Commit | Line | Data |
---|---|---|
5ca28f79 L |
1 | A Fast Method for Identifying Plain Text Files |
2 | ============================================== | |
3 | ||
4 | ||
5 | Introduction | |
6 | ------------ | |
7 | ||
8 | Given a file coming from an unknown source, it is sometimes desirable | |
9 | to find out whether the format of that file is plain text. Although | |
10 | this may appear like a simple task, a fully accurate detection of the | |
11 | file type requires heavy-duty semantic analysis on the file contents. | |
12 | It is, however, possible to obtain satisfactory results by employing | |
13 | various heuristics. | |
14 | ||
15 | Previous versions of PKZip and other zip-compatible compression tools | |
16 | were using a crude detection scheme: if more than 80% (4/5) of the bytes | |
17 | found in a certain buffer are within the range [7..127], the file is | |
18 | labeled as plain text, otherwise it is labeled as binary. A prominent | |
19 | limitation of this scheme is the restriction to Latin-based alphabets. | |
20 | Other alphabets, like Greek, Cyrillic or Asian, make extensive use of | |
21 | the bytes within the range [128..255], and texts using these alphabets | |
22 | are most often misidentified by this scheme; in other words, the rate | |
23 | of false negatives is sometimes too high, which means that the recall | |
24 | is low. Another weakness of this scheme is a reduced precision, due to | |
25 | the false positives that may occur when binary files containing large | |
26 | amounts of textual characters are misidentified as plain text. | |
27 | ||
28 | In this article we propose a new, simple detection scheme that features | |
29 | a much increased precision and a near-100% recall. This scheme is | |
30 | designed to work on ASCII, Unicode and other ASCII-derived alphabets, | |
31 | and it handles single-byte encodings (ISO-8859, MacRoman, KOI8, etc.) | |
32 | and variable-sized encodings (ISO-2022, UTF-8, etc.). Wider encodings | |
33 | (UCS-2/UTF-16 and UCS-4/UTF-32) are not handled, however. | |
34 | ||
35 | ||
36 | The Algorithm | |
37 | ------------- | |
38 | ||
39 | The algorithm works by dividing the set of bytecodes [0..255] into three | |
40 | categories: | |
41 | - The white list of textual bytecodes: | |
42 | 9 (TAB), 10 (LF), 13 (CR), 32 (SPACE) to 255. | |
43 | - The gray list of tolerated bytecodes: | |
44 | 7 (BEL), 8 (BS), 11 (VT), 12 (FF), 26 (SUB), 27 (ESC). | |
45 | - The black list of undesired, non-textual bytecodes: | |
46 | 0 (NUL) to 6, 14 to 31. | |
47 | ||
48 | If a file contains at least one byte that belongs to the white list and | |
49 | no byte that belongs to the black list, then the file is categorized as | |
50 | plain text; otherwise, it is categorized as binary. (The boundary case, | |
51 | when the file is empty, automatically falls into the latter category.) | |
52 | ||
53 | ||
54 | Rationale | |
55 | --------- | |
56 | ||
57 | The idea behind this algorithm relies on two observations. | |
58 | ||
59 | The first observation is that, although the full range of 7-bit codes | |
60 | [0..127] is properly specified by the ASCII standard, most control | |
61 | characters in the range [0..31] are not used in practice. The only | |
62 | widely-used, almost universally-portable control codes are 9 (TAB), | |
63 | 10 (LF) and 13 (CR). There are a few more control codes that are | |
64 | recognized on a reduced range of platforms and text viewers/editors: | |
65 | 7 (BEL), 8 (BS), 11 (VT), 12 (FF), 26 (SUB) and 27 (ESC); but these | |
66 | codes are rarely (if ever) used alone, without being accompanied by | |
67 | some printable text. Even the newer, portable text formats such as | |
68 | XML avoid using control characters outside the list mentioned here. | |
69 | ||
70 | The second observation is that most of the binary files tend to contain | |
71 | control characters, especially 0 (NUL). Even though the older text | |
72 | detection schemes observe the presence of non-ASCII codes from the range | |
73 | [128..255], the precision rarely has to suffer if this upper range is | |
74 | labeled as textual, because the files that are genuinely binary tend to | |
75 | contain both control characters and codes from the upper range. On the | |
76 | other hand, the upper range needs to be labeled as textual, because it | |
77 | is used by virtually all ASCII extensions. In particular, this range is | |
78 | used for encoding non-Latin scripts. | |
79 | ||
80 | Since there is no counting involved, other than simply observing the | |
81 | presence or the absence of some byte values, the algorithm produces | |
82 | consistent results, regardless what alphabet encoding is being used. | |
83 | (If counting were involved, it could be possible to obtain different | |
84 | results on a text encoded, say, using ISO-8859-16 versus UTF-8.) | |
85 | ||
86 | There is an extra category of plain text files that are "polluted" with | |
87 | one or more black-listed codes, either by mistake or by peculiar design | |
88 | considerations. In such cases, a scheme that tolerates a small fraction | |
89 | of black-listed codes would provide an increased recall (i.e. more true | |
90 | positives). This, however, incurs a reduced precision overall, since | |
91 | false positives are more likely to appear in binary files that contain | |
92 | large chunks of textual data. Furthermore, "polluted" plain text should | |
93 | be regarded as binary by general-purpose text detection schemes, because | |
94 | general-purpose text processing algorithms might not be applicable. | |
95 | Under this premise, it is safe to say that our detection method provides | |
96 | a near-100% recall. | |
97 | ||
98 | Experiments have been run on many files coming from various platforms | |
99 | and applications. We tried plain text files, system logs, source code, | |
100 | formatted office documents, compiled object code, etc. The results | |
101 | confirm the optimistic assumptions about the capabilities of this | |
102 | algorithm. | |
103 | ||
104 | ||
105 | -- | |
106 | Cosmin Truta | |
107 | Last updated: 2006-May-28 |