This post is based on our RANLP 2023 paper “Exploring the Landscape of Natural Language Processing Research”. You can read more details there.
As an efficient approach to understand, generate, and process natural language texts, research in natural language processing (NLP) has exhibited a rapid spread and wide adoption in recent years. Given the rapid developments in NLP, obtaining an overview of the domain and maintaining it is difficult. This blog post aims to provide a structured overview of different fields of study NLP and analyzes recent trends in this domain.
Fields of study are academic disciplines and concepts that usually consist of (but are not limited to) tasks or techniques.
In this article, we investigate the following questions:
- What are the different fields of study investigated in NLP?
- What are the characteristics and developments over time of the research literature in NLP?
- What are the current trends and directions of future work in NLP?
Although most fields of study in NLP are well-known and defined, there currently exists no commonly used taxonomy or categorization scheme that attempts to collect and structure these fields of study in a consistent and understandable format. Therefore, getting an overview of the entire field of NLP research is difficult. While there are lists of NLP topics in conferences and textbooks, they tend to vary considerably and are often either too broad or too specialized. Therefore, we developed a taxonomy encompassing a wide range of different fields of study in NLP. Although this taxonomy may not include all possible NLP concepts, it covers a wide range of the most popular fields of study, whereby missing fields of study may be considered as subtopics of the included fields of study. While developing the taxonomy, we found that certain lower-level fields of study had to be assigned to multiple higher-level fields of study rather than just one. Therefore, some fields of study are listed multiple times in the NLP taxonomy, but assigned to different higher-level fields of study. The final taxonomy was developed empirically in an iterative process together with domain experts.
The taxonomy serves as an overarching classification scheme in which NLP publications can be classified according to at least one of the included fields of study, even if they do not directly address one of the fields of study, but only subtopics thereof. To analyze recent developments in NLP, we trained a weakly supervised model to classify ACL Anthology papers according to the NLP taxonomy.
You can read more details about the development process of the classification model and the NLP taxonomy in our paper.
The following section provides short explanations of the fields of study concepts included in the NLP taxonomy above.
Multimodality
“Multimodality refers to the capability of a system or method to process input of different types or modalities” (Garg et al., 2022). We distinguish between systems that can process text in natural language along with visual data, speech & audio, programming languages, or structured data such as tables or graphs.
Natural Language Interfaces
“Natural language interfaces can process data based on natural language queries” (Voigt et al., 2021), usually implemented as question answering or dialogue & conversational systems.
Semantic Text Processing
This high-level field of study includes all types of concepts that attempt to derive meaning from natural language and enable machines to interpret textual data semantically. One of the most powerful fields of study in this regard are “language models that attempt to learn the joint probability function of sequences of words” (Bengio et al., 2000). “Recent advances in language model training have enabled these models to successfully perform various downstream NLP tasks” (Soni et al., 2022). In representation learning, “semantic text representations are usually learned in the form of embeddings” (Fu et al., 2022), which “can be used to compare the semantic similarity of texts in semantic search settings” (Reimers and Gurevych, 2019). Additionally, “knowledge representations, e.g., in the form of knowledge graphs, can be incorporated to improve various NLP tasks” (Schneider et al., 2022).
Sentiment Analysis
“Sentiment analysis attempts to identify and extract subjective information from texts” (Wankhade et al., 2022). Usually, studies focus on extracting opinions, emotions, or polarity from texts. More recently, aspect-based sentiment analysis emerged as a way to provide more detailed information than general sentiment analysis, as “it aims to predict the sentiment polarities of given aspects or entities in text” (Xue and Li, 2018).
Syntactic Text Processing
This high-level field of study aims at “analyzing the grammatical syntax and vocabulary of texts” (Bessmertny et al., 2016). Representative tasks in this context are syntactic parsing of word dependencies in sentences, tagging of words to their respective part-of-speech, segmentation of texts into coherent sections, or correction of erroneous texts with respect to grammar and spelling.
Linguistics & Cognitive NLP
“Linguistics & Cognitive NLP deals with natural language based on the assumptions that our linguistic abilities are firmly rooted in our cognitive abilities, that meaning is essentially conceptualization, and that grammar is shaped by usage” (Dabrowska and Divjak, 2015). Many different linguistic theories are present that generally argue that “language acquisition is governed by universal grammatical rules that are common to all typically developing humans” (Wise and Sevcik, 2017). “Psycholinguistics attempts to model how a human brain acquires and produces language, processes it, comprehends it, and provides feedback” (Balamurugan, 2018). “Cognitive modeling is concerned with modeling and simulating human cognitive processes in various forms, particularly in a computational or mathematical form” (Sun, 2020).
Responsible & Trustworthy NLP
“Responsible & trustworthy NLP is concerned with implementing methods that focus on fairness, explainability, accountability, and ethical aspects at its core” (Barredo Arrieta et al., 2020). Green & sustainable NLP is mainly focused on efficient approaches for text processing, while low-resource NLP aims to perform NLP tasks when data is scarce. Additionally, robustness in NLP attempts to develop models that are insensitive to biases, resistant to data perturbations, and reliable for out-of-distribution predictions.
Reasoning
Reasoning enables machines to draw logical conclusions and derive new knowledge based on the information available to them, using techniques such as deduction and induction. “Argument mining automatically identifies and extracts the structure of inference and reasoning expressed as arguments presented in natural language texts2 (Lawrence and Reed, 2019). “Textual inference, usually modeled as entailment problem, automatically determines whether a natural-language hypothesis can be inferred from a given premise” (MacCartney and Manning, 2007). “Commonsense reasoning bridges premises and hypotheses using world knowledge that is not explicitly provided in the text” (Ponti et al., 2020), while “numerical reasoning performs arithmetic operations” (Al-Negheimish et al., 2021). “Machine reading comprehension aims to teach machines to determine the correct answers to questions based on a given passage” (Zhang et al., 2021).
Multilinguality
Multilinguality tackles all types of NLP tasks that involve more than one natural language and is conventionally studied in machine translation. Additionally, “code-switching freely interchanges multiple languages within a single sentence or between sentences” (Diwan et al., 2021), while cross-lingual transfer techniques use data and models available for one language to solve NLP tasks in another language.
Information Retrieval
“Information retrieval is concerned with finding texts that satisfy an information need from within large collections” (Manning et al., 2008). Typically, this involves retrieving documents or passages.
Information Extraction & Text Mining
This field of study focuses on extracting structured knowledge from unstructured text and “enables the analysis and identification of patterns or correlations in data” (Hassani et al., 2020). “Text classification automatically categorizes texts into predefined classes” (Schopf et al., 2021), while “topic modeling aims to discover latent topics in document collections” (Grootendorst, 2022), often using text clustering techniques that organize semantically similar texts into the same clusters. “Summarization produces summaries of texts that include the key points of the input in less space and keep repetition to a minimum” (El-Kassas et al., 2021). Additionally, the information extraction & text mining field of study also includes “named entity recognition, which deals with the identification and categorization of named entities” (Leitner et al., 2020), “coreference resolution, which aims to identify all references to the same entity in discourse” (Yin et al., 2021), “term extraction, which aims to extract relevant terms such as keywords or keyphrases” (Rigouts Terryn et al., 2020), relation extraction that aims to extract relations between entities, and “open information extraction that facilitates the domain-independent discovery of relational tuples” (Yates et al., 2007).
Text Generation
The objective of text generation approaches is to generate texts that are both comprehensible to humans and indistinguishable from text authored by humans. Accordingly, the input usually consists of text, such as in “paraphrasing that renders the text input in a different surface form while preserving the semantics” (Niu et al., 2021), “question generation that aims to generate a fluid and relevant question given a passage and a target answer” (Song et al., 2018), or “dialogue-response generation which aims to generate natural-looking text relevant to the prompt” (Zhang et al., 2020). In many cases, however, the text is generated as a result of input from other modalities, such as in the case of “data-to-text generation that generates text based on structured data such as tables or graphs” (Kale and Rastogi, 2020), captioning of images or videos, or “speech recognition that transcribes a speech waveform into text” (Baevski et al., 2022).
Considering the literature on NLP, we start our analysis with the number of studies as an indicator of research interest. The distribution of publications over the 50-year observation period is shown in the Figure above. While the first publications appeared in 1952, the number of annual publications grew slowly until 2000. Accordingly, between 2000 and 2017, the number of publications roughly quadrupled, whereas in the subsequent five years, it has doubled again. We therefore observe a near-exponential growth in the number of NLP studies, indicating increasing attention from the research community.
