What is semantic distance?

Semantic distance is a quantitative measure of relatedness between two concepts (e.g., dog-wolf) within an n-dimensional semantic space. This sounds scarier than it really is. In the simplest case, you could consider the difference in rated scariness between dogs and wolves as a 1-dimension semantic distance. In real life, however, most people model semantic distances using much higher dimensionality.

The ‘SemanticDistance’ R package computes two complementary indices of semantic distance between different chunk sizes of your text. The package contains two embedded lookup databases with multidimensional vectors for many English words. One interesting and convenient feature of multidimensional semantic vectors for words is that they appear to be additive. That is, it is possible to average the semantic vectors for several words (or ngrams) to obtain a composite vector for a group of words. The SemanticDistance package leverages this property to yield comparisons of local (word-to-word) vs. more global (large blocks of text to word, turn-to-turn) semantic relationships.

Distance Metric 1: Experiential

The first database contains crowd-sourced human ratings for 15 dimensions (e.g., color, sound, motor, pleasantness). When you are interested in compouting semantic distance from this ‘experiential’ semantic space between two words, the program first extracts their respective semantic vectors and then compares them, reported as cosine distance. The vector for ‘dog’ against itself yields a distance of 0. The most distant possible cosine value is 2. We label the output of this distance metric as “SD15” since it relflects distance across 15 psychologically meaningful dimensions.

Distance Metric 2: Embedding

The second semantic distance metric is derived from the GLOVE word embedding model. We trained GLOVE on a vast corpus of American English. The result of this training is a lookup database with more than 100k English words characterized across 300 dimensions. These embedding dimensions are not like the psychological dimensions measured in SD15. The embeddings are mathematical constructs (hyperparameters) derived through dimension reduction reflecting co-occurence statistics between each word in the database to all other words in the database. SemanticDistance computes the second semantic distance value using these embeddings and reports the distance (dog-wolf) as ‘GLO_cosdist’. An adjusted cosine value of zero is the least distant (most similar) and two is the most distant (least similar) index between elements.

What kinds of questions can SemanticDistance answer?

1. What is the semantic distance from each word to the next across an ordered story?
   
2. What is the semantic distance turn-to-turn across two partners engaged in a conversation?
   
3. What are the semantic distanceces between all possible pairs of a bag-of-words?
   
4. What is the semantic distance from each new word to the first block of n-words in the sample?
   
5. What are the semantic distances correponding to my word pairs of interest?
   
6. What are the semantic distances corresponding to each new word relative to the previous n-words in a language sample?
   

Prepping my language sample for SemanticDistance

The first steps before you do anything with the SemanticDistance package is to figure out what format your language transcript is in and what you want it to measure. SemanticDistance offers many possible options with some default arguments that you need to understand. For example, the package requires users to clean and prep the data. You can choose to omit stopwords, lemmatize, split strings, etc. – or just tell R to leave your data alone and just split the transcript into a one word per row format.

SemanticDistance can compute pairwise semantic distance relationships in ordered and unordered language samples. Format your language samples in one of the following structures:
1. Monologues: Any ordered text sample NOT delineated by a talker/speaker (e.g., stories, narratives). The minimal requirement for a monologue is one row and one column with some text in it. Name your text column, text or some other descriptive variable name you can keep track of
2. Dialogues: An ordered language sample split by a talker/speaker/interlocutor factor. The minimum requirment is two cells with interlocutor identity and some text. Name your speaker column as speaker, talker, or person – some identity variable you can keep track of. Name their corresponding text as text or some other descriptive variable name you can keep track of
3. Word Pairs in Columns: Paired string data arrayed across two columns (e.g., Dog-Leash). Two columns of paired data. Each cell has one word in it. Name your column headers something descriptive (e.g., Word1, Word2).
4. Unordered Word Lists: Unordered list of words (nominally one column, all text in one row) that will be transformed into a distance matrix, network model, or dendrogram. This fornat just requires one undifferentiated chunk of data (e.g., a string arrayed in a text file)

Prep Data Outside the Package

1. Store your text and project files within a dedicated folder/directory (e.g., ‘mytexts/’)
   
2. Format your data as CSV or txt. Although SemanticDistance is fairly robust to different character encodings, many proograms such as Excel introduce weird hidden characters into strings.
   
3. Label your target text and metadata columns offline however you like (e.g., mytext, word, langoutput)
   
4. Import your text and associated metadata (e.g., document_id, timestamps, etc.) as a dataframe.
   
5. Identify the format of your sample (e.g., monologue, dialogue, columns, unstructured).
   
6. Install and load the SemanticDistance package
   
7. Choose a principled set of cleaning parameters (e.g., should I omit stopwords? should I lemmatize?)
   

We have created separate vignettes for specifically dealing with each of these formats.

Read your data into R

It doesn’t matter how you get your data into R, but there are some formatting considerations to keep in mind. Excel is dangerous in terms of adding hidden characters and other nonsense (e.g., formulae) that you don’t want in your dataframe. Remain vigilant and check whether the data imported into R matches your original. SemanticDistance will retain all your metadata (e.g., timestamps, speaker codes).

Clean, format, and run distances on your data

See the appropriate vignette that correposponds to your format of interest.