csv” file for the testing dataset. This file has been selected to show the variation of certainty as the considered table size increases across computations. As can be seen in the Fig. 4, when the proposed heuristic is applied, it is settled that delimiter is the equal sign “=”, since the dialects containing it divide each record into known data types: an alphanumeric field/column and a field with structured data delimited by square brackets. Increasing the table size to twenty-five (25R) induces the heuristic begins to highlight the semicolon “;” as a possible field delimiter character. Finally, the semicolon is correctly detected as a delimiter when the threshold of fifty records (50R) in the table is specified. This behavior demonstrates that the proposed methodology is strongly related to changes in the structure of tables used in dialect inference.

The results obtained after running the tests over dataset from CleverCSV are shown in Table 2. In this dataset the percentage of incorrectly detected dialects became approximately 10%. This metric indicates the presence of CSV files with unconventional structures. Notwithstanding the foregoing, dialect detection improves by 9.81% compared to CleverCSV.

CleverCSV running in verbose mode indicates that the tool failed to read 37 of the test files with errors related to the file encoding. These files, along with ones listed as “normal forms”, were excluded from the dataset, producing a really messy subset of CSV files. Executing the tests over this selective filtered subset yields the results shown in Table 3. For this subset of files there is a slight increase in the rate of incorrect detections, preserving the 10% improvement of the new methodology over CleverCSV. On average, the heuristic proposed in this research shows an improvement of 7.51% compared to CleverCSV, outperforming the latter with 10% when handling messy CSV files.

Table 2

Accuracy on dialect detection in the failed CleverCSV dataset. An erroneous detection implies that the method has failed to infer either the delimiter or quote character, or both

Method	Success rate %	Erroneous rate %
Actual (10R)	88.83	11.17
Actual (25R)	89.39	10.61
Actual (50R)	88.83	11.17
CleverCSV	79.58	20.42

Table 3

Accuracy on dialect detection over really messy CSV files. An erroneous detection implies that the method has failed to infer either the delimiter or quote character, or both

Method	Success rate %	Erroneous rate %
Actual (10R)	86.51	13.49
Actual (25R)	87.30	12.70
Actual (50R)	87.30	12.70
CleverCSV	76.98	23.02

The results after running tests over the CSVW dataset are shown in Table 4. CleverCSV exhibits a fall in accuracy, being successful in only 56.56% of CSV files, when attempting to infer dialects in this dataset. By contrast, the current research’s methodology retains a high level of success with dialects satisfactorily determined in 96.83% of the CSV files. This is a completely unexpected finding that will be the subject of a closer inspection in the discussion section.

Table 4

Accuracy on dialect detection over CSVW dataset. An erroneous detection implies that the method has failed to infer either the delimiter or quote character, or both

Method	Success rate %	Erroneous rate %
Actual (10R)	96.38	3.62
Actual (25R)	96.83	3.17
Actual (50R)	96.83	3.17
CleverCSV	56.56	43.44

7.2. Performance

Run times were quantified for both CleverCSV and the proposed new methodology. By reviewing the results shown in Table 5, we confirm that the performance of dialect detection is directly related to the amount of data loaded from the CSV files. CleverCSV is particularly affected by the required loading of all data from CSV files, as mentioned earlier in this document.

Table 5

Run time over datasets, in seconds

Method	Pollock	CleverCSV-Messy	W3C-CSVW
Actual (10R)	124.39	70.75	41.41
Actual (25R)	127.19	116.93	50.09
Actual (50R)	132.39	112.84	73.21
CleverCSV	302.07	157.48	132.5

8. Discussion

By looking closely at the results obtained, it can be deduced that there are two main categories that influence the certainty of determined dialects: the type of heuristics used, the CSV file parser behavior while producing tables using a certain dialect. In this section both categories are discussed in order to briefly qualify the experiments results.

8.1. Heuristic

In contrast to CleverCSV, in whose heuristic the detection of data types serves as a factor to scale down the score obtained by a certain pattern; the table consistency method uses data detection as a base score to be narrowed using the table consistency and data dispersion parameters. The results therefore indicate that the factors obtained are not commutative.

Since data type detection is a fundamental part of both methods, it is necessary to include a wide range of known data typologies. This factor is undoubtedly determining in dialect detection. According to Mitlohner’s research [8], with a base of 104,826 CSV files, the vast majority of data commonly stored in this type of files are numeric, tokens (words separated by spaces), entities, URLs, dates, alphanumeric fields and general text, so these data types must be recognized. Additionally, in the field of programming, there are other types of data frequently dumped in CSV files, namely: structured data with the Regex pattern $\mathrm{(}\mathrm{(}[\mathit{a}-\mathit{z}\mathit{A}-\mathit{Z}]+[\setminus \mathrm{(}[\mathit{a}-\mathit{z}\mathit{A}-\mathit{Z}]+[\setminus [\{][\wedge \setminus ]]\ast [\setminus ]\}]\mathrm{)}[\{][\wedge \setminus ]]\ast [\setminus ]\}]\mathrm{)}$ , numerical lists, tuples, arrays among others.

It is worth mentioning that dialect detection is prone to failure when the CSV file is composed of unknown data types. In these cases, the table uniformity tends to select dialects that produce registers with a single field. When reviewing the cases where CleverCSV was not able to determine the dialect, it has been observed that the common denominator has been the high count of a potential delimiter with more occurrences than the expected delimiter. In this sense, both solutions have poor performance when the space character appears in the list of potential delimiters.

The routines used by CleverCSV to obtain potential dialects behaves unexpectedly when the target CSV file contains a table with a single column, where potential delimiters are not part of the field content in any of its records. This leads to the fact that, for a file with a structure similar to that exhibited by the test165.csv file, test sample from the CSVW dataset, it is impossible to determine the dialect since the comma is not considered in any potential dialect by the referred routines, even though it is the default field separator in CSV files. In this case, the detected dialect has the space character as a field separator, ignoring the fact that it is not part of the file content. The ground truth considers this file to be comma separated due to the full knowledge that the file in question was created by the CSV on the Web Working Group following the RFC-4180 specifications.15

¹⁵

https://w3c.github.io/csvw/primer/#tabular-data

In the same vein, it was found that this type of file makes it difficult to determine dialects in CleverCSV even when the comma is present in the content of the CSV files. This is the case for files with a structure similar to that illustrated in the file test168.csv.

There are files where the threshold of records in the target table is decisive; however, tests have found that the dialect of some files is determined incorrectly as the value of this parameter is increased and more records are loaded into the table. This peculiarity allows us to conclude that the first records can adequately describe the structure of CSV files, avoiding, to a certain extent, the need to read the whole file. In this particular, it was found that CleverCSV had a running time of approximately 19 minutes before completing the tests it was subjected to. The results obtained lead to conclude that the default option when detecting dialects in CSV files should be to read only a sample of the file instead of reading its entire contents.

As pointed out earlier, the table uniformity method prefers grouped data over those that appear to be sparse data. In these cases, detection tends to depend exclusively on the data types detected in the records. This fact is evidenced by plotting the values of the uniformity parameter ${\mathit{\tau }}_0$ .

Looking at Fig. 5, it can be seen that, even though the score obtained by the semicolon dialect is very close to zero, the value of ${\mathit{\tau }}_0$ is maximum. In contrast, this value fluctuates to nearly zero for the dialect containing semicolon; it remains almost unchanged among the dialects using other fields delimiters characters. In these cases, the dialect determination is relegated to data type detection and fine-grained monitoring of changes in table structures through the ${\mathit{\tau }}_1$ parameter. It is noted that the parameters ${\mathit{\tau }}_0$ and ${\mathit{\tau }}_1$ work together for well-defined tables, selectively overriding each other when processing tables with poorly defined data structures.

8.2. CSV parser basis

The accuracy of dialect determination is intimately related to the way CSV parsers behave when confronted with atypical situations. This is because heuristics use these results to infer the configuration that returns the most suitable data structures.

One of the capabilities required for dialect determination is the recovery of data after the occurrence of a critical error. This is the case when import CSV files where there is no balanced quotation count. This situation breaks the RFC-4180 specifications and causes an import error in almost all solutions intended to work with CSV files. In this sense, the recovery of this error should include a specific message after which the loading of information should continue until the whole file is processed.

Fig. 5.

Uncertainty caused by analyzing tables with a single field across all their records.

Since the determination of dialects can be done with a few records received from a CSV file, there is a probability that some of the parameters that compose the dialect cannot be determined properly. Given this reality, it is preferable that CSV parsers be able to convert between one escaping mechanism and another instead of making the escape character mutually exclusive as established in the most relevant proposals on these topics16

¹⁶

https://specs.frictionlessdata.io/csv-dialect/

. This results in the correct interpretation of escape sequences that use the “∖” for those files in which a quote character has been detected as part of their dialect.

9. Conclusion

A method based on the uniformity of tables for CSV dialect determination has been presented. This new methodology evaluates the homogeneity and dispersion over tables structures, weighting them by detecting data types over fields. It is clear that implementing the heuristics developed on CSV parsing systems will provide high accuracy in dialects determination using only a few records form CSV files, ensuring efficient and high precision routines.

Performance in dialect determination is closely linked to the amount of information loaded. However, in the case of CleverCSV, the loss of accuracy is a limitation when trying to circumvent this drawback with a reduction in the amount of information read. Research by its creator suggests that the loss of accuracy can be as high as 20%, which is by no means negligible. This is a compelling reason why dialect detection by table uniformity is a solid alternative to be implemented in different CSV parsers.

9.1. Further work

Despite the improvement obtained in terms of accuracy, the tool can be complemented with new routines that allow adjustments to be made as the information is ingested by the systems. For this, an alternative would be to use an LLM at a late stage to take advantage of its advanced contextual understanding to validate and refine the parsed data, correcting residual errors, fill in missing values and ensure consistency. In this particular case, a traditional CSV file parser would be implemented to make the information loading more efficient. Analysts and data scientists would then be able to produce and implement a hybrid solution that allows them to reduce human intervention to a minimum, as the problems of dialect determination would be marginal. Also, the resource-intensive LLM workload would be optimized by delegating the data input to a specialised piece of CSV file uploading software. Furthermore, this type of solution would solve the problem related to the late appearance of dialect elements, such as escape sequences, that could cause anomalies and data loss on loading, avoiding the need to read the entire CSV content at dialect determination phase.