Detecting CSV File Dialects by Table Uniformity Measurement and Data Type Inference

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-4180 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) [1]. 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]. Furter more, 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 interest1, 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 [3,4]. The automation problem is to determine the delimiters (also called dialect sniffing) of a given file. Dialect sniffing requires that the field delimiter, record delimiter and escape character be determined [5].

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 Figure 1 is at risk of being misinterpreted, this is illustrated in [6]. 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. Although a well-defined file should have a header row, there are many files on the Internet that do not. [4].

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 [3]. 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 [7,8]. These trends suggest that the traditional practice of considering CSV files outside of database systems is tending to change [9].

2. Related work

The detection of dialects in CSV files is a little studied 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 system [10]. A similar treatment, based on the discovery of the table structures once the information is loaded into the RAM, is addressed by C. Christodoulakis et. al. [11]. In the latter, the methodology is based on the classification of the records present in the CSV files, to which a specific heuristic is 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 his methodology significantly improved the accuracy with which dialects were determined compared to tools such as Python’s csv module, or the intrinsic functions of the Pandas package, also in the Python programming language. The methodology implemented in CleverCSV is based on the detection of patterns in the structure of the CSV records, in addition to the inference of the data types of the fields that compose each record. In this way, the utility applies the necessary heuristics to infer the potential dialect for a given CSV file through mathematical and logical operations devised to discern between possible dialects [5].

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 obtain the dialect that produces the best table [12].

3. Problem formulation

In order to properly formulate the dialect detection problem, it is necessary to establish some basic definitions.

Definition 3.1

(CSV content). Given a CSV file ${\rm Y}$ its content is defined as $\mathit{\xi }\left\{{\xi }_1,{\xi }_2,\dots ,{\xi }_n\right\}$, where ${\mathit{\xi }}_{\mathit{i}}\in \mathit{\Omega }$, and $\mathit{\Omega }$ 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 may contain characters encoded in more than one encoder. For the purpose of this document, it is assumed that all characters share the same encoder.

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

Definition 3.2

(CSV table). The table ${\mathit{\Gamma }}_{\mathit{\delta }}$ is defined as a set of records composed of a given set of fields, which share the data typology 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 $\mathit{\Phi }\left\{{\phi }_1,{\phi }_2,\dots ,{\phi }_k\right\}$; i.e., a set of fields ${\phi }_i ;i\in \left[1,2,\dots ,k\right]$. Then, the table can be expressed as ${\mathit{\Gamma }}_{\mathit{\delta }}\left\{{\Phi }_1,{\Phi }_2,\dots ,{\Phi }_n\right\}$; i.e., a set of records ${\Phi }_i ;i\in \left[1,2,\dots ,n\right]$.

The function $\mathrm{R}\left({\rm Y}\right)$ is in charge of reading the content of the file ${\rm Y}$, while the function ${\Psi }^{-1}\left({\xi }_{\delta },{\rho }_{\delta }\right)$ is in charge of parsing and transforming the CSV content ${\xi }_{\delta }$ into a table ${\Gamma }_{\delta }$. The process of parsing and transformation is clearly out of the scope of this study, 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 3.3

(CSV dialect). Let $\mathit{\Gamma }$ be the data table from which the content $\mathit{\xi }$ of file $\mathit{{\rm Y}}$ is generated, the dialect $\mathit{\rho }$ is defined as the formatting rule to be applied to produce the output data stream.

So that, by the dialect definition, it is verified that

$${\rm Y}\leftarrow W\left(\mathit{\xi }\leftarrow \Psi \left(\Gamma ,\mathit{\rho }\right)\right);\mathit{\rho }\left\{{\upsilon }_d,{\upsilon }_{\text{q}},{\upsilon }_{\text{e}},{\upsilon }_{\text{r}}\right\}\in \mathbf{\Omega }$$

Definition 3.4

(CSV dialect determination). Given a CSV file ${\rm Y}$, determining the dialect involves identifying the dialect ${\mathit{\rho }}_{\mathit{\delta }}$ that satisfies the following statement $\mathit{\Gamma }\cong {\mathit{\Gamma }}_{\mathit{\delta }}\leftarrow {\mathit{\Psi }}^{-1}\left({\mathit{\xi }}_{\mathit{\delta }}\leftarrow R\left(\mathit{{\rm Y}}\right),{\mathit{\rho }}_{\mathit{\delta }}\right)$.

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

$$$$

1.1.1. Potential dialect boundaries

It should be noted that multiples potential dialects can produce similar tables outputs that are equal or approximately equal to the source table 𝚪. Furthermore, 𝝆_𝜹 shares the same character set as the contents 𝝃 for the CSV file 𝚼. That is, 𝝆_𝜹 can be practically any character within 𝛀 domain. Thus, it is necessary to narrow down 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 2

$${\mathit{\rho }}_{\mathit{\delta }}\{{\upsilon }_{\mathit{d}}[“,” “;” \text{TAB} “|” “:” \text{SPACE}\left],{\upsilon }_q\left[“\backslash ”“ “\text{'}” “~”\right], {\upsilon }_e\right[{\upsilon }_q “\backslash ”], {\upsilon }_r\left[\text{CRLF CR LF}\right]\}$$

4. Table uniformity

The table uniformity approach is proposed to solve the problem of dialect determination. The method is based on the measurement of the consistency of the table, which has been obtained through a dialect ${\mathit{\rho }}_{\mathit{\delta }}$, and the dispersion of records along with the inference of the raw data types from the records.

Definition 4.1

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

Definition 4.2

(Records dispersion). Let $\mathit{\Phi }$ be the sets of records from table ${\mathit{\Gamma }}_{\mathit{\delta }}$, generated when reading a CSV file $\mathit{{\rm Y}}$ using a dialect ${\mathit{\rho }}_{\mathit{\delta }}$, the records dispersion, denoted by ${\tau }_1$, is a measure describing the magnitude of the change in the records composition throughout the table.

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

The two measurements that define the table uniformity parameter $\mathit{\tau }\left\{{\tau }_0,{\tau }_1\right\},$ are related to the structure of the records $\Phi$ from the table ${\Gamma }_{\delta }$. Where ${\tau }_0$ is a direct function of the standard deviation of fields, and ${\tau }_1$ being a function measuring the weighted dispersion in records structures as a factor of the statistical segmented mode3.

$${\tau }_0=\frac{1}{1+2\sqrt{\sigma }}; {\tau }_1=2·R \left({\alpha }^2+1\right)\left(\frac{1-\beta }{M}\right)$$

where, for a given table ${\Gamma }_{\delta }$, $\sigma$ is the standard deviation of the number of fields across records; $\alpha$ represents the count of times the number of fields changes between records; $R$ is the statistical range of 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 $\beta =M/\mathit{n}$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 ${\tau }_0$ indicates the degree of consistency of the records in the table, while ${\tau }_1$ 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 records and invariant records with little dispersion in their structure. The parameter ${\tau }_0$ ranges from $0\le {\tau }_0\le 1$, being $1$ for those tables with consistent records; while ${\tau }_1$ranges from $0\le {\tau }_1<\infty$, being $0$ for those tables with invariant record structure and without dispersion.

5. Type detection

Data type detection is the core basis of the methodology implemented. Recognition of the types of data fields from each record allows us to collect information about the contents of a given table. In this context, the records scoring, $\mathit{\lambda }$, is defined as

$$\mathit{\lambda }=\frac{{\left({{\displaystyle \sum }}_{i=1}^k{S}_i\right)}^2}{100*{\mathit{k}}^2}$$

where $S_i$ is the score for ith field $\phi$ in $\Phi \left\{{\phi }_1,{\phi }_2,\dots ,{\phi }_k\right\}$ from the table ${\Gamma }_{\delta }$. If the type of the ith field $\phi$ is known $S_i=100$, $S_i=0.1$ otherwise.

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.

6. Table scoring

Once the table uniformity $\mathit{\tau }\left\{{\tau }_0, {\tau }_1\right\}$ for the records $\Phi \left\{{\phi }_1,{\phi }_2,\dots ,{\phi }_k\right\}$ contained in table ${\Gamma }_{\delta }\left\{{\mathsf{\Phi }}_1,{\mathsf{\Phi }}_2,\dots ,{\mathsf{\Phi }}_{\mathrm{n}}\right\}$, which has been generated by reading a CSV file ${\rm Y}$ using a dialect ${\mathit{\rho }}_{\mathit{\delta }}$, and the score $\mathit{\lambda }$ are computed, the table score $\mathit{\varpi }$ is defined as

$$\mathit{\varpi }=\left(\frac{{\tau }_0}{\mathsf{\Delta }}+\frac{1}{{\tau }_1+\mathit{n}}\right)*{\displaystyle \sum }_{i=1}^n{\lambda }_i; \forall \mathit{n}>1$$

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

In the case $\mathit{n}=1$ we have

$$\mathit{\varpi }=\mathit{\lambda }*\frac{\eta +\left(\frac{1}{\mathit{k}}\right)}{\mathit{k}-floor\left(\eta *\mathit{k}\right)+1}$$

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

7. Determining CSV file dialects

Algorithm 1: Dialect Determination

In order to develop a reliable methodology for determining the dialects of CSV files, the cases that produced failures in famous dialect detectors were used as a basis. This is because the dialect in well-structured files is easily determined with any simple heuristic.

The main pseudocode for dialect determination is described in Algorithm 1. At line 2 the set of predefined dialects are initialized; then, in line 4, a table ${\mathsf{\Gamma }}_{\mathsf{\delta }}$ is created by parsing the CSV content $\xi$ with each 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 ${\mathsf{\Gamma }}_{\mathsf{\delta }}$ is scored and this result is saved within the current dialect in the collection $\aleph$. At line 6, the dialect that produce the highest scored table is selected.

The table uniformity pseudocode is outlined in Algorithm 4. The method uses a set of sentinels to measure table inconsistency through monitoring table changes through the parsed records.

The parameter ${\tau }_0$ 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 [13]. However, when there are two or more dialects with a small variance, the ${\tau }_0$ parameter is not decisive. It is in this situation where the ${\tau }_1$ parameter provides support by penalizing tables with variations in its records structures, and whose structure resembles sparse data that do not maintain consistency.

The following illustration shows the preview from the modified content of one of the files used during the testing phase. It was published in the CleverCSV repository on GitHub4.

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 the file is compliant with RFC-4180 specifications. When running dialect detection, CleverCSV gets the vertical bar “|” as the delimiter as this field pattern gets a $P=93.6395$ score vs a $P= 37.647059$ from the patterns with the “,” character as delimiter. This behavior is due to the fact that the logic used strongly weights the count of delimiters over the detected data types. This behavior is due to the fact that the logic used strongly weights the delimiter count over the types of data detected, where dialects containing the comma as delimiter obtain a type score of $T=0.942647$ against the type score of $T=0.843074$ obtained by dialects with the vertical bar as delimiter.

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

Algorithm 4: Table Uniformity

8. Experiments

It was decided to code the new method and integrate it with CSV Interface5, a VBA CSV file parser. Thus, the new CSV dialect determination method will be available in a widespread programming language without overinvesting efforts.

The new solution was tested on two datasets, both on GitHub: the one provided by Gerardo Vitagliano et al., and available in the Pollock framework repository; the other provided by G. van den Burg in the CleverCSV repository. For the first dataset, one or two polluted CSV file per pollution case are used for testing, all the 99 survey 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 [14]). In addition, the dataset was enriched with data from the OpenRefine6 testing, CleverCSV failure cases and other files used at development phase serves as testing samples. In total, the solution was tested against 148 CSV files (104 MB of data) for the simple Pollock testing.

The second dataset is composed of the 256 CSV files in which CleverCSV was not able to accurately determine the dialect at the time of the research that originate the tool [15]. At the time of this research, 244 of these files were available online. A filter was applied to the files to exclude from the dataset all 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), which were used as a ground truth of our dialect detection method. Additionally, these files were subdivided to extract from them a set 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 inference7.

To set up the tests, all files were manually annotated in a file in order to verify, by comparison, the validity of the detected dialects. 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.

8.1. 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.

When using tables of ten or twenty-five records (10R, 25R) for dialect determination, the proposed method was not able to determine the 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 size of the table considered in the computations increases. As can be seen in the Figure 3, when the proposed heuristic is applied, it is settled that the 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. By increasing the table size to twenty-five (25R) 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 the changes in the structure of the tables used in dialect inference.

The results obtained after running the tests over the 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, the 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 the files listed as “normal forms”, were excluded from the dataset, producing the really messy subset of CSV files. Running 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.