Measuring the similarity of marks: an overview of suggested ideas

Introduction

A comparison of the similarity between pairs of marks is a key component of many intellectual property disputes. A key point to note is that the overall assessment of similarity needs to take account of a number of components, several of which involve subjective determinations. These components might typically include: 'inherent' characteristics of the marks in question; the meaning of any terms (i.e. conceptual analysis); their distinctiveness, strength and degree of renown; the influence of any associated logos, imagery or mark stylisation; the degree of overlap of associated goods and services; documented evidence of actual confusion; and the degree of attention paid by a typical consumer - many of which may vary between different geographical regions. All of these factors contribute to the overall assessment of the likelihood of confusion between the marks. 

Nevertheless, there are certain characteristics (generally falling under the 'inherent' category referenced above) of some types of marks which do lend themselves to a quantitative, objective measurement of similarity. The most obvious such examples are colours, and the spelling and pronunciation of word marks (which contribute to visual and aural similarity, respectively). Whilst any measurement of such characteristics cannot provide a quantification of overall mark similarity, the associated algorithms can provide a useful tool to be utilised in the assessment process. 

Algorithms to measure colour similarity, and visual and aural similarity for word marks, have a number of obvious applications. Firstly, they offer the potential for greater consistency (and greater granularity) in the assessment of the respective types of similarity across dispute cases, and secondly, they offer the potential to be able to specify quantifiable thresholds up to which IP protection might apply. In addition, they have other applications, such as the option to post-process results from trademark watching services, to (better) sort and prioritise the findings and assist with the review process

A group of suggested frameworks for formulating algorithms of this type was set out in a series of six articles^[1] recently published on the CircleID website. This overview presents a summary of the key ideas from the series.

Similarity of colours

Colours occupy a unique position in the set of mark types, due to the fact that the specification for a colour can be exactly defined. One of the most common frameworks (particularly in the context of digital display systems) is the RGB framework, where a colour is defined in terms of its red, green and blue components, each expressed as an integer value from 0 to 255 (giving 256³ or 16.8 million definable colours in total). This 'universe' of possible colours can therefore be visualised as a three-dimensional cube (or colour 'space'), with red increasing along one axis, green along another, and blue along the third, with each distinct colour occupying a unique point within the space.

Using this framework, the (degree of) difference between any two colours can be expressed in terms of their geometric distance (d) from each other in the colour space. From this, a difference score (D_col) can be formulated (by expressing d as a proportion of the maximum possible distance between two colours - i.e. between [0,0,0] (black) and [255,255,255] (white)), and from this, a colour similarity score (S_col) (equal to 1 (or 100%) minus D_col).

The concept of a colour similarity measurement metric makes most sense (in the context of disputes) if there were to exist a framework in which the protection granted under a colour mark (within appropriate categories of goods and services) registration covered not just that colour exactly, but also very similar colours (up to a specified threshold). Current guidelines suggest that protection should cover variants such that the difference between the shades is 'barely noticeable', but the use of a numerical score would provide the potential to put in place a more explicit threshold and avoid ambiguity. 

Within a framework of this type, it would also be possible (and might be convenient) to maintain a database of protected colours (or the colours of elements within broader protected marks), to help determine the existence of possible clashes between existing or proposed new protected colours. A mock-up of how this might look is shown in Table 1, for a series of colours associated with well-known brands. In the figure, the individual colours are sorted by their hue (H) values (part of an alternative (to RGB) framework for specifying colours), which orders them according to the position within the spectrum of their dominant colour, which can assist with visual review of data of this type.

Table 1: Mock-up of a database of protected colour marks, sorted by their H values

For context, the shades of orange used by Reese's and Home Depot (objectively the most similar pair of colours in the above table) have a similarity score of 95.9%.

Visual and aural similarity of word marks

For word marks (even just considering similarity in spelling (visual) and pronunciation (aural)), the situation is rather more complex. The frameworks suggested in the previous studies propose a separate similarity score for each of these two components (S_vis and S_aur, respectively), and an overall score (S_wor) reflecting both types of word similarity, which is most simply calculated as the mean of the two components (but can be differently weighted if required).

The proposed algorithm for quantifying visual (spelling) similarity is itself composed of two components (i.e. utilises two distinct metrics), reflecting different aspects of the similarity in spelling. The first metric ('fuzz.ratio') is based on a measurement called Levenshtein distance, which quantifies the number of 'edits' (i.e. character insertions, deletions, or substitutions) necessary to transform one string into the other), but with the metric also including normalisation factors to take account of the length of the strings, and the second (Jaro-Winkler similarity) is more complex, including an element which takes account of the proximity of the variations to the start of the strings (where, for example, a consumer might be more likely to be aware of any differences). 

For aural (pronunciation) similarity, the calculation is carried out by first implementing an analysis process which converts each string to its IPA (International Phonetic Alphabet) representation, and then using the 'fuzz.ratio' metric to quantify the similarity between these representations.

For illustration, the similarity score values for a range of pairs of marks which were the subject of past disputes is shown in Table 2. 

Table 2: Pairs of marks and their visual, aural and overall similarity scores

Conclusion

The algorithms proposed for quantifying the similarity of colour marks, and the visual and aural similarity of word marks, do seem to perform reasonably well, and (in the case of the word mark metrics) aligns with what might be subjectively be reckoned according to manual analysis. The formulations are, of course, just one possible option, and it would certainly be possible to 'tune' the algorithms according to specific requirements.

Algorithms of this type do offer the potential for a more granular, continuous, repeatable and quantifiable expression of similarity and, with appropriate adoption into case analysis, offer a possible route towards greater consistency in dispute decisions. 

However, it is important to reiterate the statement made in the introduction, that such metrics cannot fully assess the overall similarity between marks, or replace the existing nuanced and multi-faceted approach of considering the full range of subjective factors which contribute to an assessment of the likelihood of confusion, but can provide a useful tool to be applied in such analyses and in other contexts.

Reference

[1] For colour marks:

• 1. 'Towards a quantitative approach for objectively measuring the similarity of marks' (https://circleid.com/pdf/similarity_measurement_of_marks_part_1.pdf) (Part 1 - 'Colour, similarity, and intellectual property')
• 2. 'Further developing a colour mark similarity measurement framework: Building a database' (https://circleid.com/posts/further-developing-a-colour-mark-similarity-measurement-framework-building-a-database)
• 4. ‘'Further developing a colour mark similarity measurement framework - Part II: Defining a similarity score' (https://circleid.com/pdf/similarity_measurement_of_marks_part_4.pdf) 
• 6. 'Further developing a colour mark similarity measurement framework - Part III: A method for sorting colours' (https://circleid.com/pdf/similarity_measurement_of_marks_part_6.pdf) 

For word marks:

• 1. 'Towards a quantitative approach for objectively measuring the similarity of marks' (https://circleid.com/pdf/similarity_measurement_of_marks_part_1.pdf) (Part 2 - 'Explorations in similarity measurement for word marks')
• 3. 'Further developing a word mark similarity measurement framework' (Part 1 - 'Testing the algorithm'; Part 2 - 'Subsequences and substrings'; Part 3 - 'Phonemizing') (https://circleid.com/pdf/similarity_measurement_of_marks_part_3.pdf) 
• 5. 'Further developing a word mark similarity measurement framework - Part II: Defining an improved similarity score' (https://circleid.com/pdf/similarity_measurement_of_marks_part_5.pdf) 

This article was first published on 17 October 2024 at:

https://www.linkedin.com/pulse/measuring-similarity-marks-overview-suggested-ideas-david-barnett-zo7fe/