String similarity

About the function

String similarity is any measure of how similar any two sequences of characters are. These character strings can be strings of letters or phonemes; both of the methods of calculation included in PCT allow for calculations using either type of character. It is, therefore, a basic measure of overall form-based similarity.

String similarity finds more widespread use in areas of linguistics other than phonology; it is, for example, used in Natural Language Processing applications to determine, for example, possible alternative spellings when a word has been mistyped. It is, however, also useful for determining how phonologically close any two words might be.

String similarity could be part of a calculation of morphological relatedness, if used in conjunction with a measure of semantic similarity (see, e.g., [Hall2014b]). In particular, it can be used in conjunction with the Frequency of Alternation function of PCT (see Frequency of alternation) as a means of calculating the frequency with which two sounds alternate with each other in a language.

Some measure of string similarity is also used to calculate neighbourhood density (e.g. [Greenberg1964]; [Luce1998]; [Yao2011]), which has been shown to affect phonological processing. A phonological “neighbour” of some word X is a word that is similar in some close way to X. For example, it might differ by maximally one phone (through deletion, addition, or subsitution) from X. X’s neighbourhood density, then, is the number of words that fit the criterion for being a neighbour.

Method of calculation

Levenshtein Edit Distance

Edit distance is defined as the minimum number of one-symbol deletions, additions, and substitutions necessary to turn one string into another. For example, turn and burn would have an edit distance of 1, as the only change necessary is to turn the <t> into a <b>, while the edit distance between turn and surfs would be 3, with <t> becoming <s>, <n> becoming <f>, and ∅ becoming <s> at the end of the word. All such one-symbol changes are treated as equal in Levenshtein edit distance, unlike phonological edit distance, described in the following section. Generally speaking, the neighbourhood density of a particular lexical item is measured by summing the number of lexical items that have an edit distance of 1 from that item [Luce1998].

Phonological Edit Distance

Phonological edit distance is quite similar to Levenshtein edit distance, in that it calculates the number of one-symbol changes between strings, but it differs in that changes are weighted based on featural similarity. For example, depending on the feature system used, changing <t> to <s> might involve a single feature change (from [-cont] to [+cont]), while changing <t> to <b> might involve two (from [-voice, +cor] to [+voice, -cor]). By default, the formula for calculating the phonological distance between two segments—or between a segment and “silence”, i.e. insertion or deletion—is the one used in the Sublexical Learner [Allen2014]. When comparing two segments, the distance between them is equal to the sum of the distances between each of their feature values: the distance between two feature values that are identical is 0, while the distance between two opposing values (+/- or -/+) is 1, and the distance between two feature values in the case that just one of them is 0 (unspecified) is set to by default to 0.25. When comparing a segment to “silence” (insertion/deletion), the silence is given feature values of 0 for all features and then compared to the segment as normal.

Khorsi (2012) Similarity Metric

Khorsi (2012) proposes a particular measure of string similarity based on orthography, which he suggests can be used as a direct measure of morphological relatedness. PCT allows one to calculate this measure, which could be used, as Khorsi describes, on its own, or could be used in conjunction with other measures (e.g., semantic similarity) to create a more nuanced view.

This measure starts with the sum of the log of the inverse of the frequency of occurrence of each of the letters in the longest common shared sequence between two words, and then subtracts the sum of the log of the inverse of the frequency of the letters that are not shared, as shown below.

Formula for string similarity from [Khorsi2012]:

\(\sum_{i=1}^{\lVert LCS(w_1,w_2) \rVert} log (\frac{1}{freq(LCS(w_1,w_2)[i])}) - \sum_{i=1}^{\lVert \overline{LCS(w_1,w_2)} \rVert} log (\frac{1}{freq(\overline{LCS(w_1,w_2)}[i])})\)

Note:

  • w1, w2 are two words whose string similarity is to be measured
  • LCS(w1, w2) represents the Longest Common Shared Sequence of symbols between the two words

As with other functions, the frequency measure used for each character will be taken from the current corpus. This means that the score will be different for a given pair of words (e.g., pressed vs. pressure) depending on the frequency of the individual characters in the loaded corpus.

Calculating string similarity in the GUI

To start the analysis, click on “Analysis” / “Calculate string similarity...” in the main menu, and then follow these steps:

  1. String similarity algorithm: The first step is to choose which of the three methods described above is to be used to calculate string similarity. The options are phonological edit distance, standard (Levenshtein) edit distance, and the algorithm described above and in [Khorsi2012].
  2. Comparison type: Next, choose what kind of comparison is to be done. One can either take a single word and get its string similarity score to every other word in the corpus (useful, for example, when trying to figure out which words are most / least similar to a given word, as one might for stimuli creation), or can compare individual pairs of words (useful if a limited set of pre-determined words is of interest). For each of these, you can use words that already exist in the corpus or calculate the similarity for words (or non-words) that are not in the corpus. Note that these words will NOT be added to the corpus itself; if you want to globally add the word (and therefore have its own properties affect calculations), please use the instructions in Adding a word.
    1. One word in the corpus: To compare the similarity of one word that already exists in the corpus to every other word in the corpus, simply select “Compare one word to entire corpus” and enter the single word into the dialogue box, using its standard orthographic representation. Note that you can choose later which tier string similarity will be calculated on (spelling, transcription, etc.); this simply identifies the word for PCT.
    2. One word not in the corpus: Click on “Calculate for a word/nonword not in the corpus” and then select “Create word/nonword” to enter the new word.
      1. Spelling: Enter the spelling for your new word / nonword using the regular input keyboard on your computer.
      2. Transcription: To add in the phonetic transcription of the new word, it is best to use the provided inventory. While it is possible to type directly in to the transcription box, using the provided inventory will ensure that all characters are understood by PCT to correspond to existing characters in the corpus (with their concomitant featural interpretation). Click on “Show inventory.” (See also Edit inventory categories for more on how to set up the inventory window.) Clicking on the individual segments will add them to the transcription. Note that you do NOT need to include word boundaries at the beginning and end of the word, even when the boundary symbol is included as a member of the inventory; these will be assumed automatically by PCT.
      3. Frequency and other columns: These can be left at the default. Note that entering values will NOT affect the calculation; there is no particular need to enter anything here (it is an artifact of using the same dialogue box here as in the “Add Word” function described in Adding a word).
      4. Create word: To finish and return to the “String similarity” dialogue box, click on “Create word.”
    3. Single word pair (in or not in) the corpus: If the similarity of an individual word pair is to be calculated, one can enter the pair directly into the dialogue box. For each word that is in the corpus, simply enter its standard orthographic form. For each word that is not in the corpus, you can add it by selecting “Create word/nonword” and following the steps described immediately above in (2b).
    4. List of pairs of words (in the corpus): If there is a long list of pairs of words, one can simply create a tab-delimited plain .txt file with one word pair per line. In this case, click on “Choose file” and select the .txt file in the resulting system dialogue box. Note that this option is currently available only for words that already exist in the corpus, and that these pairs should be listed using their standard orthographic representations.
  1. Tier: The tier from which string similarity is to be calculated can be selected. Generally, one is likely to care most about either spelling or transcription, but other tiers (e.g., a vowel tier) can also be selected; in this case, all information removed from the tier is ignored. Words should always be entered orthographically (e.g., when telling PCT what word pairs to compare). If similarity is to be calculated on the basis of spelling, words that are entered are broken into their letter components. If similarity is to be calculated on the basis of transcription, the transcriptions are looked up in the corpus, or taken from the created nonword (see step # 1b above).
  2. Pronunciation variants: If the corpus contains multiple pronunciation variants for lexical items, select what strategy should be used. For details, see Pronunciation Variants. Note that here, the only choices currently available are canonical or most-frequent forms.
  3. Frequency type: If Khorsi similarity is to be calculated, the frequencies of the symbols is relevant, and so will be looked up in the currently loaded corpus. Either type frequency or token frequency can be used for the calculation. This option will not be available for either edit distance algorithm, because frequency isn’t taken into account in either one.
  4. Minimum / Maximum similarity: If one is calculating the similarity of one word to all others in the corpus, an arbitrary minimum and maximum can be set to filter out words that are particularly close or distant. For example, one could require that only words with an edit distance of both at least and at most 1 are returned, to get the members of the standard neighbourhood of a particular lexical item. (Recall that the Khorsi calculation is a measure of similarity, while edit distance and phonological edit distance are measures of difference. Thus, a minimum similarity value is analogous to a maximum distance value. PCT will automatically interpret “minimum” and “maximum” relative to the string-similarity algorithm chosen.

Here’s an example for calculating the Khorsi similarity of the pair mata (which occurs in the corpus) and mitoo [mitu] (which does not), in the sample corpus, using token frequencies and comparing transcriptions:

_images/stringsimilaritydialog.png
  1. Results: Once all options have been selected, click “Calculate string similarity.” If this is not the first calculation, and you want to add the results to a pre-existing results table, select the choice that says “add to current results table.” Otherwise, select “start new results table.” A dialogue box will open, showing a table of the results, including word 1, word 2, the result (i.e., the similarity score for Khorsi or distance score for either of the edit algorithms), what choice was made regarding pronunciation variants, whether type or token frequency was used (if the Khorsi method is selected; otherwise, N/A), and which algorithm was used. Note that the entries in the table will be written in spelling regardless of whether spelling or transcriptions were used. This file can be saved to a desired location by selecting “Save to file” at the bottom of the table.

Here’s an example result file for the above selection:

_images/stringsimilarityresults.png

To return to the function dialogue box with your most recently used selections, click on “Reopen function dialog.” Otherwise, the results table can be closed and you will be returned to your corpus view.

Classes and functions

For further details about the relevant classes and functions in PCT’s source code, please refer to Symbol similarity.